1 //===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 /// \file 9 /// This file implements a TargetTransformInfo analysis pass specific to the 10 /// X86 target machine. It uses the target's detailed information to provide 11 /// more precise answers to certain TTI queries, while letting the target 12 /// independent and default TTI implementations handle the rest. 13 /// 14 //===----------------------------------------------------------------------===// 15 /// About Cost Model numbers used below it's necessary to say the following: 16 /// the numbers correspond to some "generic" X86 CPU instead of usage of 17 /// concrete CPU model. Usually the numbers correspond to CPU where the feature 18 /// apeared at the first time. For example, if we do Subtarget.hasSSE42() in 19 /// the lookups below the cost is based on Nehalem as that was the first CPU 20 /// to support that feature level and thus has most likely the worst case cost. 21 /// Some examples of other technologies/CPUs: 22 /// SSE 3 - Pentium4 / Athlon64 23 /// SSE 4.1 - Penryn 24 /// SSE 4.2 - Nehalem 25 /// AVX - Sandy Bridge 26 /// AVX2 - Haswell 27 /// AVX-512 - Xeon Phi / Skylake 28 /// And some examples of instruction target dependent costs (latency) 29 /// divss sqrtss rsqrtss 30 /// AMD K7 11-16 19 3 31 /// Piledriver 9-24 13-15 5 32 /// Jaguar 14 16 2 33 /// Pentium II,III 18 30 2 34 /// Nehalem 7-14 7-18 3 35 /// Haswell 10-13 11 5 36 /// TODO: Develop and implement the target dependent cost model and 37 /// specialize cost numbers for different Cost Model Targets such as throughput, 38 /// code size, latency and uop count. 39 //===----------------------------------------------------------------------===// 40 41 #include "X86TargetTransformInfo.h" 42 #include "llvm/Analysis/TargetTransformInfo.h" 43 #include "llvm/CodeGen/BasicTTIImpl.h" 44 #include "llvm/CodeGen/CostTable.h" 45 #include "llvm/CodeGen/TargetLowering.h" 46 #include "llvm/IR/IntrinsicInst.h" 47 #include "llvm/Support/Debug.h" 48 49 using namespace llvm; 50 51 #define DEBUG_TYPE "x86tti" 52 53 //===----------------------------------------------------------------------===// 54 // 55 // X86 cost model. 56 // 57 //===----------------------------------------------------------------------===// 58 59 TargetTransformInfo::PopcntSupportKind 60 X86TTIImpl::getPopcntSupport(unsigned TyWidth) { 61 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2"); 62 // TODO: Currently the __builtin_popcount() implementation using SSE3 63 // instructions is inefficient. Once the problem is fixed, we should 64 // call ST->hasSSE3() instead of ST->hasPOPCNT(). 65 return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software; 66 } 67 68 llvm::Optional<unsigned> X86TTIImpl::getCacheSize( 69 TargetTransformInfo::CacheLevel Level) const { 70 switch (Level) { 71 case TargetTransformInfo::CacheLevel::L1D: 72 // - Penryn 73 // - Nehalem 74 // - Westmere 75 // - Sandy Bridge 76 // - Ivy Bridge 77 // - Haswell 78 // - Broadwell 79 // - Skylake 80 // - Kabylake 81 return 32 * 1024; // 32 KByte 82 case TargetTransformInfo::CacheLevel::L2D: 83 // - Penryn 84 // - Nehalem 85 // - Westmere 86 // - Sandy Bridge 87 // - Ivy Bridge 88 // - Haswell 89 // - Broadwell 90 // - Skylake 91 // - Kabylake 92 return 256 * 1024; // 256 KByte 93 } 94 95 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel"); 96 } 97 98 llvm::Optional<unsigned> X86TTIImpl::getCacheAssociativity( 99 TargetTransformInfo::CacheLevel Level) const { 100 // - Penryn 101 // - Nehalem 102 // - Westmere 103 // - Sandy Bridge 104 // - Ivy Bridge 105 // - Haswell 106 // - Broadwell 107 // - Skylake 108 // - Kabylake 109 switch (Level) { 110 case TargetTransformInfo::CacheLevel::L1D: 111 LLVM_FALLTHROUGH; 112 case TargetTransformInfo::CacheLevel::L2D: 113 return 8; 114 } 115 116 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel"); 117 } 118 119 unsigned X86TTIImpl::getNumberOfRegisters(unsigned ClassID) const { 120 bool Vector = (ClassID == 1); 121 if (Vector && !ST->hasSSE1()) 122 return 0; 123 124 if (ST->is64Bit()) { 125 if (Vector && ST->hasAVX512()) 126 return 32; 127 return 16; 128 } 129 return 8; 130 } 131 132 TypeSize 133 X86TTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const { 134 unsigned PreferVectorWidth = ST->getPreferVectorWidth(); 135 switch (K) { 136 case TargetTransformInfo::RGK_Scalar: 137 return TypeSize::getFixed(ST->is64Bit() ? 64 : 32); 138 case TargetTransformInfo::RGK_FixedWidthVector: 139 if (ST->hasAVX512() && PreferVectorWidth >= 512) 140 return TypeSize::getFixed(512); 141 if (ST->hasAVX() && PreferVectorWidth >= 256) 142 return TypeSize::getFixed(256); 143 if (ST->hasSSE1() && PreferVectorWidth >= 128) 144 return TypeSize::getFixed(128); 145 return TypeSize::getFixed(0); 146 case TargetTransformInfo::RGK_ScalableVector: 147 return TypeSize::getScalable(0); 148 } 149 150 llvm_unreachable("Unsupported register kind"); 151 } 152 153 unsigned X86TTIImpl::getLoadStoreVecRegBitWidth(unsigned) const { 154 return getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) 155 .getFixedSize(); 156 } 157 158 unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) { 159 // If the loop will not be vectorized, don't interleave the loop. 160 // Let regular unroll to unroll the loop, which saves the overflow 161 // check and memory check cost. 162 if (VF == 1) 163 return 1; 164 165 if (ST->isAtom()) 166 return 1; 167 168 // Sandybridge and Haswell have multiple execution ports and pipelined 169 // vector units. 170 if (ST->hasAVX()) 171 return 4; 172 173 return 2; 174 } 175 176 InstructionCost X86TTIImpl::getArithmeticInstrCost( 177 unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, 178 TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info, 179 TTI::OperandValueProperties Opd1PropInfo, 180 TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args, 181 const Instruction *CxtI) { 182 // TODO: Handle more cost kinds. 183 if (CostKind != TTI::TCK_RecipThroughput) 184 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, 185 Op2Info, Opd1PropInfo, 186 Opd2PropInfo, Args, CxtI); 187 188 // vXi8 multiplications are always promoted to vXi16. 189 if (Opcode == Instruction::Mul && Ty->isVectorTy() && 190 Ty->getScalarSizeInBits() == 8) { 191 Type *WideVecTy = 192 VectorType::getExtendedElementVectorType(cast<VectorType>(Ty)); 193 return getCastInstrCost(Instruction::ZExt, WideVecTy, Ty, 194 TargetTransformInfo::CastContextHint::None, 195 CostKind) + 196 getCastInstrCost(Instruction::Trunc, Ty, WideVecTy, 197 TargetTransformInfo::CastContextHint::None, 198 CostKind) + 199 getArithmeticInstrCost(Opcode, WideVecTy, CostKind, Op1Info, Op2Info, 200 Opd1PropInfo, Opd2PropInfo); 201 } 202 203 // Legalize the type. 204 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 205 206 int ISD = TLI->InstructionOpcodeToISD(Opcode); 207 assert(ISD && "Invalid opcode"); 208 209 if (ISD == ISD::MUL && Args.size() == 2 && LT.second.isVector() && 210 LT.second.getScalarType() == MVT::i32) { 211 // Check if the operands can be represented as a smaller datatype. 212 bool Op1Signed = false, Op2Signed = false; 213 unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed); 214 unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed); 215 unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize); 216 217 // If both are representable as i15 and at least one is constant, 218 // zero-extended, or sign-extended from vXi16 then we can treat this as 219 // PMADDWD which has the same costs as a vXi16 multiply. 220 if (OpMinSize <= 15 && !ST->isPMADDWDSlow()) { 221 bool Op1Constant = 222 isa<ConstantDataVector>(Args[0]) || isa<ConstantVector>(Args[0]); 223 bool Op2Constant = 224 isa<ConstantDataVector>(Args[1]) || isa<ConstantVector>(Args[1]); 225 bool Op1Sext16 = isa<SExtInst>(Args[0]) && Op1MinSize == 15; 226 bool Op2Sext16 = isa<SExtInst>(Args[1]) && Op2MinSize == 15; 227 228 bool IsZeroExtended = !Op1Signed || !Op2Signed; 229 bool IsConstant = Op1Constant || Op2Constant; 230 bool IsSext16 = Op1Sext16 || Op2Sext16; 231 if (IsConstant || IsZeroExtended || IsSext16) 232 LT.second = 233 MVT::getVectorVT(MVT::i16, 2 * LT.second.getVectorNumElements()); 234 } 235 } 236 237 if ((ISD == ISD::SDIV || ISD == ISD::SREM || ISD == ISD::UDIV || 238 ISD == ISD::UREM) && 239 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 240 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 241 Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) { 242 if (ISD == ISD::SDIV || ISD == ISD::SREM) { 243 // On X86, vector signed division by constants power-of-two are 244 // normally expanded to the sequence SRA + SRL + ADD + SRA. 245 // The OperandValue properties may not be the same as that of the previous 246 // operation; conservatively assume OP_None. 247 InstructionCost Cost = 248 2 * getArithmeticInstrCost(Instruction::AShr, Ty, CostKind, Op1Info, 249 Op2Info, TargetTransformInfo::OP_None, 250 TargetTransformInfo::OP_None); 251 Cost += getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, Op1Info, 252 Op2Info, TargetTransformInfo::OP_None, 253 TargetTransformInfo::OP_None); 254 Cost += getArithmeticInstrCost(Instruction::Add, Ty, CostKind, Op1Info, 255 Op2Info, TargetTransformInfo::OP_None, 256 TargetTransformInfo::OP_None); 257 258 if (ISD == ISD::SREM) { 259 // For SREM: (X % C) is the equivalent of (X - (X/C)*C) 260 Cost += getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, Op1Info, 261 Op2Info); 262 Cost += getArithmeticInstrCost(Instruction::Sub, Ty, CostKind, Op1Info, 263 Op2Info); 264 } 265 266 return Cost; 267 } 268 269 // Vector unsigned division/remainder will be simplified to shifts/masks. 270 if (ISD == ISD::UDIV) 271 return getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, Op1Info, 272 Op2Info, TargetTransformInfo::OP_None, 273 TargetTransformInfo::OP_None); 274 // UREM 275 return getArithmeticInstrCost(Instruction::And, Ty, CostKind, Op1Info, 276 Op2Info, TargetTransformInfo::OP_None, 277 TargetTransformInfo::OP_None); 278 } 279 280 static const CostTblEntry GLMCostTable[] = { 281 { ISD::FDIV, MVT::f32, 18 }, // divss 282 { ISD::FDIV, MVT::v4f32, 35 }, // divps 283 { ISD::FDIV, MVT::f64, 33 }, // divsd 284 { ISD::FDIV, MVT::v2f64, 65 }, // divpd 285 }; 286 287 if (ST->useGLMDivSqrtCosts()) 288 if (const auto *Entry = CostTableLookup(GLMCostTable, ISD, 289 LT.second)) 290 return LT.first * Entry->Cost; 291 292 static const CostTblEntry SLMCostTable[] = { 293 { ISD::MUL, MVT::v4i32, 11 }, // pmulld 294 { ISD::MUL, MVT::v8i16, 2 }, // pmullw 295 { ISD::FMUL, MVT::f64, 2 }, // mulsd 296 { ISD::FMUL, MVT::v2f64, 4 }, // mulpd 297 { ISD::FMUL, MVT::v4f32, 2 }, // mulps 298 { ISD::FDIV, MVT::f32, 17 }, // divss 299 { ISD::FDIV, MVT::v4f32, 39 }, // divps 300 { ISD::FDIV, MVT::f64, 32 }, // divsd 301 { ISD::FDIV, MVT::v2f64, 69 }, // divpd 302 { ISD::FADD, MVT::v2f64, 2 }, // addpd 303 { ISD::FSUB, MVT::v2f64, 2 }, // subpd 304 // v2i64/v4i64 mul is custom lowered as a series of long: 305 // multiplies(3), shifts(3) and adds(2) 306 // slm muldq version throughput is 2 and addq throughput 4 307 // thus: 3X2 (muldq throughput) + 3X1 (shift throughput) + 308 // 3X4 (addq throughput) = 17 309 { ISD::MUL, MVT::v2i64, 17 }, 310 // slm addq\subq throughput is 4 311 { ISD::ADD, MVT::v2i64, 4 }, 312 { ISD::SUB, MVT::v2i64, 4 }, 313 }; 314 315 if (ST->isSLM()) { 316 if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) { 317 // Check if the operands can be shrinked into a smaller datatype. 318 // TODO: Merge this into generiic vXi32 MUL patterns above. 319 bool Op1Signed = false; 320 unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed); 321 bool Op2Signed = false; 322 unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed); 323 324 bool SignedMode = Op1Signed || Op2Signed; 325 unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize); 326 327 if (OpMinSize <= 7) 328 return LT.first * 3; // pmullw/sext 329 if (!SignedMode && OpMinSize <= 8) 330 return LT.first * 3; // pmullw/zext 331 if (OpMinSize <= 15) 332 return LT.first * 5; // pmullw/pmulhw/pshuf 333 if (!SignedMode && OpMinSize <= 16) 334 return LT.first * 5; // pmullw/pmulhw/pshuf 335 } 336 337 if (const auto *Entry = CostTableLookup(SLMCostTable, ISD, 338 LT.second)) { 339 return LT.first * Entry->Cost; 340 } 341 } 342 343 static const CostTblEntry AVX512BWUniformConstCostTable[] = { 344 { ISD::SHL, MVT::v64i8, 2 }, // psllw + pand. 345 { ISD::SRL, MVT::v64i8, 2 }, // psrlw + pand. 346 { ISD::SRA, MVT::v64i8, 4 }, // psrlw, pand, pxor, psubb. 347 }; 348 349 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 350 ST->hasBWI()) { 351 if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD, 352 LT.second)) 353 return LT.first * Entry->Cost; 354 } 355 356 static const CostTblEntry AVX512UniformConstCostTable[] = { 357 { ISD::SRA, MVT::v2i64, 1 }, 358 { ISD::SRA, MVT::v4i64, 1 }, 359 { ISD::SRA, MVT::v8i64, 1 }, 360 361 { ISD::SHL, MVT::v64i8, 4 }, // psllw + pand. 362 { ISD::SRL, MVT::v64i8, 4 }, // psrlw + pand. 363 { ISD::SRA, MVT::v64i8, 8 }, // psrlw, pand, pxor, psubb. 364 365 { ISD::SDIV, MVT::v16i32, 6 }, // pmuludq sequence 366 { ISD::SREM, MVT::v16i32, 8 }, // pmuludq+mul+sub sequence 367 { ISD::UDIV, MVT::v16i32, 5 }, // pmuludq sequence 368 { ISD::UREM, MVT::v16i32, 7 }, // pmuludq+mul+sub sequence 369 }; 370 371 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 372 ST->hasAVX512()) { 373 if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD, 374 LT.second)) 375 return LT.first * Entry->Cost; 376 } 377 378 static const CostTblEntry AVX2UniformConstCostTable[] = { 379 { ISD::SHL, MVT::v32i8, 2 }, // psllw + pand. 380 { ISD::SRL, MVT::v32i8, 2 }, // psrlw + pand. 381 { ISD::SRA, MVT::v32i8, 4 }, // psrlw, pand, pxor, psubb. 382 383 { ISD::SRA, MVT::v4i64, 4 }, // 2 x psrad + shuffle. 384 385 { ISD::SDIV, MVT::v8i32, 6 }, // pmuludq sequence 386 { ISD::SREM, MVT::v8i32, 8 }, // pmuludq+mul+sub sequence 387 { ISD::UDIV, MVT::v8i32, 5 }, // pmuludq sequence 388 { ISD::UREM, MVT::v8i32, 7 }, // pmuludq+mul+sub sequence 389 }; 390 391 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 392 ST->hasAVX2()) { 393 if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD, 394 LT.second)) 395 return LT.first * Entry->Cost; 396 } 397 398 static const CostTblEntry SSE2UniformConstCostTable[] = { 399 { ISD::SHL, MVT::v16i8, 2 }, // psllw + pand. 400 { ISD::SRL, MVT::v16i8, 2 }, // psrlw + pand. 401 { ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb. 402 403 { ISD::SHL, MVT::v32i8, 4+2 }, // 2*(psllw + pand) + split. 404 { ISD::SRL, MVT::v32i8, 4+2 }, // 2*(psrlw + pand) + split. 405 { ISD::SRA, MVT::v32i8, 8+2 }, // 2*(psrlw, pand, pxor, psubb) + split. 406 407 { ISD::SDIV, MVT::v8i32, 12+2 }, // 2*pmuludq sequence + split. 408 { ISD::SREM, MVT::v8i32, 16+2 }, // 2*pmuludq+mul+sub sequence + split. 409 { ISD::SDIV, MVT::v4i32, 6 }, // pmuludq sequence 410 { ISD::SREM, MVT::v4i32, 8 }, // pmuludq+mul+sub sequence 411 { ISD::UDIV, MVT::v8i32, 10+2 }, // 2*pmuludq sequence + split. 412 { ISD::UREM, MVT::v8i32, 14+2 }, // 2*pmuludq+mul+sub sequence + split. 413 { ISD::UDIV, MVT::v4i32, 5 }, // pmuludq sequence 414 { ISD::UREM, MVT::v4i32, 7 }, // pmuludq+mul+sub sequence 415 }; 416 417 // XOP has faster vXi8 shifts. 418 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue && 419 ST->hasSSE2() && !ST->hasXOP()) { 420 if (const auto *Entry = 421 CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second)) 422 return LT.first * Entry->Cost; 423 } 424 425 static const CostTblEntry AVX512BWConstCostTable[] = { 426 { ISD::SDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence 427 { ISD::SREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 428 { ISD::UDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence 429 { ISD::UREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 430 { ISD::SDIV, MVT::v32i16, 6 }, // vpmulhw sequence 431 { ISD::SREM, MVT::v32i16, 8 }, // vpmulhw+mul+sub sequence 432 { ISD::UDIV, MVT::v32i16, 6 }, // vpmulhuw sequence 433 { ISD::UREM, MVT::v32i16, 8 }, // vpmulhuw+mul+sub sequence 434 }; 435 436 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 437 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 438 ST->hasBWI()) { 439 if (const auto *Entry = 440 CostTableLookup(AVX512BWConstCostTable, ISD, LT.second)) 441 return LT.first * Entry->Cost; 442 } 443 444 static const CostTblEntry AVX512ConstCostTable[] = { 445 { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence 446 { ISD::SREM, MVT::v16i32, 17 }, // vpmuldq+mul+sub sequence 447 { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence 448 { ISD::UREM, MVT::v16i32, 17 }, // vpmuludq+mul+sub sequence 449 { ISD::SDIV, MVT::v64i8, 28 }, // 4*ext+4*pmulhw sequence 450 { ISD::SREM, MVT::v64i8, 32 }, // 4*ext+4*pmulhw+mul+sub sequence 451 { ISD::UDIV, MVT::v64i8, 28 }, // 4*ext+4*pmulhw sequence 452 { ISD::UREM, MVT::v64i8, 32 }, // 4*ext+4*pmulhw+mul+sub sequence 453 { ISD::SDIV, MVT::v32i16, 12 }, // 2*vpmulhw sequence 454 { ISD::SREM, MVT::v32i16, 16 }, // 2*vpmulhw+mul+sub sequence 455 { ISD::UDIV, MVT::v32i16, 12 }, // 2*vpmulhuw sequence 456 { ISD::UREM, MVT::v32i16, 16 }, // 2*vpmulhuw+mul+sub sequence 457 }; 458 459 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 460 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 461 ST->hasAVX512()) { 462 if (const auto *Entry = 463 CostTableLookup(AVX512ConstCostTable, ISD, LT.second)) 464 return LT.first * Entry->Cost; 465 } 466 467 static const CostTblEntry AVX2ConstCostTable[] = { 468 { ISD::SDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence 469 { ISD::SREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 470 { ISD::UDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence 471 { ISD::UREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 472 { ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence 473 { ISD::SREM, MVT::v16i16, 8 }, // vpmulhw+mul+sub sequence 474 { ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence 475 { ISD::UREM, MVT::v16i16, 8 }, // vpmulhuw+mul+sub sequence 476 { ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence 477 { ISD::SREM, MVT::v8i32, 19 }, // vpmuldq+mul+sub sequence 478 { ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence 479 { ISD::UREM, MVT::v8i32, 19 }, // vpmuludq+mul+sub sequence 480 }; 481 482 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 483 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 484 ST->hasAVX2()) { 485 if (const auto *Entry = CostTableLookup(AVX2ConstCostTable, ISD, LT.second)) 486 return LT.first * Entry->Cost; 487 } 488 489 static const CostTblEntry SSE2ConstCostTable[] = { 490 { ISD::SDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split. 491 { ISD::SREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split. 492 { ISD::SDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence 493 { ISD::SREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 494 { ISD::UDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split. 495 { ISD::UREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split. 496 { ISD::UDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence 497 { ISD::UREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence 498 { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split. 499 { ISD::SREM, MVT::v16i16, 16+2 }, // 2*pmulhw+mul+sub sequence + split. 500 { ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence 501 { ISD::SREM, MVT::v8i16, 8 }, // pmulhw+mul+sub sequence 502 { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split. 503 { ISD::UREM, MVT::v16i16, 16+2 }, // 2*pmulhuw+mul+sub sequence + split. 504 { ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence 505 { ISD::UREM, MVT::v8i16, 8 }, // pmulhuw+mul+sub sequence 506 { ISD::SDIV, MVT::v8i32, 38+2 }, // 2*pmuludq sequence + split. 507 { ISD::SREM, MVT::v8i32, 48+2 }, // 2*pmuludq+mul+sub sequence + split. 508 { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence 509 { ISD::SREM, MVT::v4i32, 24 }, // pmuludq+mul+sub sequence 510 { ISD::UDIV, MVT::v8i32, 30+2 }, // 2*pmuludq sequence + split. 511 { ISD::UREM, MVT::v8i32, 40+2 }, // 2*pmuludq+mul+sub sequence + split. 512 { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence 513 { ISD::UREM, MVT::v4i32, 20 }, // pmuludq+mul+sub sequence 514 }; 515 516 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue || 517 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) && 518 ST->hasSSE2()) { 519 // pmuldq sequence. 520 if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX()) 521 return LT.first * 32; 522 if (ISD == ISD::SREM && LT.second == MVT::v8i32 && ST->hasAVX()) 523 return LT.first * 38; 524 if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41()) 525 return LT.first * 15; 526 if (ISD == ISD::SREM && LT.second == MVT::v4i32 && ST->hasSSE41()) 527 return LT.first * 20; 528 529 if (const auto *Entry = CostTableLookup(SSE2ConstCostTable, ISD, LT.second)) 530 return LT.first * Entry->Cost; 531 } 532 533 static const CostTblEntry AVX512BWShiftCostTable[] = { 534 { ISD::SHL, MVT::v16i8, 4 }, // extend/vpsllvw/pack sequence. 535 { ISD::SRL, MVT::v16i8, 4 }, // extend/vpsrlvw/pack sequence. 536 { ISD::SRA, MVT::v16i8, 4 }, // extend/vpsravw/pack sequence. 537 { ISD::SHL, MVT::v32i8, 4 }, // extend/vpsllvw/pack sequence. 538 { ISD::SRL, MVT::v32i8, 4 }, // extend/vpsrlvw/pack sequence. 539 { ISD::SRA, MVT::v32i8, 6 }, // extend/vpsravw/pack sequence. 540 { ISD::SHL, MVT::v64i8, 6 }, // extend/vpsllvw/pack sequence. 541 { ISD::SRL, MVT::v64i8, 7 }, // extend/vpsrlvw/pack sequence. 542 { ISD::SRA, MVT::v64i8, 15 }, // extend/vpsravw/pack sequence. 543 544 { ISD::SHL, MVT::v8i16, 1 }, // vpsllvw 545 { ISD::SRL, MVT::v8i16, 1 }, // vpsrlvw 546 { ISD::SRA, MVT::v8i16, 1 }, // vpsravw 547 { ISD::SHL, MVT::v16i16, 1 }, // vpsllvw 548 { ISD::SRL, MVT::v16i16, 1 }, // vpsrlvw 549 { ISD::SRA, MVT::v16i16, 1 }, // vpsravw 550 { ISD::SHL, MVT::v32i16, 1 }, // vpsllvw 551 { ISD::SRL, MVT::v32i16, 1 }, // vpsrlvw 552 { ISD::SRA, MVT::v32i16, 1 }, // vpsravw 553 }; 554 555 if (ST->hasBWI()) 556 if (const auto *Entry = CostTableLookup(AVX512BWShiftCostTable, ISD, LT.second)) 557 return LT.first * Entry->Cost; 558 559 static const CostTblEntry AVX2UniformCostTable[] = { 560 // Uniform splats are cheaper for the following instructions. 561 { ISD::SHL, MVT::v16i16, 1 }, // psllw. 562 { ISD::SRL, MVT::v16i16, 1 }, // psrlw. 563 { ISD::SRA, MVT::v16i16, 1 }, // psraw. 564 { ISD::SHL, MVT::v32i16, 2 }, // 2*psllw. 565 { ISD::SRL, MVT::v32i16, 2 }, // 2*psrlw. 566 { ISD::SRA, MVT::v32i16, 2 }, // 2*psraw. 567 568 { ISD::SHL, MVT::v8i32, 1 }, // pslld 569 { ISD::SRL, MVT::v8i32, 1 }, // psrld 570 { ISD::SRA, MVT::v8i32, 1 }, // psrad 571 { ISD::SHL, MVT::v4i64, 1 }, // psllq 572 { ISD::SRL, MVT::v4i64, 1 }, // psrlq 573 }; 574 575 if (ST->hasAVX2() && 576 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 577 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 578 if (const auto *Entry = 579 CostTableLookup(AVX2UniformCostTable, ISD, LT.second)) 580 return LT.first * Entry->Cost; 581 } 582 583 static const CostTblEntry SSE2UniformCostTable[] = { 584 // Uniform splats are cheaper for the following instructions. 585 { ISD::SHL, MVT::v8i16, 1 }, // psllw. 586 { ISD::SHL, MVT::v4i32, 1 }, // pslld 587 { ISD::SHL, MVT::v2i64, 1 }, // psllq. 588 589 { ISD::SRL, MVT::v8i16, 1 }, // psrlw. 590 { ISD::SRL, MVT::v4i32, 1 }, // psrld. 591 { ISD::SRL, MVT::v2i64, 1 }, // psrlq. 592 593 { ISD::SRA, MVT::v8i16, 1 }, // psraw. 594 { ISD::SRA, MVT::v4i32, 1 }, // psrad. 595 }; 596 597 if (ST->hasSSE2() && 598 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 599 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 600 if (const auto *Entry = 601 CostTableLookup(SSE2UniformCostTable, ISD, LT.second)) 602 return LT.first * Entry->Cost; 603 } 604 605 static const CostTblEntry AVX512DQCostTable[] = { 606 { ISD::MUL, MVT::v2i64, 2 }, // pmullq 607 { ISD::MUL, MVT::v4i64, 2 }, // pmullq 608 { ISD::MUL, MVT::v8i64, 2 } // pmullq 609 }; 610 611 // Look for AVX512DQ lowering tricks for custom cases. 612 if (ST->hasDQI()) 613 if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second)) 614 return LT.first * Entry->Cost; 615 616 static const CostTblEntry AVX512BWCostTable[] = { 617 { ISD::SHL, MVT::v64i8, 11 }, // vpblendvb sequence. 618 { ISD::SRL, MVT::v64i8, 11 }, // vpblendvb sequence. 619 { ISD::SRA, MVT::v64i8, 24 }, // vpblendvb sequence. 620 }; 621 622 // Look for AVX512BW lowering tricks for custom cases. 623 if (ST->hasBWI()) 624 if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second)) 625 return LT.first * Entry->Cost; 626 627 static const CostTblEntry AVX512CostTable[] = { 628 { ISD::SHL, MVT::v4i32, 1 }, 629 { ISD::SRL, MVT::v4i32, 1 }, 630 { ISD::SRA, MVT::v4i32, 1 }, 631 { ISD::SHL, MVT::v8i32, 1 }, 632 { ISD::SRL, MVT::v8i32, 1 }, 633 { ISD::SRA, MVT::v8i32, 1 }, 634 { ISD::SHL, MVT::v16i32, 1 }, 635 { ISD::SRL, MVT::v16i32, 1 }, 636 { ISD::SRA, MVT::v16i32, 1 }, 637 638 { ISD::SHL, MVT::v2i64, 1 }, 639 { ISD::SRL, MVT::v2i64, 1 }, 640 { ISD::SHL, MVT::v4i64, 1 }, 641 { ISD::SRL, MVT::v4i64, 1 }, 642 { ISD::SHL, MVT::v8i64, 1 }, 643 { ISD::SRL, MVT::v8i64, 1 }, 644 645 { ISD::SRA, MVT::v2i64, 1 }, 646 { ISD::SRA, MVT::v4i64, 1 }, 647 { ISD::SRA, MVT::v8i64, 1 }, 648 649 { ISD::MUL, MVT::v16i32, 1 }, // pmulld (Skylake from agner.org) 650 { ISD::MUL, MVT::v8i32, 1 }, // pmulld (Skylake from agner.org) 651 { ISD::MUL, MVT::v4i32, 1 }, // pmulld (Skylake from agner.org) 652 { ISD::MUL, MVT::v8i64, 6 }, // 3*pmuludq/3*shift/2*add 653 654 { ISD::FNEG, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 655 { ISD::FADD, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 656 { ISD::FSUB, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 657 { ISD::FMUL, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/ 658 { ISD::FDIV, MVT::f64, 4 }, // Skylake from http://www.agner.org/ 659 { ISD::FDIV, MVT::v2f64, 4 }, // Skylake from http://www.agner.org/ 660 { ISD::FDIV, MVT::v4f64, 8 }, // Skylake from http://www.agner.org/ 661 { ISD::FDIV, MVT::v8f64, 16 }, // Skylake from http://www.agner.org/ 662 663 { ISD::FNEG, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 664 { ISD::FADD, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 665 { ISD::FSUB, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 666 { ISD::FMUL, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/ 667 { ISD::FDIV, MVT::f32, 3 }, // Skylake from http://www.agner.org/ 668 { ISD::FDIV, MVT::v4f32, 3 }, // Skylake from http://www.agner.org/ 669 { ISD::FDIV, MVT::v8f32, 5 }, // Skylake from http://www.agner.org/ 670 { ISD::FDIV, MVT::v16f32, 10 }, // Skylake from http://www.agner.org/ 671 }; 672 673 if (ST->hasAVX512()) 674 if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second)) 675 return LT.first * Entry->Cost; 676 677 static const CostTblEntry AVX2ShiftCostTable[] = { 678 // Shifts on vXi64/vXi32 on AVX2 is legal even though we declare to 679 // customize them to detect the cases where shift amount is a scalar one. 680 { ISD::SHL, MVT::v4i32, 2 }, // vpsllvd (Haswell from agner.org) 681 { ISD::SRL, MVT::v4i32, 2 }, // vpsrlvd (Haswell from agner.org) 682 { ISD::SRA, MVT::v4i32, 2 }, // vpsravd (Haswell from agner.org) 683 { ISD::SHL, MVT::v8i32, 2 }, // vpsllvd (Haswell from agner.org) 684 { ISD::SRL, MVT::v8i32, 2 }, // vpsrlvd (Haswell from agner.org) 685 { ISD::SRA, MVT::v8i32, 2 }, // vpsravd (Haswell from agner.org) 686 { ISD::SHL, MVT::v2i64, 1 }, // vpsllvq (Haswell from agner.org) 687 { ISD::SRL, MVT::v2i64, 1 }, // vpsrlvq (Haswell from agner.org) 688 { ISD::SHL, MVT::v4i64, 1 }, // vpsllvq (Haswell from agner.org) 689 { ISD::SRL, MVT::v4i64, 1 }, // vpsrlvq (Haswell from agner.org) 690 }; 691 692 if (ST->hasAVX512()) { 693 if (ISD == ISD::SHL && LT.second == MVT::v32i16 && 694 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 695 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 696 // On AVX512, a packed v32i16 shift left by a constant build_vector 697 // is lowered into a vector multiply (vpmullw). 698 return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, 699 Op1Info, Op2Info, 700 TargetTransformInfo::OP_None, 701 TargetTransformInfo::OP_None); 702 } 703 704 // Look for AVX2 lowering tricks (XOP is always better at v4i32 shifts). 705 if (ST->hasAVX2() && !(ST->hasXOP() && LT.second == MVT::v4i32)) { 706 if (ISD == ISD::SHL && LT.second == MVT::v16i16 && 707 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 708 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 709 // On AVX2, a packed v16i16 shift left by a constant build_vector 710 // is lowered into a vector multiply (vpmullw). 711 return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, 712 Op1Info, Op2Info, 713 TargetTransformInfo::OP_None, 714 TargetTransformInfo::OP_None); 715 716 if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second)) 717 return LT.first * Entry->Cost; 718 } 719 720 static const CostTblEntry XOPShiftCostTable[] = { 721 // 128bit shifts take 1cy, but right shifts require negation beforehand. 722 { ISD::SHL, MVT::v16i8, 1 }, 723 { ISD::SRL, MVT::v16i8, 2 }, 724 { ISD::SRA, MVT::v16i8, 2 }, 725 { ISD::SHL, MVT::v8i16, 1 }, 726 { ISD::SRL, MVT::v8i16, 2 }, 727 { ISD::SRA, MVT::v8i16, 2 }, 728 { ISD::SHL, MVT::v4i32, 1 }, 729 { ISD::SRL, MVT::v4i32, 2 }, 730 { ISD::SRA, MVT::v4i32, 2 }, 731 { ISD::SHL, MVT::v2i64, 1 }, 732 { ISD::SRL, MVT::v2i64, 2 }, 733 { ISD::SRA, MVT::v2i64, 2 }, 734 // 256bit shifts require splitting if AVX2 didn't catch them above. 735 { ISD::SHL, MVT::v32i8, 2+2 }, 736 { ISD::SRL, MVT::v32i8, 4+2 }, 737 { ISD::SRA, MVT::v32i8, 4+2 }, 738 { ISD::SHL, MVT::v16i16, 2+2 }, 739 { ISD::SRL, MVT::v16i16, 4+2 }, 740 { ISD::SRA, MVT::v16i16, 4+2 }, 741 { ISD::SHL, MVT::v8i32, 2+2 }, 742 { ISD::SRL, MVT::v8i32, 4+2 }, 743 { ISD::SRA, MVT::v8i32, 4+2 }, 744 { ISD::SHL, MVT::v4i64, 2+2 }, 745 { ISD::SRL, MVT::v4i64, 4+2 }, 746 { ISD::SRA, MVT::v4i64, 4+2 }, 747 }; 748 749 // Look for XOP lowering tricks. 750 if (ST->hasXOP()) { 751 // If the right shift is constant then we'll fold the negation so 752 // it's as cheap as a left shift. 753 int ShiftISD = ISD; 754 if ((ShiftISD == ISD::SRL || ShiftISD == ISD::SRA) && 755 (Op2Info == TargetTransformInfo::OK_UniformConstantValue || 756 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)) 757 ShiftISD = ISD::SHL; 758 if (const auto *Entry = 759 CostTableLookup(XOPShiftCostTable, ShiftISD, LT.second)) 760 return LT.first * Entry->Cost; 761 } 762 763 static const CostTblEntry SSE2UniformShiftCostTable[] = { 764 // Uniform splats are cheaper for the following instructions. 765 { ISD::SHL, MVT::v16i16, 2+2 }, // 2*psllw + split. 766 { ISD::SHL, MVT::v8i32, 2+2 }, // 2*pslld + split. 767 { ISD::SHL, MVT::v4i64, 2+2 }, // 2*psllq + split. 768 769 { ISD::SRL, MVT::v16i16, 2+2 }, // 2*psrlw + split. 770 { ISD::SRL, MVT::v8i32, 2+2 }, // 2*psrld + split. 771 { ISD::SRL, MVT::v4i64, 2+2 }, // 2*psrlq + split. 772 773 { ISD::SRA, MVT::v16i16, 2+2 }, // 2*psraw + split. 774 { ISD::SRA, MVT::v8i32, 2+2 }, // 2*psrad + split. 775 { ISD::SRA, MVT::v2i64, 4 }, // 2*psrad + shuffle. 776 { ISD::SRA, MVT::v4i64, 8+2 }, // 2*(2*psrad + shuffle) + split. 777 }; 778 779 if (ST->hasSSE2() && 780 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) || 781 (Op2Info == TargetTransformInfo::OK_UniformValue))) { 782 783 // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table. 784 if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2()) 785 return LT.first * 4; // 2*psrad + shuffle. 786 787 if (const auto *Entry = 788 CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second)) 789 return LT.first * Entry->Cost; 790 } 791 792 if (ISD == ISD::SHL && 793 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) { 794 MVT VT = LT.second; 795 // Vector shift left by non uniform constant can be lowered 796 // into vector multiply. 797 if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) || 798 ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX())) 799 ISD = ISD::MUL; 800 } 801 802 static const CostTblEntry AVX2CostTable[] = { 803 { ISD::SHL, MVT::v16i8, 6 }, // vpblendvb sequence. 804 { ISD::SHL, MVT::v32i8, 6 }, // vpblendvb sequence. 805 { ISD::SHL, MVT::v64i8, 12 }, // 2*vpblendvb sequence. 806 { ISD::SHL, MVT::v8i16, 5 }, // extend/vpsrlvd/pack sequence. 807 { ISD::SHL, MVT::v16i16, 7 }, // extend/vpsrlvd/pack sequence. 808 { ISD::SHL, MVT::v32i16, 14 }, // 2*extend/vpsrlvd/pack sequence. 809 810 { ISD::SRL, MVT::v16i8, 6 }, // vpblendvb sequence. 811 { ISD::SRL, MVT::v32i8, 6 }, // vpblendvb sequence. 812 { ISD::SRL, MVT::v64i8, 12 }, // 2*vpblendvb sequence. 813 { ISD::SRL, MVT::v8i16, 5 }, // extend/vpsrlvd/pack sequence. 814 { ISD::SRL, MVT::v16i16, 7 }, // extend/vpsrlvd/pack sequence. 815 { ISD::SRL, MVT::v32i16, 14 }, // 2*extend/vpsrlvd/pack sequence. 816 817 { ISD::SRA, MVT::v16i8, 17 }, // vpblendvb sequence. 818 { ISD::SRA, MVT::v32i8, 17 }, // vpblendvb sequence. 819 { ISD::SRA, MVT::v64i8, 34 }, // 2*vpblendvb sequence. 820 { ISD::SRA, MVT::v8i16, 5 }, // extend/vpsravd/pack sequence. 821 { ISD::SRA, MVT::v16i16, 7 }, // extend/vpsravd/pack sequence. 822 { ISD::SRA, MVT::v32i16, 14 }, // 2*extend/vpsravd/pack sequence. 823 { ISD::SRA, MVT::v2i64, 2 }, // srl/xor/sub sequence. 824 { ISD::SRA, MVT::v4i64, 2 }, // srl/xor/sub sequence. 825 826 { ISD::SUB, MVT::v32i8, 1 }, // psubb 827 { ISD::ADD, MVT::v32i8, 1 }, // paddb 828 { ISD::SUB, MVT::v16i16, 1 }, // psubw 829 { ISD::ADD, MVT::v16i16, 1 }, // paddw 830 { ISD::SUB, MVT::v8i32, 1 }, // psubd 831 { ISD::ADD, MVT::v8i32, 1 }, // paddd 832 { ISD::SUB, MVT::v4i64, 1 }, // psubq 833 { ISD::ADD, MVT::v4i64, 1 }, // paddq 834 835 { ISD::MUL, MVT::v16i16, 1 }, // pmullw 836 { ISD::MUL, MVT::v8i32, 2 }, // pmulld (Haswell from agner.org) 837 { ISD::MUL, MVT::v4i64, 6 }, // 3*pmuludq/3*shift/2*add 838 839 { ISD::FNEG, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 840 { ISD::FNEG, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 841 { ISD::FADD, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 842 { ISD::FADD, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 843 { ISD::FSUB, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 844 { ISD::FSUB, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 845 { ISD::FMUL, MVT::f64, 1 }, // Haswell from http://www.agner.org/ 846 { ISD::FMUL, MVT::v2f64, 1 }, // Haswell from http://www.agner.org/ 847 { ISD::FMUL, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/ 848 { ISD::FMUL, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/ 849 850 { ISD::FDIV, MVT::f32, 7 }, // Haswell from http://www.agner.org/ 851 { ISD::FDIV, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/ 852 { ISD::FDIV, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/ 853 { ISD::FDIV, MVT::f64, 14 }, // Haswell from http://www.agner.org/ 854 { ISD::FDIV, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/ 855 { ISD::FDIV, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/ 856 }; 857 858 // Look for AVX2 lowering tricks for custom cases. 859 if (ST->hasAVX2()) 860 if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second)) 861 return LT.first * Entry->Cost; 862 863 static const CostTblEntry AVX1CostTable[] = { 864 // We don't have to scalarize unsupported ops. We can issue two half-sized 865 // operations and we only need to extract the upper YMM half. 866 // Two ops + 1 extract + 1 insert = 4. 867 { ISD::MUL, MVT::v16i16, 4 }, 868 { ISD::MUL, MVT::v8i32, 5 }, // BTVER2 from http://www.agner.org/ 869 { ISD::MUL, MVT::v4i64, 12 }, 870 871 { ISD::SUB, MVT::v32i8, 4 }, 872 { ISD::ADD, MVT::v32i8, 4 }, 873 { ISD::SUB, MVT::v16i16, 4 }, 874 { ISD::ADD, MVT::v16i16, 4 }, 875 { ISD::SUB, MVT::v8i32, 4 }, 876 { ISD::ADD, MVT::v8i32, 4 }, 877 { ISD::SUB, MVT::v4i64, 4 }, 878 { ISD::ADD, MVT::v4i64, 4 }, 879 880 { ISD::SHL, MVT::v32i8, 22 }, // pblendvb sequence + split. 881 { ISD::SHL, MVT::v8i16, 6 }, // pblendvb sequence. 882 { ISD::SHL, MVT::v16i16, 13 }, // pblendvb sequence + split. 883 { ISD::SHL, MVT::v4i32, 3 }, // pslld/paddd/cvttps2dq/pmulld 884 { ISD::SHL, MVT::v8i32, 9 }, // pslld/paddd/cvttps2dq/pmulld + split 885 { ISD::SHL, MVT::v2i64, 2 }, // Shift each lane + blend. 886 { ISD::SHL, MVT::v4i64, 6 }, // Shift each lane + blend + split. 887 888 { ISD::SRL, MVT::v32i8, 23 }, // pblendvb sequence + split. 889 { ISD::SRL, MVT::v16i16, 28 }, // pblendvb sequence + split. 890 { ISD::SRL, MVT::v4i32, 6 }, // Shift each lane + blend. 891 { ISD::SRL, MVT::v8i32, 14 }, // Shift each lane + blend + split. 892 { ISD::SRL, MVT::v2i64, 2 }, // Shift each lane + blend. 893 { ISD::SRL, MVT::v4i64, 6 }, // Shift each lane + blend + split. 894 895 { ISD::SRA, MVT::v32i8, 44 }, // pblendvb sequence + split. 896 { ISD::SRA, MVT::v16i16, 28 }, // pblendvb sequence + split. 897 { ISD::SRA, MVT::v4i32, 6 }, // Shift each lane + blend. 898 { ISD::SRA, MVT::v8i32, 14 }, // Shift each lane + blend + split. 899 { ISD::SRA, MVT::v2i64, 5 }, // Shift each lane + blend. 900 { ISD::SRA, MVT::v4i64, 12 }, // Shift each lane + blend + split. 901 902 { ISD::FNEG, MVT::v4f64, 2 }, // BTVER2 from http://www.agner.org/ 903 { ISD::FNEG, MVT::v8f32, 2 }, // BTVER2 from http://www.agner.org/ 904 905 { ISD::FMUL, MVT::f64, 2 }, // BTVER2 from http://www.agner.org/ 906 { ISD::FMUL, MVT::v2f64, 2 }, // BTVER2 from http://www.agner.org/ 907 { ISD::FMUL, MVT::v4f64, 4 }, // BTVER2 from http://www.agner.org/ 908 909 { ISD::FDIV, MVT::f32, 14 }, // SNB from http://www.agner.org/ 910 { ISD::FDIV, MVT::v4f32, 14 }, // SNB from http://www.agner.org/ 911 { ISD::FDIV, MVT::v8f32, 28 }, // SNB from http://www.agner.org/ 912 { ISD::FDIV, MVT::f64, 22 }, // SNB from http://www.agner.org/ 913 { ISD::FDIV, MVT::v2f64, 22 }, // SNB from http://www.agner.org/ 914 { ISD::FDIV, MVT::v4f64, 44 }, // SNB from http://www.agner.org/ 915 }; 916 917 if (ST->hasAVX()) 918 if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second)) 919 return LT.first * Entry->Cost; 920 921 static const CostTblEntry SSE42CostTable[] = { 922 { ISD::FADD, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 923 { ISD::FADD, MVT::f32, 1 }, // Nehalem from http://www.agner.org/ 924 { ISD::FADD, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 925 { ISD::FADD, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 926 927 { ISD::FSUB, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 928 { ISD::FSUB, MVT::f32 , 1 }, // Nehalem from http://www.agner.org/ 929 { ISD::FSUB, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 930 { ISD::FSUB, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 931 932 { ISD::FMUL, MVT::f64, 1 }, // Nehalem from http://www.agner.org/ 933 { ISD::FMUL, MVT::f32, 1 }, // Nehalem from http://www.agner.org/ 934 { ISD::FMUL, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/ 935 { ISD::FMUL, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/ 936 937 { ISD::FDIV, MVT::f32, 14 }, // Nehalem from http://www.agner.org/ 938 { ISD::FDIV, MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/ 939 { ISD::FDIV, MVT::f64, 22 }, // Nehalem from http://www.agner.org/ 940 { ISD::FDIV, MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/ 941 942 { ISD::MUL, MVT::v2i64, 6 } // 3*pmuludq/3*shift/2*add 943 }; 944 945 if (ST->hasSSE42()) 946 if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second)) 947 return LT.first * Entry->Cost; 948 949 static const CostTblEntry SSE41CostTable[] = { 950 { ISD::SHL, MVT::v16i8, 10 }, // pblendvb sequence. 951 { ISD::SHL, MVT::v8i16, 11 }, // pblendvb sequence. 952 { ISD::SHL, MVT::v4i32, 4 }, // pslld/paddd/cvttps2dq/pmulld 953 954 { ISD::SRL, MVT::v16i8, 11 }, // pblendvb sequence. 955 { ISD::SRL, MVT::v8i16, 13 }, // pblendvb sequence. 956 { ISD::SRL, MVT::v4i32, 16 }, // Shift each lane + blend. 957 958 { ISD::SRA, MVT::v16i8, 21 }, // pblendvb sequence. 959 { ISD::SRA, MVT::v8i16, 13 }, // pblendvb sequence. 960 961 { ISD::MUL, MVT::v4i32, 2 } // pmulld (Nehalem from agner.org) 962 }; 963 964 if (ST->hasSSE41()) 965 if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second)) 966 return LT.first * Entry->Cost; 967 968 static const CostTblEntry SSE2CostTable[] = { 969 // We don't correctly identify costs of casts because they are marked as 970 // custom. 971 { ISD::SHL, MVT::v16i8, 13 }, // cmpgtb sequence. 972 { ISD::SHL, MVT::v8i16, 25 }, // cmpgtw sequence. 973 { ISD::SHL, MVT::v4i32, 16 }, // pslld/paddd/cvttps2dq/pmuludq. 974 { ISD::SHL, MVT::v2i64, 4 }, // splat+shuffle sequence. 975 976 { ISD::SRL, MVT::v16i8, 14 }, // cmpgtb sequence. 977 { ISD::SRL, MVT::v8i16, 16 }, // cmpgtw sequence. 978 { ISD::SRL, MVT::v4i32, 12 }, // Shift each lane + blend. 979 { ISD::SRL, MVT::v2i64, 4 }, // splat+shuffle sequence. 980 981 { ISD::SRA, MVT::v16i8, 27 }, // unpacked cmpgtb sequence. 982 { ISD::SRA, MVT::v8i16, 16 }, // cmpgtw sequence. 983 { ISD::SRA, MVT::v4i32, 12 }, // Shift each lane + blend. 984 { ISD::SRA, MVT::v2i64, 8 }, // srl/xor/sub splat+shuffle sequence. 985 986 { ISD::MUL, MVT::v8i16, 1 }, // pmullw 987 { ISD::MUL, MVT::v4i32, 6 }, // 3*pmuludq/4*shuffle 988 { ISD::MUL, MVT::v2i64, 8 }, // 3*pmuludq/3*shift/2*add 989 990 { ISD::FDIV, MVT::f32, 23 }, // Pentium IV from http://www.agner.org/ 991 { ISD::FDIV, MVT::v4f32, 39 }, // Pentium IV from http://www.agner.org/ 992 { ISD::FDIV, MVT::f64, 38 }, // Pentium IV from http://www.agner.org/ 993 { ISD::FDIV, MVT::v2f64, 69 }, // Pentium IV from http://www.agner.org/ 994 995 { ISD::FNEG, MVT::f32, 1 }, // Pentium IV from http://www.agner.org/ 996 { ISD::FNEG, MVT::f64, 1 }, // Pentium IV from http://www.agner.org/ 997 { ISD::FNEG, MVT::v4f32, 1 }, // Pentium IV from http://www.agner.org/ 998 { ISD::FNEG, MVT::v2f64, 1 }, // Pentium IV from http://www.agner.org/ 999 1000 { ISD::FADD, MVT::f32, 2 }, // Pentium IV from http://www.agner.org/ 1001 { ISD::FADD, MVT::f64, 2 }, // Pentium IV from http://www.agner.org/ 1002 1003 { ISD::FSUB, MVT::f32, 2 }, // Pentium IV from http://www.agner.org/ 1004 { ISD::FSUB, MVT::f64, 2 }, // Pentium IV from http://www.agner.org/ 1005 }; 1006 1007 if (ST->hasSSE2()) 1008 if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second)) 1009 return LT.first * Entry->Cost; 1010 1011 static const CostTblEntry SSE1CostTable[] = { 1012 { ISD::FDIV, MVT::f32, 17 }, // Pentium III from http://www.agner.org/ 1013 { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/ 1014 1015 { ISD::FNEG, MVT::f32, 2 }, // Pentium III from http://www.agner.org/ 1016 { ISD::FNEG, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/ 1017 1018 { ISD::FADD, MVT::f32, 1 }, // Pentium III from http://www.agner.org/ 1019 { ISD::FADD, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/ 1020 1021 { ISD::FSUB, MVT::f32, 1 }, // Pentium III from http://www.agner.org/ 1022 { ISD::FSUB, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/ 1023 }; 1024 1025 if (ST->hasSSE1()) 1026 if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second)) 1027 return LT.first * Entry->Cost; 1028 1029 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 1030 { ISD::ADD, MVT::i64, 1 }, // Core (Merom) from http://www.agner.org/ 1031 { ISD::SUB, MVT::i64, 1 }, // Core (Merom) from http://www.agner.org/ 1032 { ISD::MUL, MVT::i64, 2 }, // Nehalem from http://www.agner.org/ 1033 }; 1034 1035 if (ST->is64Bit()) 1036 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, LT.second)) 1037 return LT.first * Entry->Cost; 1038 1039 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 1040 { ISD::ADD, MVT::i8, 1 }, // Pentium III from http://www.agner.org/ 1041 { ISD::ADD, MVT::i16, 1 }, // Pentium III from http://www.agner.org/ 1042 { ISD::ADD, MVT::i32, 1 }, // Pentium III from http://www.agner.org/ 1043 1044 { ISD::SUB, MVT::i8, 1 }, // Pentium III from http://www.agner.org/ 1045 { ISD::SUB, MVT::i16, 1 }, // Pentium III from http://www.agner.org/ 1046 { ISD::SUB, MVT::i32, 1 }, // Pentium III from http://www.agner.org/ 1047 }; 1048 1049 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, LT.second)) 1050 return LT.first * Entry->Cost; 1051 1052 // It is not a good idea to vectorize division. We have to scalarize it and 1053 // in the process we will often end up having to spilling regular 1054 // registers. The overhead of division is going to dominate most kernels 1055 // anyways so try hard to prevent vectorization of division - it is 1056 // generally a bad idea. Assume somewhat arbitrarily that we have to be able 1057 // to hide "20 cycles" for each lane. 1058 if (LT.second.isVector() && (ISD == ISD::SDIV || ISD == ISD::SREM || 1059 ISD == ISD::UDIV || ISD == ISD::UREM)) { 1060 InstructionCost ScalarCost = getArithmeticInstrCost( 1061 Opcode, Ty->getScalarType(), CostKind, Op1Info, Op2Info, 1062 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 1063 return 20 * LT.first * LT.second.getVectorNumElements() * ScalarCost; 1064 } 1065 1066 // Fallback to the default implementation. 1067 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info); 1068 } 1069 1070 InstructionCost X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind, 1071 VectorType *BaseTp, 1072 ArrayRef<int> Mask, int Index, 1073 VectorType *SubTp) { 1074 // 64-bit packed float vectors (v2f32) are widened to type v4f32. 1075 // 64-bit packed integer vectors (v2i32) are widened to type v4i32. 1076 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, BaseTp); 1077 1078 Kind = improveShuffleKindFromMask(Kind, Mask); 1079 // Treat Transpose as 2-op shuffles - there's no difference in lowering. 1080 if (Kind == TTI::SK_Transpose) 1081 Kind = TTI::SK_PermuteTwoSrc; 1082 1083 // For Broadcasts we are splatting the first element from the first input 1084 // register, so only need to reference that input and all the output 1085 // registers are the same. 1086 if (Kind == TTI::SK_Broadcast) 1087 LT.first = 1; 1088 1089 // Subvector extractions are free if they start at the beginning of a 1090 // vector and cheap if the subvectors are aligned. 1091 if (Kind == TTI::SK_ExtractSubvector && LT.second.isVector()) { 1092 int NumElts = LT.second.getVectorNumElements(); 1093 if ((Index % NumElts) == 0) 1094 return 0; 1095 std::pair<InstructionCost, MVT> SubLT = 1096 TLI->getTypeLegalizationCost(DL, SubTp); 1097 if (SubLT.second.isVector()) { 1098 int NumSubElts = SubLT.second.getVectorNumElements(); 1099 if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0) 1100 return SubLT.first; 1101 // Handle some cases for widening legalization. For now we only handle 1102 // cases where the original subvector was naturally aligned and evenly 1103 // fit in its legalized subvector type. 1104 // FIXME: Remove some of the alignment restrictions. 1105 // FIXME: We can use permq for 64-bit or larger extracts from 256-bit 1106 // vectors. 1107 int OrigSubElts = cast<FixedVectorType>(SubTp)->getNumElements(); 1108 if (NumSubElts > OrigSubElts && (Index % OrigSubElts) == 0 && 1109 (NumSubElts % OrigSubElts) == 0 && 1110 LT.second.getVectorElementType() == 1111 SubLT.second.getVectorElementType() && 1112 LT.second.getVectorElementType().getSizeInBits() == 1113 BaseTp->getElementType()->getPrimitiveSizeInBits()) { 1114 assert(NumElts >= NumSubElts && NumElts > OrigSubElts && 1115 "Unexpected number of elements!"); 1116 auto *VecTy = FixedVectorType::get(BaseTp->getElementType(), 1117 LT.second.getVectorNumElements()); 1118 auto *SubTy = FixedVectorType::get(BaseTp->getElementType(), 1119 SubLT.second.getVectorNumElements()); 1120 int ExtractIndex = alignDown((Index % NumElts), NumSubElts); 1121 InstructionCost ExtractCost = getShuffleCost( 1122 TTI::SK_ExtractSubvector, VecTy, None, ExtractIndex, SubTy); 1123 1124 // If the original size is 32-bits or more, we can use pshufd. Otherwise 1125 // if we have SSSE3 we can use pshufb. 1126 if (SubTp->getPrimitiveSizeInBits() >= 32 || ST->hasSSSE3()) 1127 return ExtractCost + 1; // pshufd or pshufb 1128 1129 assert(SubTp->getPrimitiveSizeInBits() == 16 && 1130 "Unexpected vector size"); 1131 1132 return ExtractCost + 2; // worst case pshufhw + pshufd 1133 } 1134 } 1135 } 1136 1137 // Subvector insertions are cheap if the subvectors are aligned. 1138 // Note that in general, the insertion starting at the beginning of a vector 1139 // isn't free, because we need to preserve the rest of the wide vector. 1140 if (Kind == TTI::SK_InsertSubvector && LT.second.isVector()) { 1141 int NumElts = LT.second.getVectorNumElements(); 1142 std::pair<InstructionCost, MVT> SubLT = 1143 TLI->getTypeLegalizationCost(DL, SubTp); 1144 if (SubLT.second.isVector()) { 1145 int NumSubElts = SubLT.second.getVectorNumElements(); 1146 if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0) 1147 return SubLT.first; 1148 } 1149 1150 // If the insertion isn't aligned, treat it like a 2-op shuffle. 1151 Kind = TTI::SK_PermuteTwoSrc; 1152 } 1153 1154 // Handle some common (illegal) sub-vector types as they are often very cheap 1155 // to shuffle even on targets without PSHUFB. 1156 EVT VT = TLI->getValueType(DL, BaseTp); 1157 if (VT.isSimple() && VT.isVector() && VT.getSizeInBits() < 128 && 1158 !ST->hasSSSE3()) { 1159 static const CostTblEntry SSE2SubVectorShuffleTbl[] = { 1160 {TTI::SK_Broadcast, MVT::v4i16, 1}, // pshuflw 1161 {TTI::SK_Broadcast, MVT::v2i16, 1}, // pshuflw 1162 {TTI::SK_Broadcast, MVT::v8i8, 2}, // punpck/pshuflw 1163 {TTI::SK_Broadcast, MVT::v4i8, 2}, // punpck/pshuflw 1164 {TTI::SK_Broadcast, MVT::v2i8, 1}, // punpck 1165 1166 {TTI::SK_Reverse, MVT::v4i16, 1}, // pshuflw 1167 {TTI::SK_Reverse, MVT::v2i16, 1}, // pshuflw 1168 {TTI::SK_Reverse, MVT::v4i8, 3}, // punpck/pshuflw/packus 1169 {TTI::SK_Reverse, MVT::v2i8, 1}, // punpck 1170 1171 {TTI::SK_PermuteTwoSrc, MVT::v4i16, 2}, // punpck/pshuflw 1172 {TTI::SK_PermuteTwoSrc, MVT::v2i16, 2}, // punpck/pshuflw 1173 {TTI::SK_PermuteTwoSrc, MVT::v8i8, 7}, // punpck/pshuflw 1174 {TTI::SK_PermuteTwoSrc, MVT::v4i8, 4}, // punpck/pshuflw 1175 {TTI::SK_PermuteTwoSrc, MVT::v2i8, 2}, // punpck 1176 1177 {TTI::SK_PermuteSingleSrc, MVT::v4i16, 1}, // pshuflw 1178 {TTI::SK_PermuteSingleSrc, MVT::v2i16, 1}, // pshuflw 1179 {TTI::SK_PermuteSingleSrc, MVT::v8i8, 5}, // punpck/pshuflw 1180 {TTI::SK_PermuteSingleSrc, MVT::v4i8, 3}, // punpck/pshuflw 1181 {TTI::SK_PermuteSingleSrc, MVT::v2i8, 1}, // punpck 1182 }; 1183 1184 if (ST->hasSSE2()) 1185 if (const auto *Entry = 1186 CostTableLookup(SSE2SubVectorShuffleTbl, Kind, VT.getSimpleVT())) 1187 return Entry->Cost; 1188 } 1189 1190 // We are going to permute multiple sources and the result will be in multiple 1191 // destinations. Providing an accurate cost only for splits where the element 1192 // type remains the same. 1193 if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) { 1194 MVT LegalVT = LT.second; 1195 if (LegalVT.isVector() && 1196 LegalVT.getVectorElementType().getSizeInBits() == 1197 BaseTp->getElementType()->getPrimitiveSizeInBits() && 1198 LegalVT.getVectorNumElements() < 1199 cast<FixedVectorType>(BaseTp)->getNumElements()) { 1200 1201 unsigned VecTySize = DL.getTypeStoreSize(BaseTp); 1202 unsigned LegalVTSize = LegalVT.getStoreSize(); 1203 // Number of source vectors after legalization: 1204 unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize; 1205 // Number of destination vectors after legalization: 1206 InstructionCost NumOfDests = LT.first; 1207 1208 auto *SingleOpTy = FixedVectorType::get(BaseTp->getElementType(), 1209 LegalVT.getVectorNumElements()); 1210 1211 InstructionCost NumOfShuffles = (NumOfSrcs - 1) * NumOfDests; 1212 return NumOfShuffles * getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy, 1213 None, 0, nullptr); 1214 } 1215 1216 return BaseT::getShuffleCost(Kind, BaseTp, Mask, Index, SubTp); 1217 } 1218 1219 // For 2-input shuffles, we must account for splitting the 2 inputs into many. 1220 if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) { 1221 // We assume that source and destination have the same vector type. 1222 InstructionCost NumOfDests = LT.first; 1223 InstructionCost NumOfShufflesPerDest = LT.first * 2 - 1; 1224 LT.first = NumOfDests * NumOfShufflesPerDest; 1225 } 1226 1227 static const CostTblEntry AVX512FP16ShuffleTbl[] = { 1228 {TTI::SK_Broadcast, MVT::v32f16, 1}, // vpbroadcastw 1229 {TTI::SK_Broadcast, MVT::v16f16, 1}, // vpbroadcastw 1230 {TTI::SK_Broadcast, MVT::v8f16, 1}, // vpbroadcastw 1231 1232 {TTI::SK_Reverse, MVT::v32f16, 2}, // vpermw 1233 {TTI::SK_Reverse, MVT::v16f16, 2}, // vpermw 1234 {TTI::SK_Reverse, MVT::v8f16, 1}, // vpshufb 1235 1236 {TTI::SK_PermuteSingleSrc, MVT::v32f16, 2}, // vpermw 1237 {TTI::SK_PermuteSingleSrc, MVT::v16f16, 2}, // vpermw 1238 {TTI::SK_PermuteSingleSrc, MVT::v8f16, 1}, // vpshufb 1239 1240 {TTI::SK_PermuteTwoSrc, MVT::v32f16, 2}, // vpermt2w 1241 {TTI::SK_PermuteTwoSrc, MVT::v16f16, 2}, // vpermt2w 1242 {TTI::SK_PermuteTwoSrc, MVT::v8f16, 2} // vpermt2w 1243 }; 1244 1245 if (!ST->useSoftFloat() && ST->hasFP16()) 1246 if (const auto *Entry = 1247 CostTableLookup(AVX512FP16ShuffleTbl, Kind, LT.second)) 1248 return LT.first * Entry->Cost; 1249 1250 static const CostTblEntry AVX512VBMIShuffleTbl[] = { 1251 {TTI::SK_Reverse, MVT::v64i8, 1}, // vpermb 1252 {TTI::SK_Reverse, MVT::v32i8, 1}, // vpermb 1253 1254 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 1}, // vpermb 1255 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 1}, // vpermb 1256 1257 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 2}, // vpermt2b 1258 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 2}, // vpermt2b 1259 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 2} // vpermt2b 1260 }; 1261 1262 if (ST->hasVBMI()) 1263 if (const auto *Entry = 1264 CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second)) 1265 return LT.first * Entry->Cost; 1266 1267 static const CostTblEntry AVX512BWShuffleTbl[] = { 1268 {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw 1269 {TTI::SK_Broadcast, MVT::v64i8, 1}, // vpbroadcastb 1270 1271 {TTI::SK_Reverse, MVT::v32i16, 2}, // vpermw 1272 {TTI::SK_Reverse, MVT::v16i16, 2}, // vpermw 1273 {TTI::SK_Reverse, MVT::v64i8, 2}, // pshufb + vshufi64x2 1274 1275 {TTI::SK_PermuteSingleSrc, MVT::v32i16, 2}, // vpermw 1276 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 2}, // vpermw 1277 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 8}, // extend to v32i16 1278 1279 {TTI::SK_PermuteTwoSrc, MVT::v32i16, 2}, // vpermt2w 1280 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 2}, // vpermt2w 1281 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 2}, // vpermt2w 1282 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 19}, // 6 * v32i8 + 1 1283 1284 {TTI::SK_Select, MVT::v32i16, 1}, // vblendmw 1285 {TTI::SK_Select, MVT::v64i8, 1}, // vblendmb 1286 }; 1287 1288 if (ST->hasBWI()) 1289 if (const auto *Entry = 1290 CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second)) 1291 return LT.first * Entry->Cost; 1292 1293 static const CostTblEntry AVX512ShuffleTbl[] = { 1294 {TTI::SK_Broadcast, MVT::v8f64, 1}, // vbroadcastpd 1295 {TTI::SK_Broadcast, MVT::v16f32, 1}, // vbroadcastps 1296 {TTI::SK_Broadcast, MVT::v8i64, 1}, // vpbroadcastq 1297 {TTI::SK_Broadcast, MVT::v16i32, 1}, // vpbroadcastd 1298 {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw 1299 {TTI::SK_Broadcast, MVT::v64i8, 1}, // vpbroadcastb 1300 1301 {TTI::SK_Reverse, MVT::v8f64, 1}, // vpermpd 1302 {TTI::SK_Reverse, MVT::v16f32, 1}, // vpermps 1303 {TTI::SK_Reverse, MVT::v8i64, 1}, // vpermq 1304 {TTI::SK_Reverse, MVT::v16i32, 1}, // vpermd 1305 {TTI::SK_Reverse, MVT::v32i16, 7}, // per mca 1306 {TTI::SK_Reverse, MVT::v64i8, 7}, // per mca 1307 1308 {TTI::SK_PermuteSingleSrc, MVT::v8f64, 1}, // vpermpd 1309 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd 1310 {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // vpermpd 1311 {TTI::SK_PermuteSingleSrc, MVT::v16f32, 1}, // vpermps 1312 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps 1313 {TTI::SK_PermuteSingleSrc, MVT::v4f32, 1}, // vpermps 1314 {TTI::SK_PermuteSingleSrc, MVT::v8i64, 1}, // vpermq 1315 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq 1316 {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // vpermq 1317 {TTI::SK_PermuteSingleSrc, MVT::v16i32, 1}, // vpermd 1318 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd 1319 {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // vpermd 1320 {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb 1321 1322 {TTI::SK_PermuteTwoSrc, MVT::v8f64, 1}, // vpermt2pd 1323 {TTI::SK_PermuteTwoSrc, MVT::v16f32, 1}, // vpermt2ps 1324 {TTI::SK_PermuteTwoSrc, MVT::v8i64, 1}, // vpermt2q 1325 {TTI::SK_PermuteTwoSrc, MVT::v16i32, 1}, // vpermt2d 1326 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 1}, // vpermt2pd 1327 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 1}, // vpermt2ps 1328 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 1}, // vpermt2q 1329 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 1}, // vpermt2d 1330 {TTI::SK_PermuteTwoSrc, MVT::v2f64, 1}, // vpermt2pd 1331 {TTI::SK_PermuteTwoSrc, MVT::v4f32, 1}, // vpermt2ps 1332 {TTI::SK_PermuteTwoSrc, MVT::v2i64, 1}, // vpermt2q 1333 {TTI::SK_PermuteTwoSrc, MVT::v4i32, 1}, // vpermt2d 1334 1335 // FIXME: This just applies the type legalization cost rules above 1336 // assuming these completely split. 1337 {TTI::SK_PermuteSingleSrc, MVT::v32i16, 14}, 1338 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 14}, 1339 {TTI::SK_PermuteTwoSrc, MVT::v32i16, 42}, 1340 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 42}, 1341 1342 {TTI::SK_Select, MVT::v32i16, 1}, // vpternlogq 1343 {TTI::SK_Select, MVT::v64i8, 1}, // vpternlogq 1344 {TTI::SK_Select, MVT::v8f64, 1}, // vblendmpd 1345 {TTI::SK_Select, MVT::v16f32, 1}, // vblendmps 1346 {TTI::SK_Select, MVT::v8i64, 1}, // vblendmq 1347 {TTI::SK_Select, MVT::v16i32, 1}, // vblendmd 1348 }; 1349 1350 if (ST->hasAVX512()) 1351 if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second)) 1352 return LT.first * Entry->Cost; 1353 1354 static const CostTblEntry AVX2ShuffleTbl[] = { 1355 {TTI::SK_Broadcast, MVT::v4f64, 1}, // vbroadcastpd 1356 {TTI::SK_Broadcast, MVT::v8f32, 1}, // vbroadcastps 1357 {TTI::SK_Broadcast, MVT::v4i64, 1}, // vpbroadcastq 1358 {TTI::SK_Broadcast, MVT::v8i32, 1}, // vpbroadcastd 1359 {TTI::SK_Broadcast, MVT::v16i16, 1}, // vpbroadcastw 1360 {TTI::SK_Broadcast, MVT::v32i8, 1}, // vpbroadcastb 1361 1362 {TTI::SK_Reverse, MVT::v4f64, 1}, // vpermpd 1363 {TTI::SK_Reverse, MVT::v8f32, 1}, // vpermps 1364 {TTI::SK_Reverse, MVT::v4i64, 1}, // vpermq 1365 {TTI::SK_Reverse, MVT::v8i32, 1}, // vpermd 1366 {TTI::SK_Reverse, MVT::v16i16, 2}, // vperm2i128 + pshufb 1367 {TTI::SK_Reverse, MVT::v32i8, 2}, // vperm2i128 + pshufb 1368 1369 {TTI::SK_Select, MVT::v16i16, 1}, // vpblendvb 1370 {TTI::SK_Select, MVT::v32i8, 1}, // vpblendvb 1371 1372 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd 1373 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps 1374 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq 1375 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd 1376 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vperm2i128 + 2*vpshufb 1377 // + vpblendvb 1378 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4}, // vperm2i128 + 2*vpshufb 1379 // + vpblendvb 1380 1381 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3}, // 2*vpermpd + vblendpd 1382 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 3}, // 2*vpermps + vblendps 1383 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3}, // 2*vpermq + vpblendd 1384 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 3}, // 2*vpermd + vpblendd 1385 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 7}, // 2*vperm2i128 + 4*vpshufb 1386 // + vpblendvb 1387 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 7}, // 2*vperm2i128 + 4*vpshufb 1388 // + vpblendvb 1389 }; 1390 1391 if (ST->hasAVX2()) 1392 if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second)) 1393 return LT.first * Entry->Cost; 1394 1395 static const CostTblEntry XOPShuffleTbl[] = { 1396 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2}, // vperm2f128 + vpermil2pd 1397 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 2}, // vperm2f128 + vpermil2ps 1398 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2}, // vperm2f128 + vpermil2pd 1399 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 2}, // vperm2f128 + vpermil2ps 1400 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vextractf128 + 2*vpperm 1401 // + vinsertf128 1402 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4}, // vextractf128 + 2*vpperm 1403 // + vinsertf128 1404 1405 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 9}, // 2*vextractf128 + 6*vpperm 1406 // + vinsertf128 1407 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 1}, // vpperm 1408 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 9}, // 2*vextractf128 + 6*vpperm 1409 // + vinsertf128 1410 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 1}, // vpperm 1411 }; 1412 1413 if (ST->hasXOP()) 1414 if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second)) 1415 return LT.first * Entry->Cost; 1416 1417 static const CostTblEntry AVX1ShuffleTbl[] = { 1418 {TTI::SK_Broadcast, MVT::v4f64, 2}, // vperm2f128 + vpermilpd 1419 {TTI::SK_Broadcast, MVT::v8f32, 2}, // vperm2f128 + vpermilps 1420 {TTI::SK_Broadcast, MVT::v4i64, 2}, // vperm2f128 + vpermilpd 1421 {TTI::SK_Broadcast, MVT::v8i32, 2}, // vperm2f128 + vpermilps 1422 {TTI::SK_Broadcast, MVT::v16i16, 3}, // vpshuflw + vpshufd + vinsertf128 1423 {TTI::SK_Broadcast, MVT::v32i8, 2}, // vpshufb + vinsertf128 1424 1425 {TTI::SK_Reverse, MVT::v4f64, 2}, // vperm2f128 + vpermilpd 1426 {TTI::SK_Reverse, MVT::v8f32, 2}, // vperm2f128 + vpermilps 1427 {TTI::SK_Reverse, MVT::v4i64, 2}, // vperm2f128 + vpermilpd 1428 {TTI::SK_Reverse, MVT::v8i32, 2}, // vperm2f128 + vpermilps 1429 {TTI::SK_Reverse, MVT::v16i16, 4}, // vextractf128 + 2*pshufb 1430 // + vinsertf128 1431 {TTI::SK_Reverse, MVT::v32i8, 4}, // vextractf128 + 2*pshufb 1432 // + vinsertf128 1433 1434 {TTI::SK_Select, MVT::v4i64, 1}, // vblendpd 1435 {TTI::SK_Select, MVT::v4f64, 1}, // vblendpd 1436 {TTI::SK_Select, MVT::v8i32, 1}, // vblendps 1437 {TTI::SK_Select, MVT::v8f32, 1}, // vblendps 1438 {TTI::SK_Select, MVT::v16i16, 3}, // vpand + vpandn + vpor 1439 {TTI::SK_Select, MVT::v32i8, 3}, // vpand + vpandn + vpor 1440 1441 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2}, // vperm2f128 + vshufpd 1442 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2}, // vperm2f128 + vshufpd 1443 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 4}, // 2*vperm2f128 + 2*vshufps 1444 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 4}, // 2*vperm2f128 + 2*vshufps 1445 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 8}, // vextractf128 + 4*pshufb 1446 // + 2*por + vinsertf128 1447 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 8}, // vextractf128 + 4*pshufb 1448 // + 2*por + vinsertf128 1449 1450 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3}, // 2*vperm2f128 + vshufpd 1451 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3}, // 2*vperm2f128 + vshufpd 1452 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 4}, // 2*vperm2f128 + 2*vshufps 1453 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 4}, // 2*vperm2f128 + 2*vshufps 1454 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 15}, // 2*vextractf128 + 8*pshufb 1455 // + 4*por + vinsertf128 1456 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 15}, // 2*vextractf128 + 8*pshufb 1457 // + 4*por + vinsertf128 1458 }; 1459 1460 if (ST->hasAVX()) 1461 if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second)) 1462 return LT.first * Entry->Cost; 1463 1464 static const CostTblEntry SSE41ShuffleTbl[] = { 1465 {TTI::SK_Select, MVT::v2i64, 1}, // pblendw 1466 {TTI::SK_Select, MVT::v2f64, 1}, // movsd 1467 {TTI::SK_Select, MVT::v4i32, 1}, // pblendw 1468 {TTI::SK_Select, MVT::v4f32, 1}, // blendps 1469 {TTI::SK_Select, MVT::v8i16, 1}, // pblendw 1470 {TTI::SK_Select, MVT::v16i8, 1} // pblendvb 1471 }; 1472 1473 if (ST->hasSSE41()) 1474 if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second)) 1475 return LT.first * Entry->Cost; 1476 1477 static const CostTblEntry SSSE3ShuffleTbl[] = { 1478 {TTI::SK_Broadcast, MVT::v8i16, 1}, // pshufb 1479 {TTI::SK_Broadcast, MVT::v16i8, 1}, // pshufb 1480 1481 {TTI::SK_Reverse, MVT::v8i16, 1}, // pshufb 1482 {TTI::SK_Reverse, MVT::v16i8, 1}, // pshufb 1483 1484 {TTI::SK_Select, MVT::v8i16, 3}, // 2*pshufb + por 1485 {TTI::SK_Select, MVT::v16i8, 3}, // 2*pshufb + por 1486 1487 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 1}, // pshufb 1488 {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb 1489 1490 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 3}, // 2*pshufb + por 1491 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 3}, // 2*pshufb + por 1492 }; 1493 1494 if (ST->hasSSSE3()) 1495 if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second)) 1496 return LT.first * Entry->Cost; 1497 1498 static const CostTblEntry SSE2ShuffleTbl[] = { 1499 {TTI::SK_Broadcast, MVT::v2f64, 1}, // shufpd 1500 {TTI::SK_Broadcast, MVT::v2i64, 1}, // pshufd 1501 {TTI::SK_Broadcast, MVT::v4i32, 1}, // pshufd 1502 {TTI::SK_Broadcast, MVT::v8i16, 2}, // pshuflw + pshufd 1503 {TTI::SK_Broadcast, MVT::v16i8, 3}, // unpck + pshuflw + pshufd 1504 1505 {TTI::SK_Reverse, MVT::v2f64, 1}, // shufpd 1506 {TTI::SK_Reverse, MVT::v2i64, 1}, // pshufd 1507 {TTI::SK_Reverse, MVT::v4i32, 1}, // pshufd 1508 {TTI::SK_Reverse, MVT::v8i16, 3}, // pshuflw + pshufhw + pshufd 1509 {TTI::SK_Reverse, MVT::v16i8, 9}, // 2*pshuflw + 2*pshufhw 1510 // + 2*pshufd + 2*unpck + packus 1511 1512 {TTI::SK_Select, MVT::v2i64, 1}, // movsd 1513 {TTI::SK_Select, MVT::v2f64, 1}, // movsd 1514 {TTI::SK_Select, MVT::v4i32, 2}, // 2*shufps 1515 {TTI::SK_Select, MVT::v8i16, 3}, // pand + pandn + por 1516 {TTI::SK_Select, MVT::v16i8, 3}, // pand + pandn + por 1517 1518 {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // shufpd 1519 {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // pshufd 1520 {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // pshufd 1521 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 5}, // 2*pshuflw + 2*pshufhw 1522 // + pshufd/unpck 1523 { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw 1524 // + 2*pshufd + 2*unpck + 2*packus 1525 1526 { TTI::SK_PermuteTwoSrc, MVT::v2f64, 1 }, // shufpd 1527 { TTI::SK_PermuteTwoSrc, MVT::v2i64, 1 }, // shufpd 1528 { TTI::SK_PermuteTwoSrc, MVT::v4i32, 2 }, // 2*{unpck,movsd,pshufd} 1529 { TTI::SK_PermuteTwoSrc, MVT::v8i16, 8 }, // blend+permute 1530 { TTI::SK_PermuteTwoSrc, MVT::v16i8, 13 }, // blend+permute 1531 }; 1532 1533 if (ST->hasSSE2()) 1534 if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second)) 1535 return LT.first * Entry->Cost; 1536 1537 static const CostTblEntry SSE1ShuffleTbl[] = { 1538 { TTI::SK_Broadcast, MVT::v4f32, 1 }, // shufps 1539 { TTI::SK_Reverse, MVT::v4f32, 1 }, // shufps 1540 { TTI::SK_Select, MVT::v4f32, 2 }, // 2*shufps 1541 { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps 1542 { TTI::SK_PermuteTwoSrc, MVT::v4f32, 2 }, // 2*shufps 1543 }; 1544 1545 if (ST->hasSSE1()) 1546 if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second)) 1547 return LT.first * Entry->Cost; 1548 1549 return BaseT::getShuffleCost(Kind, BaseTp, Mask, Index, SubTp); 1550 } 1551 1552 InstructionCost X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, 1553 Type *Src, 1554 TTI::CastContextHint CCH, 1555 TTI::TargetCostKind CostKind, 1556 const Instruction *I) { 1557 int ISD = TLI->InstructionOpcodeToISD(Opcode); 1558 assert(ISD && "Invalid opcode"); 1559 1560 // TODO: Allow non-throughput costs that aren't binary. 1561 auto AdjustCost = [&CostKind](InstructionCost Cost) -> InstructionCost { 1562 if (CostKind != TTI::TCK_RecipThroughput) 1563 return Cost == 0 ? 0 : 1; 1564 return Cost; 1565 }; 1566 1567 // The cost tables include both specific, custom (non-legal) src/dst type 1568 // conversions and generic, legalized types. We test for customs first, before 1569 // falling back to legalization. 1570 // FIXME: Need a better design of the cost table to handle non-simple types of 1571 // potential massive combinations (elem_num x src_type x dst_type). 1572 static const TypeConversionCostTblEntry AVX512BWConversionTbl[] { 1573 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 1 }, 1574 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 1 }, 1575 1576 // Mask sign extend has an instruction. 1577 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 1 }, 1578 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 1 }, 1579 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 1 }, 1580 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 1 }, 1581 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 1 }, 1582 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 1 }, 1583 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 1 }, 1584 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1585 { ISD::SIGN_EXTEND, MVT::v32i8, MVT::v32i1, 1 }, 1586 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i1, 1 }, 1587 { ISD::SIGN_EXTEND, MVT::v64i8, MVT::v64i1, 1 }, 1588 1589 // Mask zero extend is a sext + shift. 1590 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 2 }, 1591 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 2 }, 1592 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 2 }, 1593 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 2 }, 1594 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 2 }, 1595 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 2 }, 1596 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 2 }, 1597 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 }, 1598 { ISD::ZERO_EXTEND, MVT::v32i8, MVT::v32i1, 2 }, 1599 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i1, 2 }, 1600 { ISD::ZERO_EXTEND, MVT::v64i8, MVT::v64i1, 2 }, 1601 1602 { ISD::TRUNCATE, MVT::v32i8, MVT::v32i16, 2 }, 1603 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 }, // widen to zmm 1604 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 2 }, // widen to zmm 1605 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // widen to zmm 1606 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i16, 2 }, // vpmovwb 1607 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // widen to zmm 1608 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 2 }, // widen to zmm 1609 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i16, 2 }, // vpmovwb 1610 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 2 }, // widen to zmm 1611 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 2 }, // widen to zmm 1612 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 2 }, // vpmovwb 1613 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 2 }, // widen to zmm 1614 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 2 }, // widen to zmm 1615 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i8, 2 }, // widen to zmm 1616 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i16, 2 }, 1617 { ISD::TRUNCATE, MVT::v64i1, MVT::v64i8, 2 }, 1618 }; 1619 1620 static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = { 1621 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 }, 1622 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 }, 1623 1624 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 }, 1625 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 }, 1626 1627 { ISD::FP_TO_SINT, MVT::v8i64, MVT::v8f32, 1 }, 1628 { ISD::FP_TO_SINT, MVT::v8i64, MVT::v8f64, 1 }, 1629 1630 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f32, 1 }, 1631 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f64, 1 }, 1632 }; 1633 1634 // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and 1635 // 256-bit wide vectors. 1636 1637 static const TypeConversionCostTblEntry AVX512FConversionTbl[] = { 1638 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 1 }, 1639 { ISD::FP_EXTEND, MVT::v8f64, MVT::v16f32, 3 }, 1640 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 1 }, 1641 1642 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // sext+vpslld+vptestmd 1643 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 3 }, // sext+vpslld+vptestmd 1644 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 3 }, // sext+vpslld+vptestmd 1645 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 3 }, // sext+vpslld+vptestmd 1646 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 3 }, // sext+vpsllq+vptestmq 1647 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 3 }, // sext+vpsllq+vptestmq 1648 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 3 }, // sext+vpsllq+vptestmq 1649 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 3 }, // sext+vpslld+vptestmd 1650 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 2 }, // zmm vpslld+vptestmd 1651 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i32, 2 }, // zmm vpslld+vptestmd 1652 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, // zmm vpslld+vptestmd 1653 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i32, 2 }, // vpslld+vptestmd 1654 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i64, 2 }, // zmm vpsllq+vptestmq 1655 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 2 }, // zmm vpsllq+vptestmq 1656 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i64, 2 }, // vpsllq+vptestmq 1657 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i32, 2 }, // vpmovdb 1658 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 2 }, // vpmovdb 1659 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 2 }, // vpmovdb 1660 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 2 }, // vpmovdb 1661 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i64, 2 }, // vpmovqb 1662 { ISD::TRUNCATE, MVT::v2i16, MVT::v2i64, 1 }, // vpshufb 1663 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i64, 2 }, // vpmovqb 1664 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 2 }, // vpmovqw 1665 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 1 }, // vpmovqd 1666 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, // zmm vpmovqd 1667 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i64, 5 },// 2*vpmovqd+concat+vpmovdb 1668 1669 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 3 }, // extend to v16i32 1670 { ISD::TRUNCATE, MVT::v32i8, MVT::v32i16, 8 }, 1671 1672 // Sign extend is zmm vpternlogd+vptruncdb. 1673 // Zero extend is zmm broadcast load+vptruncdw. 1674 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 3 }, 1675 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 4 }, 1676 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 3 }, 1677 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 4 }, 1678 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 3 }, 1679 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 4 }, 1680 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 3 }, 1681 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 4 }, 1682 1683 // Sign extend is zmm vpternlogd+vptruncdw. 1684 // Zero extend is zmm vpternlogd+vptruncdw+vpsrlw. 1685 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 3 }, 1686 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 4 }, 1687 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 3 }, 1688 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 4 }, 1689 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 3 }, 1690 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 4 }, 1691 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 3 }, 1692 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 4 }, 1693 1694 { ISD::SIGN_EXTEND, MVT::v2i32, MVT::v2i1, 1 }, // zmm vpternlogd 1695 { ISD::ZERO_EXTEND, MVT::v2i32, MVT::v2i1, 2 }, // zmm vpternlogd+psrld 1696 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i1, 1 }, // zmm vpternlogd 1697 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i1, 2 }, // zmm vpternlogd+psrld 1698 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 1 }, // zmm vpternlogd 1699 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 2 }, // zmm vpternlogd+psrld 1700 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i1, 1 }, // zmm vpternlogq 1701 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i1, 2 }, // zmm vpternlogq+psrlq 1702 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 1 }, // zmm vpternlogq 1703 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 2 }, // zmm vpternlogq+psrlq 1704 1705 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1, 1 }, // vpternlogd 1706 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1, 2 }, // vpternlogd+psrld 1707 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i1, 1 }, // vpternlogq 1708 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i1, 2 }, // vpternlogq+psrlq 1709 1710 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 1 }, 1711 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 1 }, 1712 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 }, 1713 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 }, 1714 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 1 }, 1715 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 1 }, 1716 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 1 }, 1717 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 1 }, 1718 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i32, 1 }, 1719 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i32, 1 }, 1720 1721 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 3 }, // FIXME: May not be right 1722 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 3 }, // FIXME: May not be right 1723 1724 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 }, 1725 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 }, 1726 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v16i8, 2 }, 1727 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i8, 1 }, 1728 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 }, 1729 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 1 }, 1730 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 }, 1731 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 }, 1732 1733 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 }, 1734 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 }, 1735 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v16i8, 2 }, 1736 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i8, 1 }, 1737 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 }, 1738 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i16, 1 }, 1739 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 }, 1740 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 }, 1741 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 26 }, 1742 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 5 }, 1743 1744 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f32, 2 }, 1745 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f64, 7 }, 1746 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v32f64,15 }, 1747 { ISD::FP_TO_SINT, MVT::v64i8, MVT::v64f32,11 }, 1748 { ISD::FP_TO_SINT, MVT::v64i8, MVT::v64f64,31 }, 1749 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f64, 3 }, 1750 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v16f64, 7 }, 1751 { ISD::FP_TO_SINT, MVT::v32i16, MVT::v32f32, 5 }, 1752 { ISD::FP_TO_SINT, MVT::v32i16, MVT::v32f64,15 }, 1753 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f64, 1 }, 1754 { ISD::FP_TO_SINT, MVT::v16i32, MVT::v16f64, 3 }, 1755 1756 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f64, 1 }, 1757 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f64, 3 }, 1758 { ISD::FP_TO_UINT, MVT::v8i8, MVT::v8f64, 3 }, 1759 { ISD::FP_TO_UINT, MVT::v16i32, MVT::v16f32, 1 }, 1760 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v16f32, 3 }, 1761 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v16f32, 3 }, 1762 }; 1763 1764 static const TypeConversionCostTblEntry AVX512BWVLConversionTbl[] { 1765 // Mask sign extend has an instruction. 1766 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 1 }, 1767 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 1 }, 1768 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 1 }, 1769 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 1 }, 1770 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 1 }, 1771 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 1 }, 1772 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 1 }, 1773 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1774 { ISD::SIGN_EXTEND, MVT::v32i8, MVT::v32i1, 1 }, 1775 1776 // Mask zero extend is a sext + shift. 1777 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 2 }, 1778 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 2 }, 1779 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 2 }, 1780 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 2 }, 1781 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 2 }, 1782 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 2 }, 1783 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 2 }, 1784 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 }, 1785 { ISD::ZERO_EXTEND, MVT::v32i8, MVT::v32i1, 2 }, 1786 1787 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 }, 1788 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 2 }, // vpsllw+vptestmb 1789 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // vpsllw+vptestmw 1790 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // vpsllw+vptestmb 1791 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 2 }, // vpsllw+vptestmw 1792 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 2 }, // vpsllw+vptestmb 1793 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 2 }, // vpsllw+vptestmw 1794 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 2 }, // vpsllw+vptestmb 1795 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 2 }, // vpsllw+vptestmw 1796 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i8, 2 }, // vpsllw+vptestmb 1797 }; 1798 1799 static const TypeConversionCostTblEntry AVX512DQVLConversionTbl[] = { 1800 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 }, 1801 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 }, 1802 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 }, 1803 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 }, 1804 1805 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 }, 1806 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 }, 1807 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 }, 1808 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 }, 1809 1810 { ISD::FP_TO_SINT, MVT::v2i64, MVT::v4f32, 1 }, 1811 { ISD::FP_TO_SINT, MVT::v4i64, MVT::v4f32, 1 }, 1812 { ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f64, 1 }, 1813 { ISD::FP_TO_SINT, MVT::v4i64, MVT::v4f64, 1 }, 1814 1815 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v4f32, 1 }, 1816 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f32, 1 }, 1817 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f64, 1 }, 1818 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f64, 1 }, 1819 }; 1820 1821 static const TypeConversionCostTblEntry AVX512VLConversionTbl[] = { 1822 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // sext+vpslld+vptestmd 1823 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 3 }, // sext+vpslld+vptestmd 1824 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 3 }, // sext+vpslld+vptestmd 1825 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 8 }, // split+2*v8i8 1826 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 3 }, // sext+vpsllq+vptestmq 1827 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 3 }, // sext+vpsllq+vptestmq 1828 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 3 }, // sext+vpsllq+vptestmq 1829 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 8 }, // split+2*v8i16 1830 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 2 }, // vpslld+vptestmd 1831 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i32, 2 }, // vpslld+vptestmd 1832 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, // vpslld+vptestmd 1833 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i64, 2 }, // vpsllq+vptestmq 1834 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 2 }, // vpsllq+vptestmq 1835 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, // vpmovqd 1836 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 2 }, // vpmovqb 1837 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2 }, // vpmovqw 1838 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 }, // vpmovwb 1839 1840 // sign extend is vpcmpeq+maskedmove+vpmovdw+vpacksswb 1841 // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw+vpackuswb 1842 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 5 }, 1843 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 6 }, 1844 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 5 }, 1845 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 6 }, 1846 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 5 }, 1847 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 6 }, 1848 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 10 }, 1849 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 12 }, 1850 1851 // sign extend is vpcmpeq+maskedmove+vpmovdw 1852 // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw 1853 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 4 }, 1854 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 5 }, 1855 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 4 }, 1856 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 5 }, 1857 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 4 }, 1858 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 5 }, 1859 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 10 }, 1860 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 12 }, 1861 1862 { ISD::SIGN_EXTEND, MVT::v2i32, MVT::v2i1, 1 }, // vpternlogd 1863 { ISD::ZERO_EXTEND, MVT::v2i32, MVT::v2i1, 2 }, // vpternlogd+psrld 1864 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i1, 1 }, // vpternlogd 1865 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i1, 2 }, // vpternlogd+psrld 1866 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 1 }, // vpternlogd 1867 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 2 }, // vpternlogd+psrld 1868 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i1, 1 }, // vpternlogq 1869 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i1, 2 }, // vpternlogq+psrlq 1870 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 1 }, // vpternlogq 1871 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 2 }, // vpternlogq+psrlq 1872 1873 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v16i8, 1 }, 1874 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v16i8, 1 }, 1875 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v16i8, 1 }, 1876 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v16i8, 1 }, 1877 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 }, 1878 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1 }, 1879 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v8i16, 1 }, 1880 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v8i16, 1 }, 1881 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 1882 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 1883 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 1884 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 1885 1886 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 }, 1887 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v16i8, 1 }, 1888 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 }, 1889 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 1 }, 1890 1891 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 1 }, 1892 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 1 }, 1893 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 }, 1894 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v16i8, 1 }, 1895 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 }, 1896 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 1 }, 1897 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 }, 1898 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 1899 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 }, 1900 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 }, 1901 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 5 }, 1902 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 }, 1903 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 5 }, 1904 1905 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v8f32, 2 }, 1906 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f32, 2 }, 1907 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v32f32, 5 }, 1908 1909 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 1 }, 1910 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 1 }, 1911 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 }, 1912 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 1 }, 1913 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 1 }, 1914 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 1 }, 1915 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f64, 1 }, 1916 }; 1917 1918 static const TypeConversionCostTblEntry AVX2ConversionTbl[] = { 1919 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 3 }, 1920 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 3 }, 1921 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 3 }, 1922 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 3 }, 1923 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1924 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 1 }, 1925 1926 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v16i8, 2 }, 1927 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v16i8, 2 }, 1928 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v16i8, 2 }, 1929 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v16i8, 2 }, 1930 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 2 }, 1931 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 2 }, 1932 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v8i16, 2 }, 1933 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v8i16, 2 }, 1934 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 2 }, 1935 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 2 }, 1936 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 3 }, 1937 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 3 }, 1938 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 2 }, 1939 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 2 }, 1940 1941 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, 1942 1943 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 4 }, 1944 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 4 }, 1945 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i16, 1 }, 1946 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i32, 1 }, 1947 { ISD::TRUNCATE, MVT::v16i8, MVT::v2i64, 1 }, 1948 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i32, 4 }, 1949 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i64, 4 }, 1950 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i32, 1 }, 1951 { ISD::TRUNCATE, MVT::v8i16, MVT::v2i64, 1 }, 1952 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i64, 5 }, 1953 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, 1954 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 }, 1955 1956 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 3 }, 1957 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 3 }, 1958 1959 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v8f32, 1 }, 1960 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f64, 1 }, 1961 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f32, 1 }, 1962 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f64, 3 }, 1963 1964 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 3 }, 1965 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 3 }, 1966 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v8f32, 1 }, 1967 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 3 }, 1968 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 4 }, 1969 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4 }, 1970 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 3 }, 1971 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v4f64, 4 }, 1972 1973 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 2 }, 1974 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v16i8, 2 }, 1975 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 2 }, 1976 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 2 }, 1977 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 }, 1978 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 }, 1979 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 3 }, 1980 1981 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 2 }, 1982 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v16i8, 2 }, 1983 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 2 }, 1984 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 2 }, 1985 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 2 }, 1986 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 1 }, 1987 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 2 }, 1988 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 2 }, 1989 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 }, 1990 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 4 }, 1991 }; 1992 1993 static const TypeConversionCostTblEntry AVXConversionTbl[] = { 1994 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 6 }, 1995 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 4 }, 1996 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 7 }, 1997 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 4 }, 1998 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 4 }, 1999 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 4 }, 2000 2001 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v16i8, 3 }, 2002 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v16i8, 3 }, 2003 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v16i8, 3 }, 2004 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v16i8, 3 }, 2005 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 3 }, 2006 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 3 }, 2007 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v8i16, 3 }, 2008 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v8i16, 3 }, 2009 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 3 }, 2010 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 3 }, 2011 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 3 }, 2012 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 3 }, 2013 2014 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 4 }, 2015 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 5 }, 2016 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 4 }, 2017 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i64, 9 }, 2018 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i64, 11 }, 2019 2020 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 6 }, 2021 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 6 }, 2022 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 }, // and+extract+packuswb 2023 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i32, 5 }, 2024 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 }, 2025 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i64, 5 }, 2026 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i64, 3 }, // and+extract+2*packusdw 2027 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 }, 2028 2029 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 }, 2030 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 }, 2031 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 }, 2032 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v16i8, 4 }, 2033 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v16i8, 2 }, 2034 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 }, 2035 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v8i16, 2 }, 2036 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 2 }, 2037 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 }, 2038 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 4 }, 2039 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 5 }, 2040 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i64, 8 }, 2041 2042 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 }, 2043 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 }, 2044 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 }, 2045 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v16i8, 4 }, 2046 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v16i8, 2 }, 2047 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 }, 2048 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v8i16, 2 }, 2049 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 4 }, 2050 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 4 }, 2051 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 }, 2052 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 }, 2053 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 8 }, 2054 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 10 }, 2055 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 10 }, 2056 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 18 }, 2057 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 }, 2058 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 10 }, 2059 2060 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v8f32, 2 }, 2061 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v4f64, 2 }, 2062 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v8f32, 2 }, 2063 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v4f64, 2 }, 2064 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f32, 2 }, 2065 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v4f64, 2 }, 2066 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v8f32, 2 }, 2067 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v4f64, 2 }, 2068 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f64, 2 }, 2069 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f32, 2 }, 2070 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f64, 5 }, 2071 2072 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v8f32, 2 }, 2073 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v4f64, 2 }, 2074 { ISD::FP_TO_UINT, MVT::v32i8, MVT::v8f32, 2 }, 2075 { ISD::FP_TO_UINT, MVT::v32i8, MVT::v4f64, 2 }, 2076 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f32, 2 }, 2077 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v4f64, 2 }, 2078 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v8f32, 2 }, 2079 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v4f64, 2 }, 2080 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 3 }, 2081 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 4 }, 2082 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 6 }, 2083 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 7 }, 2084 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v4f64, 7 }, 2085 2086 { ISD::FP_EXTEND, MVT::v4f64, MVT::v4f32, 1 }, 2087 { ISD::FP_ROUND, MVT::v4f32, MVT::v4f64, 1 }, 2088 }; 2089 2090 static const TypeConversionCostTblEntry SSE41ConversionTbl[] = { 2091 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v16i8, 1 }, 2092 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v16i8, 1 }, 2093 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v16i8, 1 }, 2094 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v16i8, 1 }, 2095 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v16i8, 1 }, 2096 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v16i8, 1 }, 2097 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v8i16, 1 }, 2098 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v8i16, 1 }, 2099 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v8i16, 1 }, 2100 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v8i16, 1 }, 2101 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v4i32, 1 }, 2102 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v4i32, 1 }, 2103 2104 // These truncates end up widening elements. 2105 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 1 }, // PMOVXZBQ 2106 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 1 }, // PMOVXZWQ 2107 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 1 }, // PMOVXZBD 2108 2109 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i32, 2 }, 2110 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i32, 2 }, 2111 { ISD::TRUNCATE, MVT::v16i8, MVT::v2i64, 2 }, 2112 2113 { ISD::SINT_TO_FP, MVT::f32, MVT::i32, 1 }, 2114 { ISD::SINT_TO_FP, MVT::f64, MVT::i32, 1 }, 2115 { ISD::SINT_TO_FP, MVT::f32, MVT::i64, 1 }, 2116 { ISD::SINT_TO_FP, MVT::f64, MVT::i64, 1 }, 2117 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 1 }, 2118 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 }, 2119 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 1 }, 2120 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 }, 2121 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 2122 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 1 }, 2123 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 2 }, 2124 2125 { ISD::UINT_TO_FP, MVT::f32, MVT::i32, 1 }, 2126 { ISD::UINT_TO_FP, MVT::f64, MVT::i32, 1 }, 2127 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 4 }, 2128 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 4 }, 2129 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 1 }, 2130 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 }, 2131 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 1 }, 2132 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 }, 2133 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 3 }, 2134 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 3 }, 2135 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 2 }, 2136 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 12 }, 2137 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 22 }, 2138 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 4 }, 2139 2140 { ISD::FP_TO_SINT, MVT::i32, MVT::f32, 1 }, 2141 { ISD::FP_TO_SINT, MVT::i64, MVT::f32, 1 }, 2142 { ISD::FP_TO_SINT, MVT::i32, MVT::f64, 1 }, 2143 { ISD::FP_TO_SINT, MVT::i64, MVT::f64, 1 }, 2144 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v4f32, 2 }, 2145 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v2f64, 2 }, 2146 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v4f32, 1 }, 2147 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v2f64, 1 }, 2148 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 1 }, 2149 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v2f64, 1 }, 2150 2151 { ISD::FP_TO_UINT, MVT::i32, MVT::f32, 1 }, 2152 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 4 }, 2153 { ISD::FP_TO_UINT, MVT::i32, MVT::f64, 1 }, 2154 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 4 }, 2155 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v4f32, 2 }, 2156 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v2f64, 2 }, 2157 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v4f32, 1 }, 2158 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v2f64, 1 }, 2159 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 4 }, 2160 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 4 }, 2161 }; 2162 2163 static const TypeConversionCostTblEntry SSE2ConversionTbl[] = { 2164 // These are somewhat magic numbers justified by comparing the 2165 // output of llvm-mca for our various supported scheduler models 2166 // and basing it off the worst case scenario. 2167 { ISD::SINT_TO_FP, MVT::f32, MVT::i32, 3 }, 2168 { ISD::SINT_TO_FP, MVT::f64, MVT::i32, 3 }, 2169 { ISD::SINT_TO_FP, MVT::f32, MVT::i64, 3 }, 2170 { ISD::SINT_TO_FP, MVT::f64, MVT::i64, 3 }, 2171 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 3 }, 2172 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 4 }, 2173 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 3 }, 2174 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 4 }, 2175 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 3 }, 2176 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4 }, 2177 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 8 }, 2178 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 8 }, 2179 2180 { ISD::UINT_TO_FP, MVT::f32, MVT::i32, 3 }, 2181 { ISD::UINT_TO_FP, MVT::f64, MVT::i32, 3 }, 2182 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 8 }, 2183 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 9 }, 2184 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 4 }, 2185 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 4 }, 2186 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 4 }, 2187 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 4 }, 2188 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 7 }, 2189 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 7 }, 2190 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 }, 2191 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 15 }, 2192 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 18 }, 2193 2194 { ISD::FP_TO_SINT, MVT::i32, MVT::f32, 4 }, 2195 { ISD::FP_TO_SINT, MVT::i64, MVT::f32, 4 }, 2196 { ISD::FP_TO_SINT, MVT::i32, MVT::f64, 4 }, 2197 { ISD::FP_TO_SINT, MVT::i64, MVT::f64, 4 }, 2198 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v4f32, 6 }, 2199 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v2f64, 6 }, 2200 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v4f32, 5 }, 2201 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v2f64, 5 }, 2202 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 4 }, 2203 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v2f64, 4 }, 2204 2205 { ISD::FP_TO_UINT, MVT::i32, MVT::f32, 4 }, 2206 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 4 }, 2207 { ISD::FP_TO_UINT, MVT::i32, MVT::f64, 4 }, 2208 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 15 }, 2209 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v4f32, 6 }, 2210 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v2f64, 6 }, 2211 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v4f32, 5 }, 2212 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v2f64, 5 }, 2213 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 8 }, 2214 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 8 }, 2215 2216 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v16i8, 4 }, 2217 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v16i8, 4 }, 2218 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v16i8, 2 }, 2219 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v16i8, 3 }, 2220 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v16i8, 1 }, 2221 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v16i8, 2 }, 2222 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v8i16, 2 }, 2223 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v8i16, 3 }, 2224 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v8i16, 1 }, 2225 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v8i16, 2 }, 2226 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v4i32, 1 }, 2227 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v4i32, 2 }, 2228 2229 // These truncates are really widening elements. 2230 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 1 }, // PSHUFD 2231 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // PUNPCKLWD+DQ 2232 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // PUNPCKLBW+WD+PSHUFD 2233 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 1 }, // PUNPCKLWD 2234 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // PUNPCKLBW+WD 2235 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 1 }, // PUNPCKLBW 2236 2237 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i16, 2 }, // PAND+PACKUSWB 2238 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 3 }, 2239 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i32, 3 }, // PAND+2*PACKUSWB 2240 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 7 }, 2241 { ISD::TRUNCATE, MVT::v2i16, MVT::v2i32, 1 }, 2242 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i32, 3 }, 2243 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 }, 2244 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32,10 }, 2245 { ISD::TRUNCATE, MVT::v16i8, MVT::v2i64, 4 }, // PAND+3*PACKUSWB 2246 { ISD::TRUNCATE, MVT::v8i16, MVT::v2i64, 2 }, // PSHUFD+PSHUFLW 2247 { ISD::TRUNCATE, MVT::v4i32, MVT::v2i64, 1 }, // PSHUFD 2248 }; 2249 2250 // Attempt to map directly to (simple) MVT types to let us match custom entries. 2251 EVT SrcTy = TLI->getValueType(DL, Src); 2252 EVT DstTy = TLI->getValueType(DL, Dst); 2253 2254 // The function getSimpleVT only handles simple value types. 2255 if (SrcTy.isSimple() && DstTy.isSimple()) { 2256 MVT SimpleSrcTy = SrcTy.getSimpleVT(); 2257 MVT SimpleDstTy = DstTy.getSimpleVT(); 2258 2259 if (ST->useAVX512Regs()) { 2260 if (ST->hasBWI()) 2261 if (const auto *Entry = ConvertCostTableLookup( 2262 AVX512BWConversionTbl, ISD, SimpleDstTy, SimpleSrcTy)) 2263 return AdjustCost(Entry->Cost); 2264 2265 if (ST->hasDQI()) 2266 if (const auto *Entry = ConvertCostTableLookup( 2267 AVX512DQConversionTbl, ISD, SimpleDstTy, SimpleSrcTy)) 2268 return AdjustCost(Entry->Cost); 2269 2270 if (ST->hasAVX512()) 2271 if (const auto *Entry = ConvertCostTableLookup( 2272 AVX512FConversionTbl, ISD, SimpleDstTy, SimpleSrcTy)) 2273 return AdjustCost(Entry->Cost); 2274 } 2275 2276 if (ST->hasBWI()) 2277 if (const auto *Entry = ConvertCostTableLookup( 2278 AVX512BWVLConversionTbl, ISD, SimpleDstTy, SimpleSrcTy)) 2279 return AdjustCost(Entry->Cost); 2280 2281 if (ST->hasDQI()) 2282 if (const auto *Entry = ConvertCostTableLookup( 2283 AVX512DQVLConversionTbl, ISD, SimpleDstTy, SimpleSrcTy)) 2284 return AdjustCost(Entry->Cost); 2285 2286 if (ST->hasAVX512()) 2287 if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD, 2288 SimpleDstTy, SimpleSrcTy)) 2289 return AdjustCost(Entry->Cost); 2290 2291 if (ST->hasAVX2()) { 2292 if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD, 2293 SimpleDstTy, SimpleSrcTy)) 2294 return AdjustCost(Entry->Cost); 2295 } 2296 2297 if (ST->hasAVX()) { 2298 if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD, 2299 SimpleDstTy, SimpleSrcTy)) 2300 return AdjustCost(Entry->Cost); 2301 } 2302 2303 if (ST->hasSSE41()) { 2304 if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD, 2305 SimpleDstTy, SimpleSrcTy)) 2306 return AdjustCost(Entry->Cost); 2307 } 2308 2309 if (ST->hasSSE2()) { 2310 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD, 2311 SimpleDstTy, SimpleSrcTy)) 2312 return AdjustCost(Entry->Cost); 2313 } 2314 } 2315 2316 // Fall back to legalized types. 2317 std::pair<InstructionCost, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src); 2318 std::pair<InstructionCost, MVT> LTDest = 2319 TLI->getTypeLegalizationCost(DL, Dst); 2320 2321 if (ST->useAVX512Regs()) { 2322 if (ST->hasBWI()) 2323 if (const auto *Entry = ConvertCostTableLookup( 2324 AVX512BWConversionTbl, ISD, LTDest.second, LTSrc.second)) 2325 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2326 2327 if (ST->hasDQI()) 2328 if (const auto *Entry = ConvertCostTableLookup( 2329 AVX512DQConversionTbl, ISD, LTDest.second, LTSrc.second)) 2330 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2331 2332 if (ST->hasAVX512()) 2333 if (const auto *Entry = ConvertCostTableLookup( 2334 AVX512FConversionTbl, ISD, LTDest.second, LTSrc.second)) 2335 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2336 } 2337 2338 if (ST->hasBWI()) 2339 if (const auto *Entry = ConvertCostTableLookup(AVX512BWVLConversionTbl, ISD, 2340 LTDest.second, LTSrc.second)) 2341 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2342 2343 if (ST->hasDQI()) 2344 if (const auto *Entry = ConvertCostTableLookup(AVX512DQVLConversionTbl, ISD, 2345 LTDest.second, LTSrc.second)) 2346 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2347 2348 if (ST->hasAVX512()) 2349 if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD, 2350 LTDest.second, LTSrc.second)) 2351 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2352 2353 if (ST->hasAVX2()) 2354 if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD, 2355 LTDest.second, LTSrc.second)) 2356 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2357 2358 if (ST->hasAVX()) 2359 if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD, 2360 LTDest.second, LTSrc.second)) 2361 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2362 2363 if (ST->hasSSE41()) 2364 if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD, 2365 LTDest.second, LTSrc.second)) 2366 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2367 2368 if (ST->hasSSE2()) 2369 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD, 2370 LTDest.second, LTSrc.second)) 2371 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost); 2372 2373 // Fallback, for i8/i16 sitofp/uitofp cases we need to extend to i32 for 2374 // sitofp. 2375 if ((ISD == ISD::SINT_TO_FP || ISD == ISD::UINT_TO_FP) && 2376 1 < Src->getScalarSizeInBits() && Src->getScalarSizeInBits() < 32) { 2377 Type *ExtSrc = Src->getWithNewBitWidth(32); 2378 unsigned ExtOpc = 2379 (ISD == ISD::SINT_TO_FP) ? Instruction::SExt : Instruction::ZExt; 2380 2381 // For scalar loads the extend would be free. 2382 InstructionCost ExtCost = 0; 2383 if (!(Src->isIntegerTy() && I && isa<LoadInst>(I->getOperand(0)))) 2384 ExtCost = getCastInstrCost(ExtOpc, ExtSrc, Src, CCH, CostKind); 2385 2386 return ExtCost + getCastInstrCost(Instruction::SIToFP, Dst, ExtSrc, 2387 TTI::CastContextHint::None, CostKind); 2388 } 2389 2390 // Fallback for fptosi/fptoui i8/i16 cases we need to truncate from fptosi 2391 // i32. 2392 if ((ISD == ISD::FP_TO_SINT || ISD == ISD::FP_TO_UINT) && 2393 1 < Dst->getScalarSizeInBits() && Dst->getScalarSizeInBits() < 32) { 2394 Type *TruncDst = Dst->getWithNewBitWidth(32); 2395 return getCastInstrCost(Instruction::FPToSI, TruncDst, Src, CCH, CostKind) + 2396 getCastInstrCost(Instruction::Trunc, Dst, TruncDst, 2397 TTI::CastContextHint::None, CostKind); 2398 } 2399 2400 return AdjustCost( 2401 BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I)); 2402 } 2403 2404 InstructionCost X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 2405 Type *CondTy, 2406 CmpInst::Predicate VecPred, 2407 TTI::TargetCostKind CostKind, 2408 const Instruction *I) { 2409 // TODO: Handle other cost kinds. 2410 if (CostKind != TTI::TCK_RecipThroughput) 2411 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, 2412 I); 2413 2414 // Legalize the type. 2415 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 2416 2417 MVT MTy = LT.second; 2418 2419 int ISD = TLI->InstructionOpcodeToISD(Opcode); 2420 assert(ISD && "Invalid opcode"); 2421 2422 unsigned ExtraCost = 0; 2423 if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) { 2424 // Some vector comparison predicates cost extra instructions. 2425 // TODO: Should we invert this and assume worst case cmp costs 2426 // and reduce for particular predicates? 2427 if (MTy.isVector() && 2428 !((ST->hasXOP() && (!ST->hasAVX2() || MTy.is128BitVector())) || 2429 (ST->hasAVX512() && 32 <= MTy.getScalarSizeInBits()) || 2430 ST->hasBWI())) { 2431 // Fallback to I if a specific predicate wasn't specified. 2432 CmpInst::Predicate Pred = VecPred; 2433 if (I && (Pred == CmpInst::BAD_ICMP_PREDICATE || 2434 Pred == CmpInst::BAD_FCMP_PREDICATE)) 2435 Pred = cast<CmpInst>(I)->getPredicate(); 2436 2437 switch (Pred) { 2438 case CmpInst::Predicate::ICMP_NE: 2439 // xor(cmpeq(x,y),-1) 2440 ExtraCost = 1; 2441 break; 2442 case CmpInst::Predicate::ICMP_SGE: 2443 case CmpInst::Predicate::ICMP_SLE: 2444 // xor(cmpgt(x,y),-1) 2445 ExtraCost = 1; 2446 break; 2447 case CmpInst::Predicate::ICMP_ULT: 2448 case CmpInst::Predicate::ICMP_UGT: 2449 // cmpgt(xor(x,signbit),xor(y,signbit)) 2450 // xor(cmpeq(pmaxu(x,y),x),-1) 2451 ExtraCost = 2; 2452 break; 2453 case CmpInst::Predicate::ICMP_ULE: 2454 case CmpInst::Predicate::ICMP_UGE: 2455 if ((ST->hasSSE41() && MTy.getScalarSizeInBits() == 32) || 2456 (ST->hasSSE2() && MTy.getScalarSizeInBits() < 32)) { 2457 // cmpeq(psubus(x,y),0) 2458 // cmpeq(pminu(x,y),x) 2459 ExtraCost = 1; 2460 } else { 2461 // xor(cmpgt(xor(x,signbit),xor(y,signbit)),-1) 2462 ExtraCost = 3; 2463 } 2464 break; 2465 case CmpInst::Predicate::BAD_ICMP_PREDICATE: 2466 case CmpInst::Predicate::BAD_FCMP_PREDICATE: 2467 // Assume worst case scenario and add the maximum extra cost. 2468 ExtraCost = 3; 2469 break; 2470 default: 2471 break; 2472 } 2473 } 2474 } 2475 2476 static const CostTblEntry SLMCostTbl[] = { 2477 // slm pcmpeq/pcmpgt throughput is 2 2478 { ISD::SETCC, MVT::v2i64, 2 }, 2479 }; 2480 2481 static const CostTblEntry AVX512BWCostTbl[] = { 2482 { ISD::SETCC, MVT::v32i16, 1 }, 2483 { ISD::SETCC, MVT::v64i8, 1 }, 2484 2485 { ISD::SELECT, MVT::v32i16, 1 }, 2486 { ISD::SELECT, MVT::v64i8, 1 }, 2487 }; 2488 2489 static const CostTblEntry AVX512CostTbl[] = { 2490 { ISD::SETCC, MVT::v8i64, 1 }, 2491 { ISD::SETCC, MVT::v16i32, 1 }, 2492 { ISD::SETCC, MVT::v8f64, 1 }, 2493 { ISD::SETCC, MVT::v16f32, 1 }, 2494 2495 { ISD::SELECT, MVT::v8i64, 1 }, 2496 { ISD::SELECT, MVT::v16i32, 1 }, 2497 { ISD::SELECT, MVT::v8f64, 1 }, 2498 { ISD::SELECT, MVT::v16f32, 1 }, 2499 2500 { ISD::SETCC, MVT::v32i16, 2 }, // FIXME: should probably be 4 2501 { ISD::SETCC, MVT::v64i8, 2 }, // FIXME: should probably be 4 2502 2503 { ISD::SELECT, MVT::v32i16, 2 }, // FIXME: should be 3 2504 { ISD::SELECT, MVT::v64i8, 2 }, // FIXME: should be 3 2505 }; 2506 2507 static const CostTblEntry AVX2CostTbl[] = { 2508 { ISD::SETCC, MVT::v4i64, 1 }, 2509 { ISD::SETCC, MVT::v8i32, 1 }, 2510 { ISD::SETCC, MVT::v16i16, 1 }, 2511 { ISD::SETCC, MVT::v32i8, 1 }, 2512 2513 { ISD::SELECT, MVT::v4i64, 1 }, // pblendvb 2514 { ISD::SELECT, MVT::v8i32, 1 }, // pblendvb 2515 { ISD::SELECT, MVT::v16i16, 1 }, // pblendvb 2516 { ISD::SELECT, MVT::v32i8, 1 }, // pblendvb 2517 }; 2518 2519 static const CostTblEntry AVX1CostTbl[] = { 2520 { ISD::SETCC, MVT::v4f64, 1 }, 2521 { ISD::SETCC, MVT::v8f32, 1 }, 2522 // AVX1 does not support 8-wide integer compare. 2523 { ISD::SETCC, MVT::v4i64, 4 }, 2524 { ISD::SETCC, MVT::v8i32, 4 }, 2525 { ISD::SETCC, MVT::v16i16, 4 }, 2526 { ISD::SETCC, MVT::v32i8, 4 }, 2527 2528 { ISD::SELECT, MVT::v4f64, 1 }, // vblendvpd 2529 { ISD::SELECT, MVT::v8f32, 1 }, // vblendvps 2530 { ISD::SELECT, MVT::v4i64, 1 }, // vblendvpd 2531 { ISD::SELECT, MVT::v8i32, 1 }, // vblendvps 2532 { ISD::SELECT, MVT::v16i16, 3 }, // vandps + vandnps + vorps 2533 { ISD::SELECT, MVT::v32i8, 3 }, // vandps + vandnps + vorps 2534 }; 2535 2536 static const CostTblEntry SSE42CostTbl[] = { 2537 { ISD::SETCC, MVT::v2f64, 1 }, 2538 { ISD::SETCC, MVT::v4f32, 1 }, 2539 { ISD::SETCC, MVT::v2i64, 1 }, 2540 }; 2541 2542 static const CostTblEntry SSE41CostTbl[] = { 2543 { ISD::SELECT, MVT::v2f64, 1 }, // blendvpd 2544 { ISD::SELECT, MVT::v4f32, 1 }, // blendvps 2545 { ISD::SELECT, MVT::v2i64, 1 }, // pblendvb 2546 { ISD::SELECT, MVT::v4i32, 1 }, // pblendvb 2547 { ISD::SELECT, MVT::v8i16, 1 }, // pblendvb 2548 { ISD::SELECT, MVT::v16i8, 1 }, // pblendvb 2549 }; 2550 2551 static const CostTblEntry SSE2CostTbl[] = { 2552 { ISD::SETCC, MVT::v2f64, 2 }, 2553 { ISD::SETCC, MVT::f64, 1 }, 2554 { ISD::SETCC, MVT::v2i64, 8 }, 2555 { ISD::SETCC, MVT::v4i32, 1 }, 2556 { ISD::SETCC, MVT::v8i16, 1 }, 2557 { ISD::SETCC, MVT::v16i8, 1 }, 2558 2559 { ISD::SELECT, MVT::v2f64, 3 }, // andpd + andnpd + orpd 2560 { ISD::SELECT, MVT::v2i64, 3 }, // pand + pandn + por 2561 { ISD::SELECT, MVT::v4i32, 3 }, // pand + pandn + por 2562 { ISD::SELECT, MVT::v8i16, 3 }, // pand + pandn + por 2563 { ISD::SELECT, MVT::v16i8, 3 }, // pand + pandn + por 2564 }; 2565 2566 static const CostTblEntry SSE1CostTbl[] = { 2567 { ISD::SETCC, MVT::v4f32, 2 }, 2568 { ISD::SETCC, MVT::f32, 1 }, 2569 2570 { ISD::SELECT, MVT::v4f32, 3 }, // andps + andnps + orps 2571 }; 2572 2573 if (ST->isSLM()) 2574 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy)) 2575 return LT.first * (ExtraCost + Entry->Cost); 2576 2577 if (ST->hasBWI()) 2578 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 2579 return LT.first * (ExtraCost + Entry->Cost); 2580 2581 if (ST->hasAVX512()) 2582 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 2583 return LT.first * (ExtraCost + Entry->Cost); 2584 2585 if (ST->hasAVX2()) 2586 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 2587 return LT.first * (ExtraCost + Entry->Cost); 2588 2589 if (ST->hasAVX()) 2590 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 2591 return LT.first * (ExtraCost + Entry->Cost); 2592 2593 if (ST->hasSSE42()) 2594 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 2595 return LT.first * (ExtraCost + Entry->Cost); 2596 2597 if (ST->hasSSE41()) 2598 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 2599 return LT.first * (ExtraCost + Entry->Cost); 2600 2601 if (ST->hasSSE2()) 2602 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 2603 return LT.first * (ExtraCost + Entry->Cost); 2604 2605 if (ST->hasSSE1()) 2606 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 2607 return LT.first * (ExtraCost + Entry->Cost); 2608 2609 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I); 2610 } 2611 2612 unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; } 2613 2614 InstructionCost 2615 X86TTIImpl::getTypeBasedIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, 2616 TTI::TargetCostKind CostKind) { 2617 2618 // Costs should match the codegen from: 2619 // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll 2620 // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll 2621 // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll 2622 // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll 2623 // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll 2624 2625 // TODO: Overflow intrinsics (*ADDO, *SUBO, *MULO) with vector types are not 2626 // specialized in these tables yet. 2627 static const CostTblEntry AVX512BITALGCostTbl[] = { 2628 { ISD::CTPOP, MVT::v32i16, 1 }, 2629 { ISD::CTPOP, MVT::v64i8, 1 }, 2630 { ISD::CTPOP, MVT::v16i16, 1 }, 2631 { ISD::CTPOP, MVT::v32i8, 1 }, 2632 { ISD::CTPOP, MVT::v8i16, 1 }, 2633 { ISD::CTPOP, MVT::v16i8, 1 }, 2634 }; 2635 static const CostTblEntry AVX512VPOPCNTDQCostTbl[] = { 2636 { ISD::CTPOP, MVT::v8i64, 1 }, 2637 { ISD::CTPOP, MVT::v16i32, 1 }, 2638 { ISD::CTPOP, MVT::v4i64, 1 }, 2639 { ISD::CTPOP, MVT::v8i32, 1 }, 2640 { ISD::CTPOP, MVT::v2i64, 1 }, 2641 { ISD::CTPOP, MVT::v4i32, 1 }, 2642 }; 2643 static const CostTblEntry AVX512CDCostTbl[] = { 2644 { ISD::CTLZ, MVT::v8i64, 1 }, 2645 { ISD::CTLZ, MVT::v16i32, 1 }, 2646 { ISD::CTLZ, MVT::v32i16, 8 }, 2647 { ISD::CTLZ, MVT::v64i8, 20 }, 2648 { ISD::CTLZ, MVT::v4i64, 1 }, 2649 { ISD::CTLZ, MVT::v8i32, 1 }, 2650 { ISD::CTLZ, MVT::v16i16, 4 }, 2651 { ISD::CTLZ, MVT::v32i8, 10 }, 2652 { ISD::CTLZ, MVT::v2i64, 1 }, 2653 { ISD::CTLZ, MVT::v4i32, 1 }, 2654 { ISD::CTLZ, MVT::v8i16, 4 }, 2655 { ISD::CTLZ, MVT::v16i8, 4 }, 2656 }; 2657 static const CostTblEntry AVX512BWCostTbl[] = { 2658 { ISD::ABS, MVT::v32i16, 1 }, 2659 { ISD::ABS, MVT::v64i8, 1 }, 2660 { ISD::BITREVERSE, MVT::v8i64, 3 }, 2661 { ISD::BITREVERSE, MVT::v16i32, 3 }, 2662 { ISD::BITREVERSE, MVT::v32i16, 3 }, 2663 { ISD::BITREVERSE, MVT::v64i8, 2 }, 2664 { ISD::BSWAP, MVT::v8i64, 1 }, 2665 { ISD::BSWAP, MVT::v16i32, 1 }, 2666 { ISD::BSWAP, MVT::v32i16, 1 }, 2667 { ISD::CTLZ, MVT::v8i64, 23 }, 2668 { ISD::CTLZ, MVT::v16i32, 22 }, 2669 { ISD::CTLZ, MVT::v32i16, 18 }, 2670 { ISD::CTLZ, MVT::v64i8, 17 }, 2671 { ISD::CTPOP, MVT::v8i64, 7 }, 2672 { ISD::CTPOP, MVT::v16i32, 11 }, 2673 { ISD::CTPOP, MVT::v32i16, 9 }, 2674 { ISD::CTPOP, MVT::v64i8, 6 }, 2675 { ISD::CTTZ, MVT::v8i64, 10 }, 2676 { ISD::CTTZ, MVT::v16i32, 14 }, 2677 { ISD::CTTZ, MVT::v32i16, 12 }, 2678 { ISD::CTTZ, MVT::v64i8, 9 }, 2679 { ISD::SADDSAT, MVT::v32i16, 1 }, 2680 { ISD::SADDSAT, MVT::v64i8, 1 }, 2681 { ISD::SMAX, MVT::v32i16, 1 }, 2682 { ISD::SMAX, MVT::v64i8, 1 }, 2683 { ISD::SMIN, MVT::v32i16, 1 }, 2684 { ISD::SMIN, MVT::v64i8, 1 }, 2685 { ISD::SSUBSAT, MVT::v32i16, 1 }, 2686 { ISD::SSUBSAT, MVT::v64i8, 1 }, 2687 { ISD::UADDSAT, MVT::v32i16, 1 }, 2688 { ISD::UADDSAT, MVT::v64i8, 1 }, 2689 { ISD::UMAX, MVT::v32i16, 1 }, 2690 { ISD::UMAX, MVT::v64i8, 1 }, 2691 { ISD::UMIN, MVT::v32i16, 1 }, 2692 { ISD::UMIN, MVT::v64i8, 1 }, 2693 { ISD::USUBSAT, MVT::v32i16, 1 }, 2694 { ISD::USUBSAT, MVT::v64i8, 1 }, 2695 }; 2696 static const CostTblEntry AVX512CostTbl[] = { 2697 { ISD::ABS, MVT::v8i64, 1 }, 2698 { ISD::ABS, MVT::v16i32, 1 }, 2699 { ISD::ABS, MVT::v32i16, 2 }, // FIXME: include split 2700 { ISD::ABS, MVT::v64i8, 2 }, // FIXME: include split 2701 { ISD::ABS, MVT::v4i64, 1 }, 2702 { ISD::ABS, MVT::v2i64, 1 }, 2703 { ISD::BITREVERSE, MVT::v8i64, 36 }, 2704 { ISD::BITREVERSE, MVT::v16i32, 24 }, 2705 { ISD::BITREVERSE, MVT::v32i16, 10 }, 2706 { ISD::BITREVERSE, MVT::v64i8, 10 }, 2707 { ISD::BSWAP, MVT::v8i64, 4 }, 2708 { ISD::BSWAP, MVT::v16i32, 4 }, 2709 { ISD::BSWAP, MVT::v32i16, 4 }, 2710 { ISD::CTLZ, MVT::v8i64, 29 }, 2711 { ISD::CTLZ, MVT::v16i32, 35 }, 2712 { ISD::CTLZ, MVT::v32i16, 28 }, 2713 { ISD::CTLZ, MVT::v64i8, 18 }, 2714 { ISD::CTPOP, MVT::v8i64, 16 }, 2715 { ISD::CTPOP, MVT::v16i32, 24 }, 2716 { ISD::CTPOP, MVT::v32i16, 18 }, 2717 { ISD::CTPOP, MVT::v64i8, 12 }, 2718 { ISD::CTTZ, MVT::v8i64, 20 }, 2719 { ISD::CTTZ, MVT::v16i32, 28 }, 2720 { ISD::CTTZ, MVT::v32i16, 24 }, 2721 { ISD::CTTZ, MVT::v64i8, 18 }, 2722 { ISD::SMAX, MVT::v8i64, 1 }, 2723 { ISD::SMAX, MVT::v16i32, 1 }, 2724 { ISD::SMAX, MVT::v32i16, 2 }, // FIXME: include split 2725 { ISD::SMAX, MVT::v64i8, 2 }, // FIXME: include split 2726 { ISD::SMAX, MVT::v4i64, 1 }, 2727 { ISD::SMAX, MVT::v2i64, 1 }, 2728 { ISD::SMIN, MVT::v8i64, 1 }, 2729 { ISD::SMIN, MVT::v16i32, 1 }, 2730 { ISD::SMIN, MVT::v32i16, 2 }, // FIXME: include split 2731 { ISD::SMIN, MVT::v64i8, 2 }, // FIXME: include split 2732 { ISD::SMIN, MVT::v4i64, 1 }, 2733 { ISD::SMIN, MVT::v2i64, 1 }, 2734 { ISD::UMAX, MVT::v8i64, 1 }, 2735 { ISD::UMAX, MVT::v16i32, 1 }, 2736 { ISD::UMAX, MVT::v32i16, 2 }, // FIXME: include split 2737 { ISD::UMAX, MVT::v64i8, 2 }, // FIXME: include split 2738 { ISD::UMAX, MVT::v4i64, 1 }, 2739 { ISD::UMAX, MVT::v2i64, 1 }, 2740 { ISD::UMIN, MVT::v8i64, 1 }, 2741 { ISD::UMIN, MVT::v16i32, 1 }, 2742 { ISD::UMIN, MVT::v32i16, 2 }, // FIXME: include split 2743 { ISD::UMIN, MVT::v64i8, 2 }, // FIXME: include split 2744 { ISD::UMIN, MVT::v4i64, 1 }, 2745 { ISD::UMIN, MVT::v2i64, 1 }, 2746 { ISD::USUBSAT, MVT::v16i32, 2 }, // pmaxud + psubd 2747 { ISD::USUBSAT, MVT::v2i64, 2 }, // pmaxuq + psubq 2748 { ISD::USUBSAT, MVT::v4i64, 2 }, // pmaxuq + psubq 2749 { ISD::USUBSAT, MVT::v8i64, 2 }, // pmaxuq + psubq 2750 { ISD::UADDSAT, MVT::v16i32, 3 }, // not + pminud + paddd 2751 { ISD::UADDSAT, MVT::v2i64, 3 }, // not + pminuq + paddq 2752 { ISD::UADDSAT, MVT::v4i64, 3 }, // not + pminuq + paddq 2753 { ISD::UADDSAT, MVT::v8i64, 3 }, // not + pminuq + paddq 2754 { ISD::SADDSAT, MVT::v32i16, 2 }, // FIXME: include split 2755 { ISD::SADDSAT, MVT::v64i8, 2 }, // FIXME: include split 2756 { ISD::SSUBSAT, MVT::v32i16, 2 }, // FIXME: include split 2757 { ISD::SSUBSAT, MVT::v64i8, 2 }, // FIXME: include split 2758 { ISD::UADDSAT, MVT::v32i16, 2 }, // FIXME: include split 2759 { ISD::UADDSAT, MVT::v64i8, 2 }, // FIXME: include split 2760 { ISD::USUBSAT, MVT::v32i16, 2 }, // FIXME: include split 2761 { ISD::USUBSAT, MVT::v64i8, 2 }, // FIXME: include split 2762 { ISD::FMAXNUM, MVT::f32, 2 }, 2763 { ISD::FMAXNUM, MVT::v4f32, 2 }, 2764 { ISD::FMAXNUM, MVT::v8f32, 2 }, 2765 { ISD::FMAXNUM, MVT::v16f32, 2 }, 2766 { ISD::FMAXNUM, MVT::f64, 2 }, 2767 { ISD::FMAXNUM, MVT::v2f64, 2 }, 2768 { ISD::FMAXNUM, MVT::v4f64, 2 }, 2769 { ISD::FMAXNUM, MVT::v8f64, 2 }, 2770 }; 2771 static const CostTblEntry XOPCostTbl[] = { 2772 { ISD::BITREVERSE, MVT::v4i64, 4 }, 2773 { ISD::BITREVERSE, MVT::v8i32, 4 }, 2774 { ISD::BITREVERSE, MVT::v16i16, 4 }, 2775 { ISD::BITREVERSE, MVT::v32i8, 4 }, 2776 { ISD::BITREVERSE, MVT::v2i64, 1 }, 2777 { ISD::BITREVERSE, MVT::v4i32, 1 }, 2778 { ISD::BITREVERSE, MVT::v8i16, 1 }, 2779 { ISD::BITREVERSE, MVT::v16i8, 1 }, 2780 { ISD::BITREVERSE, MVT::i64, 3 }, 2781 { ISD::BITREVERSE, MVT::i32, 3 }, 2782 { ISD::BITREVERSE, MVT::i16, 3 }, 2783 { ISD::BITREVERSE, MVT::i8, 3 } 2784 }; 2785 static const CostTblEntry AVX2CostTbl[] = { 2786 { ISD::ABS, MVT::v4i64, 2 }, // VBLENDVPD(X,VPSUBQ(0,X),X) 2787 { ISD::ABS, MVT::v8i32, 1 }, 2788 { ISD::ABS, MVT::v16i16, 1 }, 2789 { ISD::ABS, MVT::v32i8, 1 }, 2790 { ISD::BITREVERSE, MVT::v2i64, 3 }, 2791 { ISD::BITREVERSE, MVT::v4i64, 3 }, 2792 { ISD::BITREVERSE, MVT::v4i32, 3 }, 2793 { ISD::BITREVERSE, MVT::v8i32, 3 }, 2794 { ISD::BITREVERSE, MVT::v8i16, 3 }, 2795 { ISD::BITREVERSE, MVT::v16i16, 3 }, 2796 { ISD::BITREVERSE, MVT::v16i8, 3 }, 2797 { ISD::BITREVERSE, MVT::v32i8, 3 }, 2798 { ISD::BSWAP, MVT::v4i64, 1 }, 2799 { ISD::BSWAP, MVT::v8i32, 1 }, 2800 { ISD::BSWAP, MVT::v16i16, 1 }, 2801 { ISD::CTLZ, MVT::v2i64, 7 }, 2802 { ISD::CTLZ, MVT::v4i64, 7 }, 2803 { ISD::CTLZ, MVT::v4i32, 5 }, 2804 { ISD::CTLZ, MVT::v8i32, 5 }, 2805 { ISD::CTLZ, MVT::v8i16, 4 }, 2806 { ISD::CTLZ, MVT::v16i16, 4 }, 2807 { ISD::CTLZ, MVT::v16i8, 3 }, 2808 { ISD::CTLZ, MVT::v32i8, 3 }, 2809 { ISD::CTPOP, MVT::v2i64, 3 }, 2810 { ISD::CTPOP, MVT::v4i64, 3 }, 2811 { ISD::CTPOP, MVT::v4i32, 7 }, 2812 { ISD::CTPOP, MVT::v8i32, 7 }, 2813 { ISD::CTPOP, MVT::v8i16, 3 }, 2814 { ISD::CTPOP, MVT::v16i16, 3 }, 2815 { ISD::CTPOP, MVT::v16i8, 2 }, 2816 { ISD::CTPOP, MVT::v32i8, 2 }, 2817 { ISD::CTTZ, MVT::v2i64, 4 }, 2818 { ISD::CTTZ, MVT::v4i64, 4 }, 2819 { ISD::CTTZ, MVT::v4i32, 7 }, 2820 { ISD::CTTZ, MVT::v8i32, 7 }, 2821 { ISD::CTTZ, MVT::v8i16, 4 }, 2822 { ISD::CTTZ, MVT::v16i16, 4 }, 2823 { ISD::CTTZ, MVT::v16i8, 3 }, 2824 { ISD::CTTZ, MVT::v32i8, 3 }, 2825 { ISD::SADDSAT, MVT::v16i16, 1 }, 2826 { ISD::SADDSAT, MVT::v32i8, 1 }, 2827 { ISD::SMAX, MVT::v8i32, 1 }, 2828 { ISD::SMAX, MVT::v16i16, 1 }, 2829 { ISD::SMAX, MVT::v32i8, 1 }, 2830 { ISD::SMIN, MVT::v8i32, 1 }, 2831 { ISD::SMIN, MVT::v16i16, 1 }, 2832 { ISD::SMIN, MVT::v32i8, 1 }, 2833 { ISD::SSUBSAT, MVT::v16i16, 1 }, 2834 { ISD::SSUBSAT, MVT::v32i8, 1 }, 2835 { ISD::UADDSAT, MVT::v16i16, 1 }, 2836 { ISD::UADDSAT, MVT::v32i8, 1 }, 2837 { ISD::UADDSAT, MVT::v8i32, 3 }, // not + pminud + paddd 2838 { ISD::UMAX, MVT::v8i32, 1 }, 2839 { ISD::UMAX, MVT::v16i16, 1 }, 2840 { ISD::UMAX, MVT::v32i8, 1 }, 2841 { ISD::UMIN, MVT::v8i32, 1 }, 2842 { ISD::UMIN, MVT::v16i16, 1 }, 2843 { ISD::UMIN, MVT::v32i8, 1 }, 2844 { ISD::USUBSAT, MVT::v16i16, 1 }, 2845 { ISD::USUBSAT, MVT::v32i8, 1 }, 2846 { ISD::USUBSAT, MVT::v8i32, 2 }, // pmaxud + psubd 2847 { ISD::FMAXNUM, MVT::v8f32, 3 }, // MAXPS + CMPUNORDPS + BLENDVPS 2848 { ISD::FMAXNUM, MVT::v4f64, 3 }, // MAXPD + CMPUNORDPD + BLENDVPD 2849 { ISD::FSQRT, MVT::f32, 7 }, // Haswell from http://www.agner.org/ 2850 { ISD::FSQRT, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/ 2851 { ISD::FSQRT, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/ 2852 { ISD::FSQRT, MVT::f64, 14 }, // Haswell from http://www.agner.org/ 2853 { ISD::FSQRT, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/ 2854 { ISD::FSQRT, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/ 2855 }; 2856 static const CostTblEntry AVX1CostTbl[] = { 2857 { ISD::ABS, MVT::v4i64, 5 }, // VBLENDVPD(X,VPSUBQ(0,X),X) 2858 { ISD::ABS, MVT::v8i32, 3 }, 2859 { ISD::ABS, MVT::v16i16, 3 }, 2860 { ISD::ABS, MVT::v32i8, 3 }, 2861 { ISD::BITREVERSE, MVT::v4i64, 12 }, // 2 x 128-bit Op + extract/insert 2862 { ISD::BITREVERSE, MVT::v8i32, 12 }, // 2 x 128-bit Op + extract/insert 2863 { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert 2864 { ISD::BITREVERSE, MVT::v32i8, 12 }, // 2 x 128-bit Op + extract/insert 2865 { ISD::BSWAP, MVT::v4i64, 4 }, 2866 { ISD::BSWAP, MVT::v8i32, 4 }, 2867 { ISD::BSWAP, MVT::v16i16, 4 }, 2868 { ISD::CTLZ, MVT::v4i64, 48 }, // 2 x 128-bit Op + extract/insert 2869 { ISD::CTLZ, MVT::v8i32, 38 }, // 2 x 128-bit Op + extract/insert 2870 { ISD::CTLZ, MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert 2871 { ISD::CTLZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert 2872 { ISD::CTPOP, MVT::v4i64, 16 }, // 2 x 128-bit Op + extract/insert 2873 { ISD::CTPOP, MVT::v8i32, 24 }, // 2 x 128-bit Op + extract/insert 2874 { ISD::CTPOP, MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert 2875 { ISD::CTPOP, MVT::v32i8, 14 }, // 2 x 128-bit Op + extract/insert 2876 { ISD::CTTZ, MVT::v4i64, 22 }, // 2 x 128-bit Op + extract/insert 2877 { ISD::CTTZ, MVT::v8i32, 30 }, // 2 x 128-bit Op + extract/insert 2878 { ISD::CTTZ, MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert 2879 { ISD::CTTZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert 2880 { ISD::SADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2881 { ISD::SADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2882 { ISD::SMAX, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2883 { ISD::SMAX, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2884 { ISD::SMAX, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2885 { ISD::SMIN, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2886 { ISD::SMIN, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2887 { ISD::SMIN, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2888 { ISD::SSUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2889 { ISD::SSUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2890 { ISD::UADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2891 { ISD::UADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2892 { ISD::UADDSAT, MVT::v8i32, 8 }, // 2 x 128-bit Op + extract/insert 2893 { ISD::UMAX, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2894 { ISD::UMAX, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2895 { ISD::UMAX, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2896 { ISD::UMIN, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert 2897 { ISD::UMIN, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2898 { ISD::UMIN, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2899 { ISD::USUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert 2900 { ISD::USUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert 2901 { ISD::USUBSAT, MVT::v8i32, 6 }, // 2 x 128-bit Op + extract/insert 2902 { ISD::FMAXNUM, MVT::f32, 3 }, // MAXSS + CMPUNORDSS + BLENDVPS 2903 { ISD::FMAXNUM, MVT::v4f32, 3 }, // MAXPS + CMPUNORDPS + BLENDVPS 2904 { ISD::FMAXNUM, MVT::v8f32, 5 }, // MAXPS + CMPUNORDPS + BLENDVPS + ? 2905 { ISD::FMAXNUM, MVT::f64, 3 }, // MAXSD + CMPUNORDSD + BLENDVPD 2906 { ISD::FMAXNUM, MVT::v2f64, 3 }, // MAXPD + CMPUNORDPD + BLENDVPD 2907 { ISD::FMAXNUM, MVT::v4f64, 5 }, // MAXPD + CMPUNORDPD + BLENDVPD + ? 2908 { ISD::FSQRT, MVT::f32, 14 }, // SNB from http://www.agner.org/ 2909 { ISD::FSQRT, MVT::v4f32, 14 }, // SNB from http://www.agner.org/ 2910 { ISD::FSQRT, MVT::v8f32, 28 }, // SNB from http://www.agner.org/ 2911 { ISD::FSQRT, MVT::f64, 21 }, // SNB from http://www.agner.org/ 2912 { ISD::FSQRT, MVT::v2f64, 21 }, // SNB from http://www.agner.org/ 2913 { ISD::FSQRT, MVT::v4f64, 43 }, // SNB from http://www.agner.org/ 2914 }; 2915 static const CostTblEntry GLMCostTbl[] = { 2916 { ISD::FSQRT, MVT::f32, 19 }, // sqrtss 2917 { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps 2918 { ISD::FSQRT, MVT::f64, 34 }, // sqrtsd 2919 { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd 2920 }; 2921 static const CostTblEntry SLMCostTbl[] = { 2922 { ISD::FSQRT, MVT::f32, 20 }, // sqrtss 2923 { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps 2924 { ISD::FSQRT, MVT::f64, 35 }, // sqrtsd 2925 { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd 2926 }; 2927 static const CostTblEntry SSE42CostTbl[] = { 2928 { ISD::USUBSAT, MVT::v4i32, 2 }, // pmaxud + psubd 2929 { ISD::UADDSAT, MVT::v4i32, 3 }, // not + pminud + paddd 2930 { ISD::FSQRT, MVT::f32, 18 }, // Nehalem from http://www.agner.org/ 2931 { ISD::FSQRT, MVT::v4f32, 18 }, // Nehalem from http://www.agner.org/ 2932 }; 2933 static const CostTblEntry SSE41CostTbl[] = { 2934 { ISD::ABS, MVT::v2i64, 2 }, // BLENDVPD(X,PSUBQ(0,X),X) 2935 { ISD::SMAX, MVT::v4i32, 1 }, 2936 { ISD::SMAX, MVT::v16i8, 1 }, 2937 { ISD::SMIN, MVT::v4i32, 1 }, 2938 { ISD::SMIN, MVT::v16i8, 1 }, 2939 { ISD::UMAX, MVT::v4i32, 1 }, 2940 { ISD::UMAX, MVT::v8i16, 1 }, 2941 { ISD::UMIN, MVT::v4i32, 1 }, 2942 { ISD::UMIN, MVT::v8i16, 1 }, 2943 }; 2944 static const CostTblEntry SSSE3CostTbl[] = { 2945 { ISD::ABS, MVT::v4i32, 1 }, 2946 { ISD::ABS, MVT::v8i16, 1 }, 2947 { ISD::ABS, MVT::v16i8, 1 }, 2948 { ISD::BITREVERSE, MVT::v2i64, 5 }, 2949 { ISD::BITREVERSE, MVT::v4i32, 5 }, 2950 { ISD::BITREVERSE, MVT::v8i16, 5 }, 2951 { ISD::BITREVERSE, MVT::v16i8, 5 }, 2952 { ISD::BSWAP, MVT::v2i64, 1 }, 2953 { ISD::BSWAP, MVT::v4i32, 1 }, 2954 { ISD::BSWAP, MVT::v8i16, 1 }, 2955 { ISD::CTLZ, MVT::v2i64, 23 }, 2956 { ISD::CTLZ, MVT::v4i32, 18 }, 2957 { ISD::CTLZ, MVT::v8i16, 14 }, 2958 { ISD::CTLZ, MVT::v16i8, 9 }, 2959 { ISD::CTPOP, MVT::v2i64, 7 }, 2960 { ISD::CTPOP, MVT::v4i32, 11 }, 2961 { ISD::CTPOP, MVT::v8i16, 9 }, 2962 { ISD::CTPOP, MVT::v16i8, 6 }, 2963 { ISD::CTTZ, MVT::v2i64, 10 }, 2964 { ISD::CTTZ, MVT::v4i32, 14 }, 2965 { ISD::CTTZ, MVT::v8i16, 12 }, 2966 { ISD::CTTZ, MVT::v16i8, 9 } 2967 }; 2968 static const CostTblEntry SSE2CostTbl[] = { 2969 { ISD::ABS, MVT::v2i64, 4 }, 2970 { ISD::ABS, MVT::v4i32, 3 }, 2971 { ISD::ABS, MVT::v8i16, 2 }, 2972 { ISD::ABS, MVT::v16i8, 2 }, 2973 { ISD::BITREVERSE, MVT::v2i64, 29 }, 2974 { ISD::BITREVERSE, MVT::v4i32, 27 }, 2975 { ISD::BITREVERSE, MVT::v8i16, 27 }, 2976 { ISD::BITREVERSE, MVT::v16i8, 20 }, 2977 { ISD::BSWAP, MVT::v2i64, 7 }, 2978 { ISD::BSWAP, MVT::v4i32, 7 }, 2979 { ISD::BSWAP, MVT::v8i16, 7 }, 2980 { ISD::CTLZ, MVT::v2i64, 25 }, 2981 { ISD::CTLZ, MVT::v4i32, 26 }, 2982 { ISD::CTLZ, MVT::v8i16, 20 }, 2983 { ISD::CTLZ, MVT::v16i8, 17 }, 2984 { ISD::CTPOP, MVT::v2i64, 12 }, 2985 { ISD::CTPOP, MVT::v4i32, 15 }, 2986 { ISD::CTPOP, MVT::v8i16, 13 }, 2987 { ISD::CTPOP, MVT::v16i8, 10 }, 2988 { ISD::CTTZ, MVT::v2i64, 14 }, 2989 { ISD::CTTZ, MVT::v4i32, 18 }, 2990 { ISD::CTTZ, MVT::v8i16, 16 }, 2991 { ISD::CTTZ, MVT::v16i8, 13 }, 2992 { ISD::SADDSAT, MVT::v8i16, 1 }, 2993 { ISD::SADDSAT, MVT::v16i8, 1 }, 2994 { ISD::SMAX, MVT::v8i16, 1 }, 2995 { ISD::SMIN, MVT::v8i16, 1 }, 2996 { ISD::SSUBSAT, MVT::v8i16, 1 }, 2997 { ISD::SSUBSAT, MVT::v16i8, 1 }, 2998 { ISD::UADDSAT, MVT::v8i16, 1 }, 2999 { ISD::UADDSAT, MVT::v16i8, 1 }, 3000 { ISD::UMAX, MVT::v8i16, 2 }, 3001 { ISD::UMAX, MVT::v16i8, 1 }, 3002 { ISD::UMIN, MVT::v8i16, 2 }, 3003 { ISD::UMIN, MVT::v16i8, 1 }, 3004 { ISD::USUBSAT, MVT::v8i16, 1 }, 3005 { ISD::USUBSAT, MVT::v16i8, 1 }, 3006 { ISD::FMAXNUM, MVT::f64, 4 }, 3007 { ISD::FMAXNUM, MVT::v2f64, 4 }, 3008 { ISD::FSQRT, MVT::f64, 32 }, // Nehalem from http://www.agner.org/ 3009 { ISD::FSQRT, MVT::v2f64, 32 }, // Nehalem from http://www.agner.org/ 3010 }; 3011 static const CostTblEntry SSE1CostTbl[] = { 3012 { ISD::FMAXNUM, MVT::f32, 4 }, 3013 { ISD::FMAXNUM, MVT::v4f32, 4 }, 3014 { ISD::FSQRT, MVT::f32, 28 }, // Pentium III from http://www.agner.org/ 3015 { ISD::FSQRT, MVT::v4f32, 56 }, // Pentium III from http://www.agner.org/ 3016 }; 3017 static const CostTblEntry BMI64CostTbl[] = { // 64-bit targets 3018 { ISD::CTTZ, MVT::i64, 1 }, 3019 }; 3020 static const CostTblEntry BMI32CostTbl[] = { // 32 or 64-bit targets 3021 { ISD::CTTZ, MVT::i32, 1 }, 3022 { ISD::CTTZ, MVT::i16, 1 }, 3023 { ISD::CTTZ, MVT::i8, 1 }, 3024 }; 3025 static const CostTblEntry LZCNT64CostTbl[] = { // 64-bit targets 3026 { ISD::CTLZ, MVT::i64, 1 }, 3027 }; 3028 static const CostTblEntry LZCNT32CostTbl[] = { // 32 or 64-bit targets 3029 { ISD::CTLZ, MVT::i32, 1 }, 3030 { ISD::CTLZ, MVT::i16, 1 }, 3031 { ISD::CTLZ, MVT::i8, 1 }, 3032 }; 3033 static const CostTblEntry POPCNT64CostTbl[] = { // 64-bit targets 3034 { ISD::CTPOP, MVT::i64, 1 }, 3035 }; 3036 static const CostTblEntry POPCNT32CostTbl[] = { // 32 or 64-bit targets 3037 { ISD::CTPOP, MVT::i32, 1 }, 3038 { ISD::CTPOP, MVT::i16, 1 }, 3039 { ISD::CTPOP, MVT::i8, 1 }, 3040 }; 3041 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 3042 { ISD::ABS, MVT::i64, 2 }, // SUB+CMOV 3043 { ISD::BITREVERSE, MVT::i64, 14 }, 3044 { ISD::BSWAP, MVT::i64, 1 }, 3045 { ISD::CTLZ, MVT::i64, 4 }, // BSR+XOR or BSR+XOR+CMOV 3046 { ISD::CTTZ, MVT::i64, 3 }, // TEST+BSF+CMOV/BRANCH 3047 { ISD::CTPOP, MVT::i64, 10 }, 3048 { ISD::SADDO, MVT::i64, 1 }, 3049 { ISD::UADDO, MVT::i64, 1 }, 3050 { ISD::UMULO, MVT::i64, 2 }, // mulq + seto 3051 }; 3052 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 3053 { ISD::ABS, MVT::i32, 2 }, // SUB+CMOV 3054 { ISD::ABS, MVT::i16, 2 }, // SUB+CMOV 3055 { ISD::BITREVERSE, MVT::i32, 14 }, 3056 { ISD::BITREVERSE, MVT::i16, 14 }, 3057 { ISD::BITREVERSE, MVT::i8, 11 }, 3058 { ISD::BSWAP, MVT::i32, 1 }, 3059 { ISD::BSWAP, MVT::i16, 1 }, // ROL 3060 { ISD::CTLZ, MVT::i32, 4 }, // BSR+XOR or BSR+XOR+CMOV 3061 { ISD::CTLZ, MVT::i16, 4 }, // BSR+XOR or BSR+XOR+CMOV 3062 { ISD::CTLZ, MVT::i8, 4 }, // BSR+XOR or BSR+XOR+CMOV 3063 { ISD::CTTZ, MVT::i32, 3 }, // TEST+BSF+CMOV/BRANCH 3064 { ISD::CTTZ, MVT::i16, 3 }, // TEST+BSF+CMOV/BRANCH 3065 { ISD::CTTZ, MVT::i8, 3 }, // TEST+BSF+CMOV/BRANCH 3066 { ISD::CTPOP, MVT::i32, 8 }, 3067 { ISD::CTPOP, MVT::i16, 9 }, 3068 { ISD::CTPOP, MVT::i8, 7 }, 3069 { ISD::SADDO, MVT::i32, 1 }, 3070 { ISD::SADDO, MVT::i16, 1 }, 3071 { ISD::SADDO, MVT::i8, 1 }, 3072 { ISD::UADDO, MVT::i32, 1 }, 3073 { ISD::UADDO, MVT::i16, 1 }, 3074 { ISD::UADDO, MVT::i8, 1 }, 3075 { ISD::UMULO, MVT::i32, 2 }, // mul + seto 3076 { ISD::UMULO, MVT::i16, 2 }, 3077 { ISD::UMULO, MVT::i8, 2 }, 3078 }; 3079 3080 Type *RetTy = ICA.getReturnType(); 3081 Type *OpTy = RetTy; 3082 Intrinsic::ID IID = ICA.getID(); 3083 unsigned ISD = ISD::DELETED_NODE; 3084 switch (IID) { 3085 default: 3086 break; 3087 case Intrinsic::abs: 3088 ISD = ISD::ABS; 3089 break; 3090 case Intrinsic::bitreverse: 3091 ISD = ISD::BITREVERSE; 3092 break; 3093 case Intrinsic::bswap: 3094 ISD = ISD::BSWAP; 3095 break; 3096 case Intrinsic::ctlz: 3097 ISD = ISD::CTLZ; 3098 break; 3099 case Intrinsic::ctpop: 3100 ISD = ISD::CTPOP; 3101 break; 3102 case Intrinsic::cttz: 3103 ISD = ISD::CTTZ; 3104 break; 3105 case Intrinsic::maxnum: 3106 case Intrinsic::minnum: 3107 // FMINNUM has same costs so don't duplicate. 3108 ISD = ISD::FMAXNUM; 3109 break; 3110 case Intrinsic::sadd_sat: 3111 ISD = ISD::SADDSAT; 3112 break; 3113 case Intrinsic::smax: 3114 ISD = ISD::SMAX; 3115 break; 3116 case Intrinsic::smin: 3117 ISD = ISD::SMIN; 3118 break; 3119 case Intrinsic::ssub_sat: 3120 ISD = ISD::SSUBSAT; 3121 break; 3122 case Intrinsic::uadd_sat: 3123 ISD = ISD::UADDSAT; 3124 break; 3125 case Intrinsic::umax: 3126 ISD = ISD::UMAX; 3127 break; 3128 case Intrinsic::umin: 3129 ISD = ISD::UMIN; 3130 break; 3131 case Intrinsic::usub_sat: 3132 ISD = ISD::USUBSAT; 3133 break; 3134 case Intrinsic::sqrt: 3135 ISD = ISD::FSQRT; 3136 break; 3137 case Intrinsic::sadd_with_overflow: 3138 case Intrinsic::ssub_with_overflow: 3139 // SSUBO has same costs so don't duplicate. 3140 ISD = ISD::SADDO; 3141 OpTy = RetTy->getContainedType(0); 3142 break; 3143 case Intrinsic::uadd_with_overflow: 3144 case Intrinsic::usub_with_overflow: 3145 // USUBO has same costs so don't duplicate. 3146 ISD = ISD::UADDO; 3147 OpTy = RetTy->getContainedType(0); 3148 break; 3149 case Intrinsic::umul_with_overflow: 3150 case Intrinsic::smul_with_overflow: 3151 // SMULO has same costs so don't duplicate. 3152 ISD = ISD::UMULO; 3153 OpTy = RetTy->getContainedType(0); 3154 break; 3155 } 3156 3157 if (ISD != ISD::DELETED_NODE) { 3158 // Legalize the type. 3159 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, OpTy); 3160 MVT MTy = LT.second; 3161 3162 // Attempt to lookup cost. 3163 if (ISD == ISD::BITREVERSE && ST->hasGFNI() && ST->hasSSSE3() && 3164 MTy.isVector()) { 3165 // With PSHUFB the code is very similar for all types. If we have integer 3166 // byte operations, we just need a GF2P8AFFINEQB for vXi8. For other types 3167 // we also need a PSHUFB. 3168 unsigned Cost = MTy.getVectorElementType() == MVT::i8 ? 1 : 2; 3169 3170 // Without byte operations, we need twice as many GF2P8AFFINEQB and PSHUFB 3171 // instructions. We also need an extract and an insert. 3172 if (!(MTy.is128BitVector() || (ST->hasAVX2() && MTy.is256BitVector()) || 3173 (ST->hasBWI() && MTy.is512BitVector()))) 3174 Cost = Cost * 2 + 2; 3175 3176 return LT.first * Cost; 3177 } 3178 3179 auto adjustTableCost = [](const CostTblEntry &Entry, 3180 InstructionCost LegalizationCost, 3181 FastMathFlags FMF) { 3182 // If there are no NANs to deal with, then these are reduced to a 3183 // single MIN** or MAX** instruction instead of the MIN/CMP/SELECT that we 3184 // assume is used in the non-fast case. 3185 if (Entry.ISD == ISD::FMAXNUM || Entry.ISD == ISD::FMINNUM) { 3186 if (FMF.noNaNs()) 3187 return LegalizationCost * 1; 3188 } 3189 return LegalizationCost * (int)Entry.Cost; 3190 }; 3191 3192 if (ST->useGLMDivSqrtCosts()) 3193 if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy)) 3194 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3195 3196 if (ST->isSLM()) 3197 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy)) 3198 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3199 3200 if (ST->hasBITALG()) 3201 if (const auto *Entry = CostTableLookup(AVX512BITALGCostTbl, ISD, MTy)) 3202 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3203 3204 if (ST->hasVPOPCNTDQ()) 3205 if (const auto *Entry = CostTableLookup(AVX512VPOPCNTDQCostTbl, ISD, MTy)) 3206 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3207 3208 if (ST->hasCDI()) 3209 if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy)) 3210 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3211 3212 if (ST->hasBWI()) 3213 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 3214 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3215 3216 if (ST->hasAVX512()) 3217 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 3218 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3219 3220 if (ST->hasXOP()) 3221 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy)) 3222 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3223 3224 if (ST->hasAVX2()) 3225 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 3226 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3227 3228 if (ST->hasAVX()) 3229 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 3230 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3231 3232 if (ST->hasSSE42()) 3233 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 3234 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3235 3236 if (ST->hasSSE41()) 3237 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 3238 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3239 3240 if (ST->hasSSSE3()) 3241 if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy)) 3242 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3243 3244 if (ST->hasSSE2()) 3245 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 3246 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3247 3248 if (ST->hasSSE1()) 3249 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 3250 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3251 3252 if (ST->hasBMI()) { 3253 if (ST->is64Bit()) 3254 if (const auto *Entry = CostTableLookup(BMI64CostTbl, ISD, MTy)) 3255 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3256 3257 if (const auto *Entry = CostTableLookup(BMI32CostTbl, ISD, MTy)) 3258 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3259 } 3260 3261 if (ST->hasLZCNT()) { 3262 if (ST->is64Bit()) 3263 if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy)) 3264 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3265 3266 if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy)) 3267 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3268 } 3269 3270 if (ST->hasPOPCNT()) { 3271 if (ST->is64Bit()) 3272 if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy)) 3273 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3274 3275 if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy)) 3276 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3277 } 3278 3279 if (ISD == ISD::BSWAP && ST->hasMOVBE() && ST->hasFastMOVBE()) { 3280 if (const Instruction *II = ICA.getInst()) { 3281 if (II->hasOneUse() && isa<StoreInst>(II->user_back())) 3282 return TTI::TCC_Free; 3283 if (auto *LI = dyn_cast<LoadInst>(II->getOperand(0))) { 3284 if (LI->hasOneUse()) 3285 return TTI::TCC_Free; 3286 } 3287 } 3288 } 3289 3290 // TODO - add BMI (TZCNT) scalar handling 3291 3292 if (ST->is64Bit()) 3293 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy)) 3294 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3295 3296 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy)) 3297 return adjustTableCost(*Entry, LT.first, ICA.getFlags()); 3298 } 3299 3300 return BaseT::getIntrinsicInstrCost(ICA, CostKind); 3301 } 3302 3303 InstructionCost 3304 X86TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, 3305 TTI::TargetCostKind CostKind) { 3306 if (ICA.isTypeBasedOnly()) 3307 return getTypeBasedIntrinsicInstrCost(ICA, CostKind); 3308 3309 static const CostTblEntry AVX512CostTbl[] = { 3310 { ISD::ROTL, MVT::v8i64, 1 }, 3311 { ISD::ROTL, MVT::v4i64, 1 }, 3312 { ISD::ROTL, MVT::v2i64, 1 }, 3313 { ISD::ROTL, MVT::v16i32, 1 }, 3314 { ISD::ROTL, MVT::v8i32, 1 }, 3315 { ISD::ROTL, MVT::v4i32, 1 }, 3316 { ISD::ROTR, MVT::v8i64, 1 }, 3317 { ISD::ROTR, MVT::v4i64, 1 }, 3318 { ISD::ROTR, MVT::v2i64, 1 }, 3319 { ISD::ROTR, MVT::v16i32, 1 }, 3320 { ISD::ROTR, MVT::v8i32, 1 }, 3321 { ISD::ROTR, MVT::v4i32, 1 } 3322 }; 3323 // XOP: ROTL = VPROT(X,Y), ROTR = VPROT(X,SUB(0,Y)) 3324 static const CostTblEntry XOPCostTbl[] = { 3325 { ISD::ROTL, MVT::v4i64, 4 }, 3326 { ISD::ROTL, MVT::v8i32, 4 }, 3327 { ISD::ROTL, MVT::v16i16, 4 }, 3328 { ISD::ROTL, MVT::v32i8, 4 }, 3329 { ISD::ROTL, MVT::v2i64, 1 }, 3330 { ISD::ROTL, MVT::v4i32, 1 }, 3331 { ISD::ROTL, MVT::v8i16, 1 }, 3332 { ISD::ROTL, MVT::v16i8, 1 }, 3333 { ISD::ROTR, MVT::v4i64, 6 }, 3334 { ISD::ROTR, MVT::v8i32, 6 }, 3335 { ISD::ROTR, MVT::v16i16, 6 }, 3336 { ISD::ROTR, MVT::v32i8, 6 }, 3337 { ISD::ROTR, MVT::v2i64, 2 }, 3338 { ISD::ROTR, MVT::v4i32, 2 }, 3339 { ISD::ROTR, MVT::v8i16, 2 }, 3340 { ISD::ROTR, MVT::v16i8, 2 } 3341 }; 3342 static const CostTblEntry X64CostTbl[] = { // 64-bit targets 3343 { ISD::ROTL, MVT::i64, 1 }, 3344 { ISD::ROTR, MVT::i64, 1 }, 3345 { ISD::FSHL, MVT::i64, 4 } 3346 }; 3347 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets 3348 { ISD::ROTL, MVT::i32, 1 }, 3349 { ISD::ROTL, MVT::i16, 1 }, 3350 { ISD::ROTL, MVT::i8, 1 }, 3351 { ISD::ROTR, MVT::i32, 1 }, 3352 { ISD::ROTR, MVT::i16, 1 }, 3353 { ISD::ROTR, MVT::i8, 1 }, 3354 { ISD::FSHL, MVT::i32, 4 }, 3355 { ISD::FSHL, MVT::i16, 4 }, 3356 { ISD::FSHL, MVT::i8, 4 } 3357 }; 3358 3359 Intrinsic::ID IID = ICA.getID(); 3360 Type *RetTy = ICA.getReturnType(); 3361 const SmallVectorImpl<const Value *> &Args = ICA.getArgs(); 3362 unsigned ISD = ISD::DELETED_NODE; 3363 switch (IID) { 3364 default: 3365 break; 3366 case Intrinsic::fshl: 3367 ISD = ISD::FSHL; 3368 if (Args[0] == Args[1]) 3369 ISD = ISD::ROTL; 3370 break; 3371 case Intrinsic::fshr: 3372 // FSHR has same costs so don't duplicate. 3373 ISD = ISD::FSHL; 3374 if (Args[0] == Args[1]) 3375 ISD = ISD::ROTR; 3376 break; 3377 } 3378 3379 if (ISD != ISD::DELETED_NODE) { 3380 // Legalize the type. 3381 std::pair<InstructionCost, MVT> LT = 3382 TLI->getTypeLegalizationCost(DL, RetTy); 3383 MVT MTy = LT.second; 3384 3385 // Attempt to lookup cost. 3386 if (ST->hasAVX512()) 3387 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 3388 return LT.first * Entry->Cost; 3389 3390 if (ST->hasXOP()) 3391 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy)) 3392 return LT.first * Entry->Cost; 3393 3394 if (ST->is64Bit()) 3395 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy)) 3396 return LT.first * Entry->Cost; 3397 3398 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy)) 3399 return LT.first * Entry->Cost; 3400 } 3401 3402 return BaseT::getIntrinsicInstrCost(ICA, CostKind); 3403 } 3404 3405 InstructionCost X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, 3406 unsigned Index) { 3407 static const CostTblEntry SLMCostTbl[] = { 3408 { ISD::EXTRACT_VECTOR_ELT, MVT::i8, 4 }, 3409 { ISD::EXTRACT_VECTOR_ELT, MVT::i16, 4 }, 3410 { ISD::EXTRACT_VECTOR_ELT, MVT::i32, 4 }, 3411 { ISD::EXTRACT_VECTOR_ELT, MVT::i64, 7 } 3412 }; 3413 3414 assert(Val->isVectorTy() && "This must be a vector type"); 3415 Type *ScalarType = Val->getScalarType(); 3416 int RegisterFileMoveCost = 0; 3417 3418 // Non-immediate extraction/insertion can be handled as a sequence of 3419 // aliased loads+stores via the stack. 3420 if (Index == -1U && (Opcode == Instruction::ExtractElement || 3421 Opcode == Instruction::InsertElement)) { 3422 // TODO: On some SSE41+ targets, we expand to cmp+splat+select patterns: 3423 // inselt N0, N1, N2 --> select (SplatN2 == {0,1,2...}) ? SplatN1 : N0. 3424 3425 // TODO: Move this to BasicTTIImpl.h? We'd need better gep + index handling. 3426 assert(isa<FixedVectorType>(Val) && "Fixed vector type expected"); 3427 Align VecAlign = DL.getPrefTypeAlign(Val); 3428 Align SclAlign = DL.getPrefTypeAlign(ScalarType); 3429 3430 // Extract - store vector to stack, load scalar. 3431 if (Opcode == Instruction::ExtractElement) { 3432 return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0, 3433 TTI::TargetCostKind::TCK_RecipThroughput) + 3434 getMemoryOpCost(Instruction::Load, ScalarType, SclAlign, 0, 3435 TTI::TargetCostKind::TCK_RecipThroughput); 3436 } 3437 // Insert - store vector to stack, store scalar, load vector. 3438 if (Opcode == Instruction::InsertElement) { 3439 return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0, 3440 TTI::TargetCostKind::TCK_RecipThroughput) + 3441 getMemoryOpCost(Instruction::Store, ScalarType, SclAlign, 0, 3442 TTI::TargetCostKind::TCK_RecipThroughput) + 3443 getMemoryOpCost(Instruction::Load, Val, VecAlign, 0, 3444 TTI::TargetCostKind::TCK_RecipThroughput); 3445 } 3446 } 3447 3448 if (Index != -1U && (Opcode == Instruction::ExtractElement || 3449 Opcode == Instruction::InsertElement)) { 3450 // Legalize the type. 3451 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Val); 3452 3453 // This type is legalized to a scalar type. 3454 if (!LT.second.isVector()) 3455 return 0; 3456 3457 // The type may be split. Normalize the index to the new type. 3458 unsigned NumElts = LT.second.getVectorNumElements(); 3459 unsigned SubNumElts = NumElts; 3460 Index = Index % NumElts; 3461 3462 // For >128-bit vectors, we need to extract higher 128-bit subvectors. 3463 // For inserts, we also need to insert the subvector back. 3464 if (LT.second.getSizeInBits() > 128) { 3465 assert((LT.second.getSizeInBits() % 128) == 0 && "Illegal vector"); 3466 unsigned NumSubVecs = LT.second.getSizeInBits() / 128; 3467 SubNumElts = NumElts / NumSubVecs; 3468 if (SubNumElts <= Index) { 3469 RegisterFileMoveCost += (Opcode == Instruction::InsertElement ? 2 : 1); 3470 Index %= SubNumElts; 3471 } 3472 } 3473 3474 if (Index == 0) { 3475 // Floating point scalars are already located in index #0. 3476 // Many insertions to #0 can fold away for scalar fp-ops, so let's assume 3477 // true for all. 3478 if (ScalarType->isFloatingPointTy()) 3479 return RegisterFileMoveCost; 3480 3481 // Assume movd/movq XMM -> GPR is relatively cheap on all targets. 3482 if (ScalarType->isIntegerTy() && Opcode == Instruction::ExtractElement) 3483 return 1 + RegisterFileMoveCost; 3484 } 3485 3486 int ISD = TLI->InstructionOpcodeToISD(Opcode); 3487 assert(ISD && "Unexpected vector opcode"); 3488 MVT MScalarTy = LT.second.getScalarType(); 3489 if (ST->isSLM()) 3490 if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy)) 3491 return Entry->Cost + RegisterFileMoveCost; 3492 3493 // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets. 3494 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) || 3495 (MScalarTy.isInteger() && ST->hasSSE41())) 3496 return 1 + RegisterFileMoveCost; 3497 3498 // Assume insertps is relatively cheap on all targets. 3499 if (MScalarTy == MVT::f32 && ST->hasSSE41() && 3500 Opcode == Instruction::InsertElement) 3501 return 1 + RegisterFileMoveCost; 3502 3503 // For extractions we just need to shuffle the element to index 0, which 3504 // should be very cheap (assume cost = 1). For insertions we need to shuffle 3505 // the elements to its destination. In both cases we must handle the 3506 // subvector move(s). 3507 // If the vector type is already less than 128-bits then don't reduce it. 3508 // TODO: Under what circumstances should we shuffle using the full width? 3509 InstructionCost ShuffleCost = 1; 3510 if (Opcode == Instruction::InsertElement) { 3511 auto *SubTy = cast<VectorType>(Val); 3512 EVT VT = TLI->getValueType(DL, Val); 3513 if (VT.getScalarType() != MScalarTy || VT.getSizeInBits() >= 128) 3514 SubTy = FixedVectorType::get(ScalarType, SubNumElts); 3515 ShuffleCost = 3516 getShuffleCost(TTI::SK_PermuteTwoSrc, SubTy, None, 0, SubTy); 3517 } 3518 int IntOrFpCost = ScalarType->isFloatingPointTy() ? 0 : 1; 3519 return ShuffleCost + IntOrFpCost + RegisterFileMoveCost; 3520 } 3521 3522 // Add to the base cost if we know that the extracted element of a vector is 3523 // destined to be moved to and used in the integer register file. 3524 if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy()) 3525 RegisterFileMoveCost += 1; 3526 3527 return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost; 3528 } 3529 3530 InstructionCost X86TTIImpl::getScalarizationOverhead(VectorType *Ty, 3531 const APInt &DemandedElts, 3532 bool Insert, 3533 bool Extract) { 3534 InstructionCost Cost = 0; 3535 3536 // For insertions, a ISD::BUILD_VECTOR style vector initialization can be much 3537 // cheaper than an accumulation of ISD::INSERT_VECTOR_ELT. 3538 if (Insert) { 3539 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 3540 MVT MScalarTy = LT.second.getScalarType(); 3541 3542 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) || 3543 (MScalarTy.isInteger() && ST->hasSSE41()) || 3544 (MScalarTy == MVT::f32 && ST->hasSSE41())) { 3545 // For types we can insert directly, insertion into 128-bit sub vectors is 3546 // cheap, followed by a cheap chain of concatenations. 3547 if (LT.second.getSizeInBits() <= 128) { 3548 Cost += 3549 BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, false); 3550 } else { 3551 // In each 128-lane, if at least one index is demanded but not all 3552 // indices are demanded and this 128-lane is not the first 128-lane of 3553 // the legalized-vector, then this 128-lane needs a extracti128; If in 3554 // each 128-lane, there is at least one demanded index, this 128-lane 3555 // needs a inserti128. 3556 3557 // The following cases will help you build a better understanding: 3558 // Assume we insert several elements into a v8i32 vector in avx2, 3559 // Case#1: inserting into 1th index needs vpinsrd + inserti128. 3560 // Case#2: inserting into 5th index needs extracti128 + vpinsrd + 3561 // inserti128. 3562 // Case#3: inserting into 4,5,6,7 index needs 4*vpinsrd + inserti128. 3563 const int CostValue = *LT.first.getValue(); 3564 assert(CostValue >= 0 && "Negative cost!"); 3565 unsigned Num128Lanes = LT.second.getSizeInBits() / 128 * CostValue; 3566 unsigned NumElts = LT.second.getVectorNumElements() * CostValue; 3567 APInt WidenedDemandedElts = DemandedElts.zextOrSelf(NumElts); 3568 unsigned Scale = NumElts / Num128Lanes; 3569 // We iterate each 128-lane, and check if we need a 3570 // extracti128/inserti128 for this 128-lane. 3571 for (unsigned I = 0; I < NumElts; I += Scale) { 3572 APInt Mask = WidenedDemandedElts.getBitsSet(NumElts, I, I + Scale); 3573 APInt MaskedDE = Mask & WidenedDemandedElts; 3574 unsigned Population = MaskedDE.countPopulation(); 3575 Cost += (Population > 0 && Population != Scale && 3576 I % LT.second.getVectorNumElements() != 0); 3577 Cost += Population > 0; 3578 } 3579 Cost += DemandedElts.countPopulation(); 3580 3581 // For vXf32 cases, insertion into the 0'th index in each v4f32 3582 // 128-bit vector is free. 3583 // NOTE: This assumes legalization widens vXf32 vectors. 3584 if (MScalarTy == MVT::f32) 3585 for (unsigned i = 0, e = cast<FixedVectorType>(Ty)->getNumElements(); 3586 i < e; i += 4) 3587 if (DemandedElts[i]) 3588 Cost--; 3589 } 3590 } else if (LT.second.isVector()) { 3591 // Without fast insertion, we need to use MOVD/MOVQ to pass each demanded 3592 // integer element as a SCALAR_TO_VECTOR, then we build the vector as a 3593 // series of UNPCK followed by CONCAT_VECTORS - all of these can be 3594 // considered cheap. 3595 if (Ty->isIntOrIntVectorTy()) 3596 Cost += DemandedElts.countPopulation(); 3597 3598 // Get the smaller of the legalized or original pow2-extended number of 3599 // vector elements, which represents the number of unpacks we'll end up 3600 // performing. 3601 unsigned NumElts = LT.second.getVectorNumElements(); 3602 unsigned Pow2Elts = 3603 PowerOf2Ceil(cast<FixedVectorType>(Ty)->getNumElements()); 3604 Cost += (std::min<unsigned>(NumElts, Pow2Elts) - 1) * LT.first; 3605 } 3606 } 3607 3608 // TODO: Use default extraction for now, but we should investigate extending this 3609 // to handle repeated subvector extraction. 3610 if (Extract) 3611 Cost += BaseT::getScalarizationOverhead(Ty, DemandedElts, false, Extract); 3612 3613 return Cost; 3614 } 3615 3616 InstructionCost X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, 3617 MaybeAlign Alignment, 3618 unsigned AddressSpace, 3619 TTI::TargetCostKind CostKind, 3620 const Instruction *I) { 3621 // TODO: Handle other cost kinds. 3622 if (CostKind != TTI::TCK_RecipThroughput) { 3623 if (auto *SI = dyn_cast_or_null<StoreInst>(I)) { 3624 // Store instruction with index and scale costs 2 Uops. 3625 // Check the preceding GEP to identify non-const indices. 3626 if (auto *GEP = dyn_cast<GetElementPtrInst>(SI->getPointerOperand())) { 3627 if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); })) 3628 return TTI::TCC_Basic * 2; 3629 } 3630 } 3631 return TTI::TCC_Basic; 3632 } 3633 3634 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) && 3635 "Invalid Opcode"); 3636 // Type legalization can't handle structs 3637 if (TLI->getValueType(DL, Src, true) == MVT::Other) 3638 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, 3639 CostKind); 3640 3641 // Legalize the type. 3642 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Src); 3643 3644 auto *VTy = dyn_cast<FixedVectorType>(Src); 3645 3646 // Handle the simple case of non-vectors. 3647 // NOTE: this assumes that legalization never creates vector from scalars! 3648 if (!VTy || !LT.second.isVector()) 3649 // Each load/store unit costs 1. 3650 return LT.first * 1; 3651 3652 bool IsLoad = Opcode == Instruction::Load; 3653 3654 Type *EltTy = VTy->getElementType(); 3655 3656 const int EltTyBits = DL.getTypeSizeInBits(EltTy); 3657 3658 InstructionCost Cost = 0; 3659 3660 // Source of truth: how many elements were there in the original IR vector? 3661 const unsigned SrcNumElt = VTy->getNumElements(); 3662 3663 // How far have we gotten? 3664 int NumEltRemaining = SrcNumElt; 3665 // Note that we intentionally capture by-reference, NumEltRemaining changes. 3666 auto NumEltDone = [&]() { return SrcNumElt - NumEltRemaining; }; 3667 3668 const int MaxLegalOpSizeBytes = divideCeil(LT.second.getSizeInBits(), 8); 3669 3670 // Note that even if we can store 64 bits of an XMM, we still operate on XMM. 3671 const unsigned XMMBits = 128; 3672 if (XMMBits % EltTyBits != 0) 3673 // Vector size must be a multiple of the element size. I.e. no padding. 3674 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, 3675 CostKind); 3676 const int NumEltPerXMM = XMMBits / EltTyBits; 3677 3678 auto *XMMVecTy = FixedVectorType::get(EltTy, NumEltPerXMM); 3679 3680 for (int CurrOpSizeBytes = MaxLegalOpSizeBytes, SubVecEltsLeft = 0; 3681 NumEltRemaining > 0; CurrOpSizeBytes /= 2) { 3682 // How many elements would a single op deal with at once? 3683 if ((8 * CurrOpSizeBytes) % EltTyBits != 0) 3684 // Vector size must be a multiple of the element size. I.e. no padding. 3685 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, 3686 CostKind); 3687 int CurrNumEltPerOp = (8 * CurrOpSizeBytes) / EltTyBits; 3688 3689 assert(CurrOpSizeBytes > 0 && CurrNumEltPerOp > 0 && "How'd we get here?"); 3690 assert((((NumEltRemaining * EltTyBits) < (2 * 8 * CurrOpSizeBytes)) || 3691 (CurrOpSizeBytes == MaxLegalOpSizeBytes)) && 3692 "Unless we haven't halved the op size yet, " 3693 "we have less than two op's sized units of work left."); 3694 3695 auto *CurrVecTy = CurrNumEltPerOp > NumEltPerXMM 3696 ? FixedVectorType::get(EltTy, CurrNumEltPerOp) 3697 : XMMVecTy; 3698 3699 assert(CurrVecTy->getNumElements() % CurrNumEltPerOp == 0 && 3700 "After halving sizes, the vector elt count is no longer a multiple " 3701 "of number of elements per operation?"); 3702 auto *CoalescedVecTy = 3703 CurrNumEltPerOp == 1 3704 ? CurrVecTy 3705 : FixedVectorType::get( 3706 IntegerType::get(Src->getContext(), 3707 EltTyBits * CurrNumEltPerOp), 3708 CurrVecTy->getNumElements() / CurrNumEltPerOp); 3709 assert(DL.getTypeSizeInBits(CoalescedVecTy) == 3710 DL.getTypeSizeInBits(CurrVecTy) && 3711 "coalesciing elements doesn't change vector width."); 3712 3713 while (NumEltRemaining > 0) { 3714 assert(SubVecEltsLeft >= 0 && "Subreg element count overconsumtion?"); 3715 3716 // Can we use this vector size, as per the remaining element count? 3717 // Iff the vector is naturally aligned, we can do a wide load regardless. 3718 if (NumEltRemaining < CurrNumEltPerOp && 3719 (!IsLoad || Alignment.valueOrOne() < CurrOpSizeBytes) && 3720 CurrOpSizeBytes != 1) 3721 break; // Try smalled vector size. 3722 3723 bool Is0thSubVec = (NumEltDone() % LT.second.getVectorNumElements()) == 0; 3724 3725 // If we have fully processed the previous reg, we need to replenish it. 3726 if (SubVecEltsLeft == 0) { 3727 SubVecEltsLeft += CurrVecTy->getNumElements(); 3728 // And that's free only for the 0'th subvector of a legalized vector. 3729 if (!Is0thSubVec) 3730 Cost += getShuffleCost(IsLoad ? TTI::ShuffleKind::SK_InsertSubvector 3731 : TTI::ShuffleKind::SK_ExtractSubvector, 3732 VTy, None, NumEltDone(), CurrVecTy); 3733 } 3734 3735 // While we can directly load/store ZMM, YMM, and 64-bit halves of XMM, 3736 // for smaller widths (32/16/8) we have to insert/extract them separately. 3737 // Again, it's free for the 0'th subreg (if op is 32/64 bit wide, 3738 // but let's pretend that it is also true for 16/8 bit wide ops...) 3739 if (CurrOpSizeBytes <= 32 / 8 && !Is0thSubVec) { 3740 int NumEltDoneInCurrXMM = NumEltDone() % NumEltPerXMM; 3741 assert(NumEltDoneInCurrXMM % CurrNumEltPerOp == 0 && ""); 3742 int CoalescedVecEltIdx = NumEltDoneInCurrXMM / CurrNumEltPerOp; 3743 APInt DemandedElts = 3744 APInt::getBitsSet(CoalescedVecTy->getNumElements(), 3745 CoalescedVecEltIdx, CoalescedVecEltIdx + 1); 3746 assert(DemandedElts.countPopulation() == 1 && "Inserting single value"); 3747 Cost += getScalarizationOverhead(CoalescedVecTy, DemandedElts, IsLoad, 3748 !IsLoad); 3749 } 3750 3751 // This isn't exactly right. We're using slow unaligned 32-byte accesses 3752 // as a proxy for a double-pumped AVX memory interface such as on 3753 // Sandybridge. 3754 if (CurrOpSizeBytes == 32 && ST->isUnalignedMem32Slow()) 3755 Cost += 2; 3756 else 3757 Cost += 1; 3758 3759 SubVecEltsLeft -= CurrNumEltPerOp; 3760 NumEltRemaining -= CurrNumEltPerOp; 3761 Alignment = commonAlignment(Alignment.valueOrOne(), CurrOpSizeBytes); 3762 } 3763 } 3764 3765 assert(NumEltRemaining <= 0 && "Should have processed all the elements."); 3766 3767 return Cost; 3768 } 3769 3770 InstructionCost 3771 X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy, Align Alignment, 3772 unsigned AddressSpace, 3773 TTI::TargetCostKind CostKind) { 3774 bool IsLoad = (Instruction::Load == Opcode); 3775 bool IsStore = (Instruction::Store == Opcode); 3776 3777 auto *SrcVTy = dyn_cast<FixedVectorType>(SrcTy); 3778 if (!SrcVTy) 3779 // To calculate scalar take the regular cost, without mask 3780 return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace, CostKind); 3781 3782 unsigned NumElem = SrcVTy->getNumElements(); 3783 auto *MaskTy = 3784 FixedVectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem); 3785 if ((IsLoad && !isLegalMaskedLoad(SrcVTy, Alignment)) || 3786 (IsStore && !isLegalMaskedStore(SrcVTy, Alignment))) { 3787 // Scalarization 3788 APInt DemandedElts = APInt::getAllOnes(NumElem); 3789 InstructionCost MaskSplitCost = 3790 getScalarizationOverhead(MaskTy, DemandedElts, false, true); 3791 InstructionCost ScalarCompareCost = getCmpSelInstrCost( 3792 Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr, 3793 CmpInst::BAD_ICMP_PREDICATE, CostKind); 3794 InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind); 3795 InstructionCost MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost); 3796 InstructionCost ValueSplitCost = 3797 getScalarizationOverhead(SrcVTy, DemandedElts, IsLoad, IsStore); 3798 InstructionCost MemopCost = 3799 NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 3800 Alignment, AddressSpace, CostKind); 3801 return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost; 3802 } 3803 3804 // Legalize the type. 3805 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy); 3806 auto VT = TLI->getValueType(DL, SrcVTy); 3807 InstructionCost Cost = 0; 3808 if (VT.isSimple() && LT.second != VT.getSimpleVT() && 3809 LT.second.getVectorNumElements() == NumElem) 3810 // Promotion requires extend/truncate for data and a shuffle for mask. 3811 Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVTy, None, 0, nullptr) + 3812 getShuffleCost(TTI::SK_PermuteTwoSrc, MaskTy, None, 0, nullptr); 3813 3814 else if (LT.first * LT.second.getVectorNumElements() > NumElem) { 3815 auto *NewMaskTy = FixedVectorType::get(MaskTy->getElementType(), 3816 LT.second.getVectorNumElements()); 3817 // Expanding requires fill mask with zeroes 3818 Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, None, 0, MaskTy); 3819 } 3820 3821 // Pre-AVX512 - each maskmov load costs 2 + store costs ~8. 3822 if (!ST->hasAVX512()) 3823 return Cost + LT.first * (IsLoad ? 2 : 8); 3824 3825 // AVX-512 masked load/store is cheapper 3826 return Cost + LT.first; 3827 } 3828 3829 InstructionCost X86TTIImpl::getAddressComputationCost(Type *Ty, 3830 ScalarEvolution *SE, 3831 const SCEV *Ptr) { 3832 // Address computations in vectorized code with non-consecutive addresses will 3833 // likely result in more instructions compared to scalar code where the 3834 // computation can more often be merged into the index mode. The resulting 3835 // extra micro-ops can significantly decrease throughput. 3836 const unsigned NumVectorInstToHideOverhead = 10; 3837 3838 // Cost modeling of Strided Access Computation is hidden by the indexing 3839 // modes of X86 regardless of the stride value. We dont believe that there 3840 // is a difference between constant strided access in gerenal and constant 3841 // strided value which is less than or equal to 64. 3842 // Even in the case of (loop invariant) stride whose value is not known at 3843 // compile time, the address computation will not incur more than one extra 3844 // ADD instruction. 3845 if (Ty->isVectorTy() && SE) { 3846 if (!BaseT::isStridedAccess(Ptr)) 3847 return NumVectorInstToHideOverhead; 3848 if (!BaseT::getConstantStrideStep(SE, Ptr)) 3849 return 1; 3850 } 3851 3852 return BaseT::getAddressComputationCost(Ty, SE, Ptr); 3853 } 3854 3855 InstructionCost 3856 X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy, 3857 Optional<FastMathFlags> FMF, 3858 TTI::TargetCostKind CostKind) { 3859 if (TTI::requiresOrderedReduction(FMF)) 3860 return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind); 3861 3862 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 3863 // and make it as the cost. 3864 3865 static const CostTblEntry SLMCostTblNoPairWise[] = { 3866 { ISD::FADD, MVT::v2f64, 3 }, 3867 { ISD::ADD, MVT::v2i64, 5 }, 3868 }; 3869 3870 static const CostTblEntry SSE2CostTblNoPairWise[] = { 3871 { ISD::FADD, MVT::v2f64, 2 }, 3872 { ISD::FADD, MVT::v2f32, 2 }, 3873 { ISD::FADD, MVT::v4f32, 4 }, 3874 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6". 3875 { ISD::ADD, MVT::v2i32, 2 }, // FIXME: chosen to be less than v4i32 3876 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3". 3877 { ISD::ADD, MVT::v2i16, 2 }, // The data reported by the IACA tool is "4.3". 3878 { ISD::ADD, MVT::v4i16, 3 }, // The data reported by the IACA tool is "4.3". 3879 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3". 3880 { ISD::ADD, MVT::v2i8, 2 }, 3881 { ISD::ADD, MVT::v4i8, 2 }, 3882 { ISD::ADD, MVT::v8i8, 2 }, 3883 { ISD::ADD, MVT::v16i8, 3 }, 3884 }; 3885 3886 static const CostTblEntry AVX1CostTblNoPairWise[] = { 3887 { ISD::FADD, MVT::v4f64, 3 }, 3888 { ISD::FADD, MVT::v4f32, 3 }, 3889 { ISD::FADD, MVT::v8f32, 4 }, 3890 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5". 3891 { ISD::ADD, MVT::v4i64, 3 }, 3892 { ISD::ADD, MVT::v8i32, 5 }, 3893 { ISD::ADD, MVT::v16i16, 5 }, 3894 { ISD::ADD, MVT::v32i8, 4 }, 3895 }; 3896 3897 int ISD = TLI->InstructionOpcodeToISD(Opcode); 3898 assert(ISD && "Invalid opcode"); 3899 3900 // Before legalizing the type, give a chance to look up illegal narrow types 3901 // in the table. 3902 // FIXME: Is there a better way to do this? 3903 EVT VT = TLI->getValueType(DL, ValTy); 3904 if (VT.isSimple()) { 3905 MVT MTy = VT.getSimpleVT(); 3906 if (ST->isSLM()) 3907 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy)) 3908 return Entry->Cost; 3909 3910 if (ST->hasAVX()) 3911 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3912 return Entry->Cost; 3913 3914 if (ST->hasSSE2()) 3915 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3916 return Entry->Cost; 3917 } 3918 3919 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 3920 3921 MVT MTy = LT.second; 3922 3923 auto *ValVTy = cast<FixedVectorType>(ValTy); 3924 3925 // Special case: vXi8 mul reductions are performed as vXi16. 3926 if (ISD == ISD::MUL && MTy.getScalarType() == MVT::i8) { 3927 auto *WideSclTy = IntegerType::get(ValVTy->getContext(), 16); 3928 auto *WideVecTy = FixedVectorType::get(WideSclTy, ValVTy->getNumElements()); 3929 return getCastInstrCost(Instruction::ZExt, WideVecTy, ValTy, 3930 TargetTransformInfo::CastContextHint::None, 3931 CostKind) + 3932 getArithmeticReductionCost(Opcode, WideVecTy, FMF, CostKind); 3933 } 3934 3935 InstructionCost ArithmeticCost = 0; 3936 if (LT.first != 1 && MTy.isVector() && 3937 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 3938 // Type needs to be split. We need LT.first - 1 arithmetic ops. 3939 auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(), 3940 MTy.getVectorNumElements()); 3941 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind); 3942 ArithmeticCost *= LT.first - 1; 3943 } 3944 3945 if (ST->isSLM()) 3946 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy)) 3947 return ArithmeticCost + Entry->Cost; 3948 3949 if (ST->hasAVX()) 3950 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 3951 return ArithmeticCost + Entry->Cost; 3952 3953 if (ST->hasSSE2()) 3954 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 3955 return ArithmeticCost + Entry->Cost; 3956 3957 // FIXME: These assume a naive kshift+binop lowering, which is probably 3958 // conservative in most cases. 3959 static const CostTblEntry AVX512BoolReduction[] = { 3960 { ISD::AND, MVT::v2i1, 3 }, 3961 { ISD::AND, MVT::v4i1, 5 }, 3962 { ISD::AND, MVT::v8i1, 7 }, 3963 { ISD::AND, MVT::v16i1, 9 }, 3964 { ISD::AND, MVT::v32i1, 11 }, 3965 { ISD::AND, MVT::v64i1, 13 }, 3966 { ISD::OR, MVT::v2i1, 3 }, 3967 { ISD::OR, MVT::v4i1, 5 }, 3968 { ISD::OR, MVT::v8i1, 7 }, 3969 { ISD::OR, MVT::v16i1, 9 }, 3970 { ISD::OR, MVT::v32i1, 11 }, 3971 { ISD::OR, MVT::v64i1, 13 }, 3972 }; 3973 3974 static const CostTblEntry AVX2BoolReduction[] = { 3975 { ISD::AND, MVT::v16i16, 2 }, // vpmovmskb + cmp 3976 { ISD::AND, MVT::v32i8, 2 }, // vpmovmskb + cmp 3977 { ISD::OR, MVT::v16i16, 2 }, // vpmovmskb + cmp 3978 { ISD::OR, MVT::v32i8, 2 }, // vpmovmskb + cmp 3979 }; 3980 3981 static const CostTblEntry AVX1BoolReduction[] = { 3982 { ISD::AND, MVT::v4i64, 2 }, // vmovmskpd + cmp 3983 { ISD::AND, MVT::v8i32, 2 }, // vmovmskps + cmp 3984 { ISD::AND, MVT::v16i16, 4 }, // vextractf128 + vpand + vpmovmskb + cmp 3985 { ISD::AND, MVT::v32i8, 4 }, // vextractf128 + vpand + vpmovmskb + cmp 3986 { ISD::OR, MVT::v4i64, 2 }, // vmovmskpd + cmp 3987 { ISD::OR, MVT::v8i32, 2 }, // vmovmskps + cmp 3988 { ISD::OR, MVT::v16i16, 4 }, // vextractf128 + vpor + vpmovmskb + cmp 3989 { ISD::OR, MVT::v32i8, 4 }, // vextractf128 + vpor + vpmovmskb + cmp 3990 }; 3991 3992 static const CostTblEntry SSE2BoolReduction[] = { 3993 { ISD::AND, MVT::v2i64, 2 }, // movmskpd + cmp 3994 { ISD::AND, MVT::v4i32, 2 }, // movmskps + cmp 3995 { ISD::AND, MVT::v8i16, 2 }, // pmovmskb + cmp 3996 { ISD::AND, MVT::v16i8, 2 }, // pmovmskb + cmp 3997 { ISD::OR, MVT::v2i64, 2 }, // movmskpd + cmp 3998 { ISD::OR, MVT::v4i32, 2 }, // movmskps + cmp 3999 { ISD::OR, MVT::v8i16, 2 }, // pmovmskb + cmp 4000 { ISD::OR, MVT::v16i8, 2 }, // pmovmskb + cmp 4001 }; 4002 4003 // Handle bool allof/anyof patterns. 4004 if (ValVTy->getElementType()->isIntegerTy(1)) { 4005 InstructionCost ArithmeticCost = 0; 4006 if (LT.first != 1 && MTy.isVector() && 4007 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 4008 // Type needs to be split. We need LT.first - 1 arithmetic ops. 4009 auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(), 4010 MTy.getVectorNumElements()); 4011 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind); 4012 ArithmeticCost *= LT.first - 1; 4013 } 4014 4015 if (ST->hasAVX512()) 4016 if (const auto *Entry = CostTableLookup(AVX512BoolReduction, ISD, MTy)) 4017 return ArithmeticCost + Entry->Cost; 4018 if (ST->hasAVX2()) 4019 if (const auto *Entry = CostTableLookup(AVX2BoolReduction, ISD, MTy)) 4020 return ArithmeticCost + Entry->Cost; 4021 if (ST->hasAVX()) 4022 if (const auto *Entry = CostTableLookup(AVX1BoolReduction, ISD, MTy)) 4023 return ArithmeticCost + Entry->Cost; 4024 if (ST->hasSSE2()) 4025 if (const auto *Entry = CostTableLookup(SSE2BoolReduction, ISD, MTy)) 4026 return ArithmeticCost + Entry->Cost; 4027 4028 return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind); 4029 } 4030 4031 unsigned NumVecElts = ValVTy->getNumElements(); 4032 unsigned ScalarSize = ValVTy->getScalarSizeInBits(); 4033 4034 // Special case power of 2 reductions where the scalar type isn't changed 4035 // by type legalization. 4036 if (!isPowerOf2_32(NumVecElts) || ScalarSize != MTy.getScalarSizeInBits()) 4037 return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind); 4038 4039 InstructionCost ReductionCost = 0; 4040 4041 auto *Ty = ValVTy; 4042 if (LT.first != 1 && MTy.isVector() && 4043 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 4044 // Type needs to be split. We need LT.first - 1 arithmetic ops. 4045 Ty = FixedVectorType::get(ValVTy->getElementType(), 4046 MTy.getVectorNumElements()); 4047 ReductionCost = getArithmeticInstrCost(Opcode, Ty, CostKind); 4048 ReductionCost *= LT.first - 1; 4049 NumVecElts = MTy.getVectorNumElements(); 4050 } 4051 4052 // Now handle reduction with the legal type, taking into account size changes 4053 // at each level. 4054 while (NumVecElts > 1) { 4055 // Determine the size of the remaining vector we need to reduce. 4056 unsigned Size = NumVecElts * ScalarSize; 4057 NumVecElts /= 2; 4058 // If we're reducing from 256/512 bits, use an extract_subvector. 4059 if (Size > 128) { 4060 auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts); 4061 ReductionCost += 4062 getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy); 4063 Ty = SubTy; 4064 } else if (Size == 128) { 4065 // Reducing from 128 bits is a permute of v2f64/v2i64. 4066 FixedVectorType *ShufTy; 4067 if (ValVTy->isFloatingPointTy()) 4068 ShufTy = 4069 FixedVectorType::get(Type::getDoubleTy(ValVTy->getContext()), 2); 4070 else 4071 ShufTy = 4072 FixedVectorType::get(Type::getInt64Ty(ValVTy->getContext()), 2); 4073 ReductionCost += 4074 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr); 4075 } else if (Size == 64) { 4076 // Reducing from 64 bits is a shuffle of v4f32/v4i32. 4077 FixedVectorType *ShufTy; 4078 if (ValVTy->isFloatingPointTy()) 4079 ShufTy = 4080 FixedVectorType::get(Type::getFloatTy(ValVTy->getContext()), 4); 4081 else 4082 ShufTy = 4083 FixedVectorType::get(Type::getInt32Ty(ValVTy->getContext()), 4); 4084 ReductionCost += 4085 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr); 4086 } else { 4087 // Reducing from smaller size is a shift by immediate. 4088 auto *ShiftTy = FixedVectorType::get( 4089 Type::getIntNTy(ValVTy->getContext(), Size), 128 / Size); 4090 ReductionCost += getArithmeticInstrCost( 4091 Instruction::LShr, ShiftTy, CostKind, 4092 TargetTransformInfo::OK_AnyValue, 4093 TargetTransformInfo::OK_UniformConstantValue, 4094 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 4095 } 4096 4097 // Add the arithmetic op for this level. 4098 ReductionCost += getArithmeticInstrCost(Opcode, Ty, CostKind); 4099 } 4100 4101 // Add the final extract element to the cost. 4102 return ReductionCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0); 4103 } 4104 4105 InstructionCost X86TTIImpl::getMinMaxCost(Type *Ty, Type *CondTy, 4106 bool IsUnsigned) { 4107 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 4108 4109 MVT MTy = LT.second; 4110 4111 int ISD; 4112 if (Ty->isIntOrIntVectorTy()) { 4113 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN; 4114 } else { 4115 assert(Ty->isFPOrFPVectorTy() && 4116 "Expected float point or integer vector type."); 4117 ISD = ISD::FMINNUM; 4118 } 4119 4120 static const CostTblEntry SSE1CostTbl[] = { 4121 {ISD::FMINNUM, MVT::v4f32, 1}, 4122 }; 4123 4124 static const CostTblEntry SSE2CostTbl[] = { 4125 {ISD::FMINNUM, MVT::v2f64, 1}, 4126 {ISD::SMIN, MVT::v8i16, 1}, 4127 {ISD::UMIN, MVT::v16i8, 1}, 4128 }; 4129 4130 static const CostTblEntry SSE41CostTbl[] = { 4131 {ISD::SMIN, MVT::v4i32, 1}, 4132 {ISD::UMIN, MVT::v4i32, 1}, 4133 {ISD::UMIN, MVT::v8i16, 1}, 4134 {ISD::SMIN, MVT::v16i8, 1}, 4135 }; 4136 4137 static const CostTblEntry SSE42CostTbl[] = { 4138 {ISD::UMIN, MVT::v2i64, 3}, // xor+pcmpgtq+blendvpd 4139 }; 4140 4141 static const CostTblEntry AVX1CostTbl[] = { 4142 {ISD::FMINNUM, MVT::v8f32, 1}, 4143 {ISD::FMINNUM, MVT::v4f64, 1}, 4144 {ISD::SMIN, MVT::v8i32, 3}, 4145 {ISD::UMIN, MVT::v8i32, 3}, 4146 {ISD::SMIN, MVT::v16i16, 3}, 4147 {ISD::UMIN, MVT::v16i16, 3}, 4148 {ISD::SMIN, MVT::v32i8, 3}, 4149 {ISD::UMIN, MVT::v32i8, 3}, 4150 }; 4151 4152 static const CostTblEntry AVX2CostTbl[] = { 4153 {ISD::SMIN, MVT::v8i32, 1}, 4154 {ISD::UMIN, MVT::v8i32, 1}, 4155 {ISD::SMIN, MVT::v16i16, 1}, 4156 {ISD::UMIN, MVT::v16i16, 1}, 4157 {ISD::SMIN, MVT::v32i8, 1}, 4158 {ISD::UMIN, MVT::v32i8, 1}, 4159 }; 4160 4161 static const CostTblEntry AVX512CostTbl[] = { 4162 {ISD::FMINNUM, MVT::v16f32, 1}, 4163 {ISD::FMINNUM, MVT::v8f64, 1}, 4164 {ISD::SMIN, MVT::v2i64, 1}, 4165 {ISD::UMIN, MVT::v2i64, 1}, 4166 {ISD::SMIN, MVT::v4i64, 1}, 4167 {ISD::UMIN, MVT::v4i64, 1}, 4168 {ISD::SMIN, MVT::v8i64, 1}, 4169 {ISD::UMIN, MVT::v8i64, 1}, 4170 {ISD::SMIN, MVT::v16i32, 1}, 4171 {ISD::UMIN, MVT::v16i32, 1}, 4172 }; 4173 4174 static const CostTblEntry AVX512BWCostTbl[] = { 4175 {ISD::SMIN, MVT::v32i16, 1}, 4176 {ISD::UMIN, MVT::v32i16, 1}, 4177 {ISD::SMIN, MVT::v64i8, 1}, 4178 {ISD::UMIN, MVT::v64i8, 1}, 4179 }; 4180 4181 // If we have a native MIN/MAX instruction for this type, use it. 4182 if (ST->hasBWI()) 4183 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy)) 4184 return LT.first * Entry->Cost; 4185 4186 if (ST->hasAVX512()) 4187 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy)) 4188 return LT.first * Entry->Cost; 4189 4190 if (ST->hasAVX2()) 4191 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy)) 4192 return LT.first * Entry->Cost; 4193 4194 if (ST->hasAVX()) 4195 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy)) 4196 return LT.first * Entry->Cost; 4197 4198 if (ST->hasSSE42()) 4199 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy)) 4200 return LT.first * Entry->Cost; 4201 4202 if (ST->hasSSE41()) 4203 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy)) 4204 return LT.first * Entry->Cost; 4205 4206 if (ST->hasSSE2()) 4207 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy)) 4208 return LT.first * Entry->Cost; 4209 4210 if (ST->hasSSE1()) 4211 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy)) 4212 return LT.first * Entry->Cost; 4213 4214 unsigned CmpOpcode; 4215 if (Ty->isFPOrFPVectorTy()) { 4216 CmpOpcode = Instruction::FCmp; 4217 } else { 4218 assert(Ty->isIntOrIntVectorTy() && 4219 "expecting floating point or integer type for min/max reduction"); 4220 CmpOpcode = Instruction::ICmp; 4221 } 4222 4223 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; 4224 // Otherwise fall back to cmp+select. 4225 InstructionCost Result = 4226 getCmpSelInstrCost(CmpOpcode, Ty, CondTy, CmpInst::BAD_ICMP_PREDICATE, 4227 CostKind) + 4228 getCmpSelInstrCost(Instruction::Select, Ty, CondTy, 4229 CmpInst::BAD_ICMP_PREDICATE, CostKind); 4230 return Result; 4231 } 4232 4233 InstructionCost 4234 X86TTIImpl::getMinMaxReductionCost(VectorType *ValTy, VectorType *CondTy, 4235 bool IsUnsigned, 4236 TTI::TargetCostKind CostKind) { 4237 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 4238 4239 MVT MTy = LT.second; 4240 4241 int ISD; 4242 if (ValTy->isIntOrIntVectorTy()) { 4243 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN; 4244 } else { 4245 assert(ValTy->isFPOrFPVectorTy() && 4246 "Expected float point or integer vector type."); 4247 ISD = ISD::FMINNUM; 4248 } 4249 4250 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput 4251 // and make it as the cost. 4252 4253 static const CostTblEntry SSE2CostTblNoPairWise[] = { 4254 {ISD::UMIN, MVT::v2i16, 5}, // need pxors to use pminsw/pmaxsw 4255 {ISD::UMIN, MVT::v4i16, 7}, // need pxors to use pminsw/pmaxsw 4256 {ISD::UMIN, MVT::v8i16, 9}, // need pxors to use pminsw/pmaxsw 4257 }; 4258 4259 static const CostTblEntry SSE41CostTblNoPairWise[] = { 4260 {ISD::SMIN, MVT::v2i16, 3}, // same as sse2 4261 {ISD::SMIN, MVT::v4i16, 5}, // same as sse2 4262 {ISD::UMIN, MVT::v2i16, 5}, // same as sse2 4263 {ISD::UMIN, MVT::v4i16, 7}, // same as sse2 4264 {ISD::SMIN, MVT::v8i16, 4}, // phminposuw+xor 4265 {ISD::UMIN, MVT::v8i16, 4}, // FIXME: umin is cheaper than umax 4266 {ISD::SMIN, MVT::v2i8, 3}, // pminsb 4267 {ISD::SMIN, MVT::v4i8, 5}, // pminsb 4268 {ISD::SMIN, MVT::v8i8, 7}, // pminsb 4269 {ISD::SMIN, MVT::v16i8, 6}, 4270 {ISD::UMIN, MVT::v2i8, 3}, // same as sse2 4271 {ISD::UMIN, MVT::v4i8, 5}, // same as sse2 4272 {ISD::UMIN, MVT::v8i8, 7}, // same as sse2 4273 {ISD::UMIN, MVT::v16i8, 6}, // FIXME: umin is cheaper than umax 4274 }; 4275 4276 static const CostTblEntry AVX1CostTblNoPairWise[] = { 4277 {ISD::SMIN, MVT::v16i16, 6}, 4278 {ISD::UMIN, MVT::v16i16, 6}, // FIXME: umin is cheaper than umax 4279 {ISD::SMIN, MVT::v32i8, 8}, 4280 {ISD::UMIN, MVT::v32i8, 8}, 4281 }; 4282 4283 static const CostTblEntry AVX512BWCostTblNoPairWise[] = { 4284 {ISD::SMIN, MVT::v32i16, 8}, 4285 {ISD::UMIN, MVT::v32i16, 8}, // FIXME: umin is cheaper than umax 4286 {ISD::SMIN, MVT::v64i8, 10}, 4287 {ISD::UMIN, MVT::v64i8, 10}, 4288 }; 4289 4290 // Before legalizing the type, give a chance to look up illegal narrow types 4291 // in the table. 4292 // FIXME: Is there a better way to do this? 4293 EVT VT = TLI->getValueType(DL, ValTy); 4294 if (VT.isSimple()) { 4295 MVT MTy = VT.getSimpleVT(); 4296 if (ST->hasBWI()) 4297 if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy)) 4298 return Entry->Cost; 4299 4300 if (ST->hasAVX()) 4301 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 4302 return Entry->Cost; 4303 4304 if (ST->hasSSE41()) 4305 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy)) 4306 return Entry->Cost; 4307 4308 if (ST->hasSSE2()) 4309 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 4310 return Entry->Cost; 4311 } 4312 4313 auto *ValVTy = cast<FixedVectorType>(ValTy); 4314 unsigned NumVecElts = ValVTy->getNumElements(); 4315 4316 auto *Ty = ValVTy; 4317 InstructionCost MinMaxCost = 0; 4318 if (LT.first != 1 && MTy.isVector() && 4319 MTy.getVectorNumElements() < ValVTy->getNumElements()) { 4320 // Type needs to be split. We need LT.first - 1 operations ops. 4321 Ty = FixedVectorType::get(ValVTy->getElementType(), 4322 MTy.getVectorNumElements()); 4323 auto *SubCondTy = FixedVectorType::get(CondTy->getElementType(), 4324 MTy.getVectorNumElements()); 4325 MinMaxCost = getMinMaxCost(Ty, SubCondTy, IsUnsigned); 4326 MinMaxCost *= LT.first - 1; 4327 NumVecElts = MTy.getVectorNumElements(); 4328 } 4329 4330 if (ST->hasBWI()) 4331 if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy)) 4332 return MinMaxCost + Entry->Cost; 4333 4334 if (ST->hasAVX()) 4335 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy)) 4336 return MinMaxCost + Entry->Cost; 4337 4338 if (ST->hasSSE41()) 4339 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy)) 4340 return MinMaxCost + Entry->Cost; 4341 4342 if (ST->hasSSE2()) 4343 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy)) 4344 return MinMaxCost + Entry->Cost; 4345 4346 unsigned ScalarSize = ValTy->getScalarSizeInBits(); 4347 4348 // Special case power of 2 reductions where the scalar type isn't changed 4349 // by type legalization. 4350 if (!isPowerOf2_32(ValVTy->getNumElements()) || 4351 ScalarSize != MTy.getScalarSizeInBits()) 4352 return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsUnsigned, CostKind); 4353 4354 // Now handle reduction with the legal type, taking into account size changes 4355 // at each level. 4356 while (NumVecElts > 1) { 4357 // Determine the size of the remaining vector we need to reduce. 4358 unsigned Size = NumVecElts * ScalarSize; 4359 NumVecElts /= 2; 4360 // If we're reducing from 256/512 bits, use an extract_subvector. 4361 if (Size > 128) { 4362 auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts); 4363 MinMaxCost += 4364 getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy); 4365 Ty = SubTy; 4366 } else if (Size == 128) { 4367 // Reducing from 128 bits is a permute of v2f64/v2i64. 4368 VectorType *ShufTy; 4369 if (ValTy->isFloatingPointTy()) 4370 ShufTy = 4371 FixedVectorType::get(Type::getDoubleTy(ValTy->getContext()), 2); 4372 else 4373 ShufTy = FixedVectorType::get(Type::getInt64Ty(ValTy->getContext()), 2); 4374 MinMaxCost += 4375 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr); 4376 } else if (Size == 64) { 4377 // Reducing from 64 bits is a shuffle of v4f32/v4i32. 4378 FixedVectorType *ShufTy; 4379 if (ValTy->isFloatingPointTy()) 4380 ShufTy = FixedVectorType::get(Type::getFloatTy(ValTy->getContext()), 4); 4381 else 4382 ShufTy = FixedVectorType::get(Type::getInt32Ty(ValTy->getContext()), 4); 4383 MinMaxCost += 4384 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr); 4385 } else { 4386 // Reducing from smaller size is a shift by immediate. 4387 auto *ShiftTy = FixedVectorType::get( 4388 Type::getIntNTy(ValTy->getContext(), Size), 128 / Size); 4389 MinMaxCost += getArithmeticInstrCost( 4390 Instruction::LShr, ShiftTy, TTI::TCK_RecipThroughput, 4391 TargetTransformInfo::OK_AnyValue, 4392 TargetTransformInfo::OK_UniformConstantValue, 4393 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None); 4394 } 4395 4396 // Add the arithmetic op for this level. 4397 auto *SubCondTy = 4398 FixedVectorType::get(CondTy->getElementType(), Ty->getNumElements()); 4399 MinMaxCost += getMinMaxCost(Ty, SubCondTy, IsUnsigned); 4400 } 4401 4402 // Add the final extract element to the cost. 4403 return MinMaxCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0); 4404 } 4405 4406 /// Calculate the cost of materializing a 64-bit value. This helper 4407 /// method might only calculate a fraction of a larger immediate. Therefore it 4408 /// is valid to return a cost of ZERO. 4409 InstructionCost X86TTIImpl::getIntImmCost(int64_t Val) { 4410 if (Val == 0) 4411 return TTI::TCC_Free; 4412 4413 if (isInt<32>(Val)) 4414 return TTI::TCC_Basic; 4415 4416 return 2 * TTI::TCC_Basic; 4417 } 4418 4419 InstructionCost X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty, 4420 TTI::TargetCostKind CostKind) { 4421 assert(Ty->isIntegerTy()); 4422 4423 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 4424 if (BitSize == 0) 4425 return ~0U; 4426 4427 // Never hoist constants larger than 128bit, because this might lead to 4428 // incorrect code generation or assertions in codegen. 4429 // Fixme: Create a cost model for types larger than i128 once the codegen 4430 // issues have been fixed. 4431 if (BitSize > 128) 4432 return TTI::TCC_Free; 4433 4434 if (Imm == 0) 4435 return TTI::TCC_Free; 4436 4437 // Sign-extend all constants to a multiple of 64-bit. 4438 APInt ImmVal = Imm; 4439 if (BitSize % 64 != 0) 4440 ImmVal = Imm.sext(alignTo(BitSize, 64)); 4441 4442 // Split the constant into 64-bit chunks and calculate the cost for each 4443 // chunk. 4444 InstructionCost Cost = 0; 4445 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) { 4446 APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64); 4447 int64_t Val = Tmp.getSExtValue(); 4448 Cost += getIntImmCost(Val); 4449 } 4450 // We need at least one instruction to materialize the constant. 4451 return std::max<InstructionCost>(1, Cost); 4452 } 4453 4454 InstructionCost X86TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx, 4455 const APInt &Imm, Type *Ty, 4456 TTI::TargetCostKind CostKind, 4457 Instruction *Inst) { 4458 assert(Ty->isIntegerTy()); 4459 4460 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 4461 // There is no cost model for constants with a bit size of 0. Return TCC_Free 4462 // here, so that constant hoisting will ignore this constant. 4463 if (BitSize == 0) 4464 return TTI::TCC_Free; 4465 4466 unsigned ImmIdx = ~0U; 4467 switch (Opcode) { 4468 default: 4469 return TTI::TCC_Free; 4470 case Instruction::GetElementPtr: 4471 // Always hoist the base address of a GetElementPtr. This prevents the 4472 // creation of new constants for every base constant that gets constant 4473 // folded with the offset. 4474 if (Idx == 0) 4475 return 2 * TTI::TCC_Basic; 4476 return TTI::TCC_Free; 4477 case Instruction::Store: 4478 ImmIdx = 0; 4479 break; 4480 case Instruction::ICmp: 4481 // This is an imperfect hack to prevent constant hoisting of 4482 // compares that might be trying to check if a 64-bit value fits in 4483 // 32-bits. The backend can optimize these cases using a right shift by 32. 4484 // Ideally we would check the compare predicate here. There also other 4485 // similar immediates the backend can use shifts for. 4486 if (Idx == 1 && Imm.getBitWidth() == 64) { 4487 uint64_t ImmVal = Imm.getZExtValue(); 4488 if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff) 4489 return TTI::TCC_Free; 4490 } 4491 ImmIdx = 1; 4492 break; 4493 case Instruction::And: 4494 // We support 64-bit ANDs with immediates with 32-bits of leading zeroes 4495 // by using a 32-bit operation with implicit zero extension. Detect such 4496 // immediates here as the normal path expects bit 31 to be sign extended. 4497 if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue())) 4498 return TTI::TCC_Free; 4499 ImmIdx = 1; 4500 break; 4501 case Instruction::Add: 4502 case Instruction::Sub: 4503 // For add/sub, we can use the opposite instruction for INT32_MIN. 4504 if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000) 4505 return TTI::TCC_Free; 4506 ImmIdx = 1; 4507 break; 4508 case Instruction::UDiv: 4509 case Instruction::SDiv: 4510 case Instruction::URem: 4511 case Instruction::SRem: 4512 // Division by constant is typically expanded later into a different 4513 // instruction sequence. This completely changes the constants. 4514 // Report them as "free" to stop ConstantHoist from marking them as opaque. 4515 return TTI::TCC_Free; 4516 case Instruction::Mul: 4517 case Instruction::Or: 4518 case Instruction::Xor: 4519 ImmIdx = 1; 4520 break; 4521 // Always return TCC_Free for the shift value of a shift instruction. 4522 case Instruction::Shl: 4523 case Instruction::LShr: 4524 case Instruction::AShr: 4525 if (Idx == 1) 4526 return TTI::TCC_Free; 4527 break; 4528 case Instruction::Trunc: 4529 case Instruction::ZExt: 4530 case Instruction::SExt: 4531 case Instruction::IntToPtr: 4532 case Instruction::PtrToInt: 4533 case Instruction::BitCast: 4534 case Instruction::PHI: 4535 case Instruction::Call: 4536 case Instruction::Select: 4537 case Instruction::Ret: 4538 case Instruction::Load: 4539 break; 4540 } 4541 4542 if (Idx == ImmIdx) { 4543 int NumConstants = divideCeil(BitSize, 64); 4544 InstructionCost Cost = X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 4545 return (Cost <= NumConstants * TTI::TCC_Basic) 4546 ? static_cast<int>(TTI::TCC_Free) 4547 : Cost; 4548 } 4549 4550 return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 4551 } 4552 4553 InstructionCost X86TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, 4554 const APInt &Imm, Type *Ty, 4555 TTI::TargetCostKind CostKind) { 4556 assert(Ty->isIntegerTy()); 4557 4558 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 4559 // There is no cost model for constants with a bit size of 0. Return TCC_Free 4560 // here, so that constant hoisting will ignore this constant. 4561 if (BitSize == 0) 4562 return TTI::TCC_Free; 4563 4564 switch (IID) { 4565 default: 4566 return TTI::TCC_Free; 4567 case Intrinsic::sadd_with_overflow: 4568 case Intrinsic::uadd_with_overflow: 4569 case Intrinsic::ssub_with_overflow: 4570 case Intrinsic::usub_with_overflow: 4571 case Intrinsic::smul_with_overflow: 4572 case Intrinsic::umul_with_overflow: 4573 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue())) 4574 return TTI::TCC_Free; 4575 break; 4576 case Intrinsic::experimental_stackmap: 4577 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 4578 return TTI::TCC_Free; 4579 break; 4580 case Intrinsic::experimental_patchpoint_void: 4581 case Intrinsic::experimental_patchpoint_i64: 4582 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 4583 return TTI::TCC_Free; 4584 break; 4585 } 4586 return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind); 4587 } 4588 4589 InstructionCost X86TTIImpl::getCFInstrCost(unsigned Opcode, 4590 TTI::TargetCostKind CostKind, 4591 const Instruction *I) { 4592 if (CostKind != TTI::TCK_RecipThroughput) 4593 return Opcode == Instruction::PHI ? 0 : 1; 4594 // Branches are assumed to be predicted. 4595 return 0; 4596 } 4597 4598 int X86TTIImpl::getGatherOverhead() const { 4599 // Some CPUs have more overhead for gather. The specified overhead is relative 4600 // to the Load operation. "2" is the number provided by Intel architects. This 4601 // parameter is used for cost estimation of Gather Op and comparison with 4602 // other alternatives. 4603 // TODO: Remove the explicit hasAVX512()?, That would mean we would only 4604 // enable gather with a -march. 4605 if (ST->hasAVX512() || (ST->hasAVX2() && ST->hasFastGather())) 4606 return 2; 4607 4608 return 1024; 4609 } 4610 4611 int X86TTIImpl::getScatterOverhead() const { 4612 if (ST->hasAVX512()) 4613 return 2; 4614 4615 return 1024; 4616 } 4617 4618 // Return an average cost of Gather / Scatter instruction, maybe improved later. 4619 // FIXME: Add TargetCostKind support. 4620 InstructionCost X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy, 4621 const Value *Ptr, Align Alignment, 4622 unsigned AddressSpace) { 4623 4624 assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost"); 4625 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements(); 4626 4627 // Try to reduce index size from 64 bit (default for GEP) 4628 // to 32. It is essential for VF 16. If the index can't be reduced to 32, the 4629 // operation will use 16 x 64 indices which do not fit in a zmm and needs 4630 // to split. Also check that the base pointer is the same for all lanes, 4631 // and that there's at most one variable index. 4632 auto getIndexSizeInBits = [](const Value *Ptr, const DataLayout &DL) { 4633 unsigned IndexSize = DL.getPointerSizeInBits(); 4634 const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr); 4635 if (IndexSize < 64 || !GEP) 4636 return IndexSize; 4637 4638 unsigned NumOfVarIndices = 0; 4639 const Value *Ptrs = GEP->getPointerOperand(); 4640 if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs)) 4641 return IndexSize; 4642 for (unsigned i = 1; i < GEP->getNumOperands(); ++i) { 4643 if (isa<Constant>(GEP->getOperand(i))) 4644 continue; 4645 Type *IndxTy = GEP->getOperand(i)->getType(); 4646 if (auto *IndexVTy = dyn_cast<VectorType>(IndxTy)) 4647 IndxTy = IndexVTy->getElementType(); 4648 if ((IndxTy->getPrimitiveSizeInBits() == 64 && 4649 !isa<SExtInst>(GEP->getOperand(i))) || 4650 ++NumOfVarIndices > 1) 4651 return IndexSize; // 64 4652 } 4653 return (unsigned)32; 4654 }; 4655 4656 // Trying to reduce IndexSize to 32 bits for vector 16. 4657 // By default the IndexSize is equal to pointer size. 4658 unsigned IndexSize = (ST->hasAVX512() && VF >= 16) 4659 ? getIndexSizeInBits(Ptr, DL) 4660 : DL.getPointerSizeInBits(); 4661 4662 auto *IndexVTy = FixedVectorType::get( 4663 IntegerType::get(SrcVTy->getContext(), IndexSize), VF); 4664 std::pair<InstructionCost, MVT> IdxsLT = 4665 TLI->getTypeLegalizationCost(DL, IndexVTy); 4666 std::pair<InstructionCost, MVT> SrcLT = 4667 TLI->getTypeLegalizationCost(DL, SrcVTy); 4668 InstructionCost::CostType SplitFactor = 4669 *std::max(IdxsLT.first, SrcLT.first).getValue(); 4670 if (SplitFactor > 1) { 4671 // Handle splitting of vector of pointers 4672 auto *SplitSrcTy = 4673 FixedVectorType::get(SrcVTy->getScalarType(), VF / SplitFactor); 4674 return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment, 4675 AddressSpace); 4676 } 4677 4678 // The gather / scatter cost is given by Intel architects. It is a rough 4679 // number since we are looking at one instruction in a time. 4680 const int GSOverhead = (Opcode == Instruction::Load) 4681 ? getGatherOverhead() 4682 : getScatterOverhead(); 4683 return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 4684 MaybeAlign(Alignment), AddressSpace, 4685 TTI::TCK_RecipThroughput); 4686 } 4687 4688 /// Return the cost of full scalarization of gather / scatter operation. 4689 /// 4690 /// Opcode - Load or Store instruction. 4691 /// SrcVTy - The type of the data vector that should be gathered or scattered. 4692 /// VariableMask - The mask is non-constant at compile time. 4693 /// Alignment - Alignment for one element. 4694 /// AddressSpace - pointer[s] address space. 4695 /// 4696 /// FIXME: Add TargetCostKind support. 4697 InstructionCost X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy, 4698 bool VariableMask, Align Alignment, 4699 unsigned AddressSpace) { 4700 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements(); 4701 APInt DemandedElts = APInt::getAllOnes(VF); 4702 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; 4703 4704 InstructionCost MaskUnpackCost = 0; 4705 if (VariableMask) { 4706 auto *MaskTy = 4707 FixedVectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF); 4708 MaskUnpackCost = 4709 getScalarizationOverhead(MaskTy, DemandedElts, false, true); 4710 InstructionCost ScalarCompareCost = getCmpSelInstrCost( 4711 Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()), nullptr, 4712 CmpInst::BAD_ICMP_PREDICATE, CostKind); 4713 InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind); 4714 MaskUnpackCost += VF * (BranchCost + ScalarCompareCost); 4715 } 4716 4717 // The cost of the scalar loads/stores. 4718 InstructionCost MemoryOpCost = 4719 VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(), 4720 MaybeAlign(Alignment), AddressSpace, CostKind); 4721 4722 InstructionCost InsertExtractCost = 0; 4723 if (Opcode == Instruction::Load) 4724 for (unsigned i = 0; i < VF; ++i) 4725 // Add the cost of inserting each scalar load into the vector 4726 InsertExtractCost += 4727 getVectorInstrCost(Instruction::InsertElement, SrcVTy, i); 4728 else 4729 for (unsigned i = 0; i < VF; ++i) 4730 // Add the cost of extracting each element out of the data vector 4731 InsertExtractCost += 4732 getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i); 4733 4734 return MemoryOpCost + MaskUnpackCost + InsertExtractCost; 4735 } 4736 4737 /// Calculate the cost of Gather / Scatter operation 4738 InstructionCost X86TTIImpl::getGatherScatterOpCost( 4739 unsigned Opcode, Type *SrcVTy, const Value *Ptr, bool VariableMask, 4740 Align Alignment, TTI::TargetCostKind CostKind, 4741 const Instruction *I = nullptr) { 4742 if (CostKind != TTI::TCK_RecipThroughput) { 4743 if ((Opcode == Instruction::Load && 4744 isLegalMaskedGather(SrcVTy, Align(Alignment))) || 4745 (Opcode == Instruction::Store && 4746 isLegalMaskedScatter(SrcVTy, Align(Alignment)))) 4747 return 1; 4748 return BaseT::getGatherScatterOpCost(Opcode, SrcVTy, Ptr, VariableMask, 4749 Alignment, CostKind, I); 4750 } 4751 4752 assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter"); 4753 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType()); 4754 if (!PtrTy && Ptr->getType()->isVectorTy()) 4755 PtrTy = dyn_cast<PointerType>( 4756 cast<VectorType>(Ptr->getType())->getElementType()); 4757 assert(PtrTy && "Unexpected type for Ptr argument"); 4758 unsigned AddressSpace = PtrTy->getAddressSpace(); 4759 4760 if ((Opcode == Instruction::Load && 4761 !isLegalMaskedGather(SrcVTy, Align(Alignment))) || 4762 (Opcode == Instruction::Store && 4763 !isLegalMaskedScatter(SrcVTy, Align(Alignment)))) 4764 return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment, 4765 AddressSpace); 4766 4767 return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace); 4768 } 4769 4770 bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1, 4771 TargetTransformInfo::LSRCost &C2) { 4772 // X86 specific here are "instruction number 1st priority". 4773 return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost, 4774 C1.NumIVMuls, C1.NumBaseAdds, 4775 C1.ScaleCost, C1.ImmCost, C1.SetupCost) < 4776 std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost, 4777 C2.NumIVMuls, C2.NumBaseAdds, 4778 C2.ScaleCost, C2.ImmCost, C2.SetupCost); 4779 } 4780 4781 bool X86TTIImpl::canMacroFuseCmp() { 4782 return ST->hasMacroFusion() || ST->hasBranchFusion(); 4783 } 4784 4785 bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) { 4786 if (!ST->hasAVX()) 4787 return false; 4788 4789 // The backend can't handle a single element vector. 4790 if (isa<VectorType>(DataTy) && 4791 cast<FixedVectorType>(DataTy)->getNumElements() == 1) 4792 return false; 4793 Type *ScalarTy = DataTy->getScalarType(); 4794 4795 if (ScalarTy->isPointerTy()) 4796 return true; 4797 4798 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4799 return true; 4800 4801 if (ScalarTy->isHalfTy() && ST->hasBWI() && ST->hasFP16()) 4802 return true; 4803 4804 if (!ScalarTy->isIntegerTy()) 4805 return false; 4806 4807 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4808 return IntWidth == 32 || IntWidth == 64 || 4809 ((IntWidth == 8 || IntWidth == 16) && ST->hasBWI()); 4810 } 4811 4812 bool X86TTIImpl::isLegalMaskedStore(Type *DataType, Align Alignment) { 4813 return isLegalMaskedLoad(DataType, Alignment); 4814 } 4815 4816 bool X86TTIImpl::isLegalNTLoad(Type *DataType, Align Alignment) { 4817 unsigned DataSize = DL.getTypeStoreSize(DataType); 4818 // The only supported nontemporal loads are for aligned vectors of 16 or 32 4819 // bytes. Note that 32-byte nontemporal vector loads are supported by AVX2 4820 // (the equivalent stores only require AVX). 4821 if (Alignment >= DataSize && (DataSize == 16 || DataSize == 32)) 4822 return DataSize == 16 ? ST->hasSSE1() : ST->hasAVX2(); 4823 4824 return false; 4825 } 4826 4827 bool X86TTIImpl::isLegalNTStore(Type *DataType, Align Alignment) { 4828 unsigned DataSize = DL.getTypeStoreSize(DataType); 4829 4830 // SSE4A supports nontemporal stores of float and double at arbitrary 4831 // alignment. 4832 if (ST->hasSSE4A() && (DataType->isFloatTy() || DataType->isDoubleTy())) 4833 return true; 4834 4835 // Besides the SSE4A subtarget exception above, only aligned stores are 4836 // available nontemporaly on any other subtarget. And only stores with a size 4837 // of 4..32 bytes (powers of 2, only) are permitted. 4838 if (Alignment < DataSize || DataSize < 4 || DataSize > 32 || 4839 !isPowerOf2_32(DataSize)) 4840 return false; 4841 4842 // 32-byte vector nontemporal stores are supported by AVX (the equivalent 4843 // loads require AVX2). 4844 if (DataSize == 32) 4845 return ST->hasAVX(); 4846 if (DataSize == 16) 4847 return ST->hasSSE1(); 4848 return true; 4849 } 4850 4851 bool X86TTIImpl::isLegalMaskedExpandLoad(Type *DataTy) { 4852 if (!isa<VectorType>(DataTy)) 4853 return false; 4854 4855 if (!ST->hasAVX512()) 4856 return false; 4857 4858 // The backend can't handle a single element vector. 4859 if (cast<FixedVectorType>(DataTy)->getNumElements() == 1) 4860 return false; 4861 4862 Type *ScalarTy = cast<VectorType>(DataTy)->getElementType(); 4863 4864 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4865 return true; 4866 4867 if (!ScalarTy->isIntegerTy()) 4868 return false; 4869 4870 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4871 return IntWidth == 32 || IntWidth == 64 || 4872 ((IntWidth == 8 || IntWidth == 16) && ST->hasVBMI2()); 4873 } 4874 4875 bool X86TTIImpl::isLegalMaskedCompressStore(Type *DataTy) { 4876 return isLegalMaskedExpandLoad(DataTy); 4877 } 4878 4879 bool X86TTIImpl::isLegalMaskedGather(Type *DataTy, Align Alignment) { 4880 // Some CPUs have better gather performance than others. 4881 // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only 4882 // enable gather with a -march. 4883 if (!(ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2()))) 4884 return false; 4885 4886 // This function is called now in two cases: from the Loop Vectorizer 4887 // and from the Scalarizer. 4888 // When the Loop Vectorizer asks about legality of the feature, 4889 // the vectorization factor is not calculated yet. The Loop Vectorizer 4890 // sends a scalar type and the decision is based on the width of the 4891 // scalar element. 4892 // Later on, the cost model will estimate usage this intrinsic based on 4893 // the vector type. 4894 // The Scalarizer asks again about legality. It sends a vector type. 4895 // In this case we can reject non-power-of-2 vectors. 4896 // We also reject single element vectors as the type legalizer can't 4897 // scalarize it. 4898 if (auto *DataVTy = dyn_cast<FixedVectorType>(DataTy)) { 4899 unsigned NumElts = DataVTy->getNumElements(); 4900 if (NumElts == 1) 4901 return false; 4902 // Gather / Scatter for vector 2 is not profitable on KNL / SKX 4903 // Vector-4 of gather/scatter instruction does not exist on KNL. 4904 // We can extend it to 8 elements, but zeroing upper bits of 4905 // the mask vector will add more instructions. Right now we give the scalar 4906 // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter 4907 // instruction is better in the VariableMask case. 4908 if (ST->hasAVX512() && (NumElts == 2 || (NumElts == 4 && !ST->hasVLX()))) 4909 return false; 4910 } 4911 Type *ScalarTy = DataTy->getScalarType(); 4912 if (ScalarTy->isPointerTy()) 4913 return true; 4914 4915 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) 4916 return true; 4917 4918 if (!ScalarTy->isIntegerTy()) 4919 return false; 4920 4921 unsigned IntWidth = ScalarTy->getIntegerBitWidth(); 4922 return IntWidth == 32 || IntWidth == 64; 4923 } 4924 4925 bool X86TTIImpl::isLegalMaskedScatter(Type *DataType, Align Alignment) { 4926 // AVX2 doesn't support scatter 4927 if (!ST->hasAVX512()) 4928 return false; 4929 return isLegalMaskedGather(DataType, Alignment); 4930 } 4931 4932 bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) { 4933 EVT VT = TLI->getValueType(DL, DataType); 4934 return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT); 4935 } 4936 4937 bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) { 4938 return false; 4939 } 4940 4941 bool X86TTIImpl::areInlineCompatible(const Function *Caller, 4942 const Function *Callee) const { 4943 const TargetMachine &TM = getTLI()->getTargetMachine(); 4944 4945 // Work this as a subsetting of subtarget features. 4946 const FeatureBitset &CallerBits = 4947 TM.getSubtargetImpl(*Caller)->getFeatureBits(); 4948 const FeatureBitset &CalleeBits = 4949 TM.getSubtargetImpl(*Callee)->getFeatureBits(); 4950 4951 FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList; 4952 FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList; 4953 return (RealCallerBits & RealCalleeBits) == RealCalleeBits; 4954 } 4955 4956 bool X86TTIImpl::areFunctionArgsABICompatible( 4957 const Function *Caller, const Function *Callee, 4958 SmallPtrSetImpl<Argument *> &Args) const { 4959 if (!BaseT::areFunctionArgsABICompatible(Caller, Callee, Args)) 4960 return false; 4961 4962 // If we get here, we know the target features match. If one function 4963 // considers 512-bit vectors legal and the other does not, consider them 4964 // incompatible. 4965 const TargetMachine &TM = getTLI()->getTargetMachine(); 4966 4967 if (TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() == 4968 TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs()) 4969 return true; 4970 4971 // Consider the arguments compatible if they aren't vectors or aggregates. 4972 // FIXME: Look at the size of vectors. 4973 // FIXME: Look at the element types of aggregates to see if there are vectors. 4974 // FIXME: The API of this function seems intended to allow arguments 4975 // to be removed from the set, but the caller doesn't check if the set 4976 // becomes empty so that may not work in practice. 4977 return llvm::none_of(Args, [](Argument *A) { 4978 auto *EltTy = cast<PointerType>(A->getType())->getElementType(); 4979 return EltTy->isVectorTy() || EltTy->isAggregateType(); 4980 }); 4981 } 4982 4983 X86TTIImpl::TTI::MemCmpExpansionOptions 4984 X86TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { 4985 TTI::MemCmpExpansionOptions Options; 4986 Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize); 4987 Options.NumLoadsPerBlock = 2; 4988 // All GPR and vector loads can be unaligned. 4989 Options.AllowOverlappingLoads = true; 4990 if (IsZeroCmp) { 4991 // Only enable vector loads for equality comparison. Right now the vector 4992 // version is not as fast for three way compare (see #33329). 4993 const unsigned PreferredWidth = ST->getPreferVectorWidth(); 4994 if (PreferredWidth >= 512 && ST->hasAVX512()) Options.LoadSizes.push_back(64); 4995 if (PreferredWidth >= 256 && ST->hasAVX()) Options.LoadSizes.push_back(32); 4996 if (PreferredWidth >= 128 && ST->hasSSE2()) Options.LoadSizes.push_back(16); 4997 } 4998 if (ST->is64Bit()) { 4999 Options.LoadSizes.push_back(8); 5000 } 5001 Options.LoadSizes.push_back(4); 5002 Options.LoadSizes.push_back(2); 5003 Options.LoadSizes.push_back(1); 5004 return Options; 5005 } 5006 5007 bool X86TTIImpl::enableInterleavedAccessVectorization() { 5008 // TODO: We expect this to be beneficial regardless of arch, 5009 // but there are currently some unexplained performance artifacts on Atom. 5010 // As a temporary solution, disable on Atom. 5011 return !(ST->isAtom()); 5012 } 5013 5014 // Get estimation for interleaved load/store operations for AVX2. 5015 // \p Factor is the interleaved-access factor (stride) - number of 5016 // (interleaved) elements in the group. 5017 // \p Indices contains the indices for a strided load: when the 5018 // interleaved load has gaps they indicate which elements are used. 5019 // If Indices is empty (or if the number of indices is equal to the size 5020 // of the interleaved-access as given in \p Factor) the access has no gaps. 5021 // 5022 // As opposed to AVX-512, AVX2 does not have generic shuffles that allow 5023 // computing the cost using a generic formula as a function of generic 5024 // shuffles. We therefore use a lookup table instead, filled according to 5025 // the instruction sequences that codegen currently generates. 5026 InstructionCost X86TTIImpl::getInterleavedMemoryOpCostAVX2( 5027 unsigned Opcode, FixedVectorType *VecTy, unsigned Factor, 5028 ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace, 5029 TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) { 5030 5031 if (UseMaskForCond || UseMaskForGaps) 5032 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5033 Alignment, AddressSpace, CostKind, 5034 UseMaskForCond, UseMaskForGaps); 5035 5036 // We currently Support only fully-interleaved groups, with no gaps. 5037 // TODO: Support also strided loads (interleaved-groups with gaps). 5038 if (Indices.size() && Indices.size() != Factor) 5039 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5040 Alignment, AddressSpace, CostKind); 5041 5042 // VecTy for interleave memop is <VF*Factor x Elt>. 5043 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have 5044 // VecTy = <12 x i32>. 5045 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second; 5046 5047 // This function can be called with VecTy=<6xi128>, Factor=3, in which case 5048 // the VF=2, while v2i128 is an unsupported MVT vector type 5049 // (see MachineValueType.h::getVectorVT()). 5050 if (!LegalVT.isVector()) 5051 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5052 Alignment, AddressSpace, CostKind); 5053 5054 unsigned VF = VecTy->getNumElements() / Factor; 5055 Type *ScalarTy = VecTy->getElementType(); 5056 // Deduplicate entries, model floats/pointers as appropriately-sized integers. 5057 if (!ScalarTy->isIntegerTy()) 5058 ScalarTy = 5059 Type::getIntNTy(ScalarTy->getContext(), DL.getTypeSizeInBits(ScalarTy)); 5060 5061 // Get the cost of all the memory operations. 5062 InstructionCost MemOpCosts = getMemoryOpCost( 5063 Opcode, VecTy, MaybeAlign(Alignment), AddressSpace, CostKind); 5064 5065 auto *VT = FixedVectorType::get(ScalarTy, VF); 5066 EVT ETy = TLI->getValueType(DL, VT); 5067 if (!ETy.isSimple()) 5068 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5069 Alignment, AddressSpace, CostKind); 5070 5071 // TODO: Complete for other data-types and strides. 5072 // Each combination of Stride, element bit width and VF results in a different 5073 // sequence; The cost tables are therefore accessed with: 5074 // Factor (stride) and VectorType=VFxiN. 5075 // The Cost accounts only for the shuffle sequence; 5076 // The cost of the loads/stores is accounted for separately. 5077 // 5078 static const CostTblEntry AVX2InterleavedLoadTbl[] = { 5079 {2, MVT::v2i8, 2}, // (load 4i8 and) deinterleave into 2 x 2i8 5080 {2, MVT::v4i8, 2}, // (load 8i8 and) deinterleave into 2 x 4i8 5081 {2, MVT::v8i8, 2}, // (load 16i8 and) deinterleave into 2 x 8i8 5082 {2, MVT::v16i8, 4}, // (load 32i8 and) deinterleave into 2 x 16i8 5083 {2, MVT::v32i8, 6}, // (load 64i8 and) deinterleave into 2 x 32i8 5084 5085 {2, MVT::v2i16, 2}, // (load 4i16 and) deinterleave into 2 x 2i16 5086 {2, MVT::v4i16, 2}, // (load 8i16 and) deinterleave into 2 x 4i16 5087 {2, MVT::v8i16, 6}, // (load 16i16 and) deinterleave into 2 x 8i16 5088 {2, MVT::v16i16, 9}, // (load 32i16 and) deinterleave into 2 x 16i16 5089 {2, MVT::v32i16, 18}, // (load 64i16 and) deinterleave into 2 x 32i16 5090 5091 {2, MVT::v2i32, 2}, // (load 4i32 and) deinterleave into 2 x 2i32 5092 {2, MVT::v4i32, 2}, // (load 8i32 and) deinterleave into 2 x 4i32 5093 {2, MVT::v8i32, 4}, // (load 16i32 and) deinterleave into 2 x 8i32 5094 {2, MVT::v16i32, 8}, // (load 32i32 and) deinterleave into 2 x 16i32 5095 {2, MVT::v32i32, 16}, // (load 64i32 and) deinterleave into 2 x 32i32 5096 5097 {2, MVT::v2i64, 2}, // (load 4i64 and) deinterleave into 2 x 2i64 5098 {2, MVT::v4i64, 4}, // (load 8i64 and) deinterleave into 2 x 4i64 5099 {2, MVT::v8i64, 8}, // (load 16i64 and) deinterleave into 2 x 8i64 5100 {2, MVT::v16i64, 16}, // (load 32i64 and) deinterleave into 2 x 16i64 5101 5102 {3, MVT::v2i8, 3}, // (load 6i8 and) deinterleave into 3 x 2i8 5103 {3, MVT::v4i8, 3}, // (load 12i8 and) deinterleave into 3 x 4i8 5104 {3, MVT::v8i8, 6}, // (load 24i8 and) deinterleave into 3 x 8i8 5105 {3, MVT::v16i8, 11}, // (load 48i8 and) deinterleave into 3 x 16i8 5106 {3, MVT::v32i8, 14}, // (load 96i8 and) deinterleave into 3 x 32i8 5107 5108 {3, MVT::v2i16, 5}, // (load 6i16 and) deinterleave into 3 x 2i16 5109 {3, MVT::v4i16, 7}, // (load 12i16 and) deinterleave into 3 x 4i16 5110 {3, MVT::v8i16, 9}, // (load 24i16 and) deinterleave into 3 x 8i16 5111 {3, MVT::v16i16, 28}, // (load 48i16 and) deinterleave into 3 x 16i16 5112 {3, MVT::v32i16, 56}, // (load 96i16 and) deinterleave into 3 x 32i16 5113 5114 {3, MVT::v2i32, 3}, // (load 6i32 and) deinterleave into 3 x 2i32 5115 {3, MVT::v4i32, 3}, // (load 12i32 and) deinterleave into 3 x 4i32 5116 {3, MVT::v8i32, 7}, // (load 24i32 and) deinterleave into 3 x 8i32 5117 {3, MVT::v16i32, 14}, // (load 48i32 and) deinterleave into 3 x 16i32 5118 5119 {3, MVT::v2i64, 1}, // (load 6i64 and) deinterleave into 3 x 2i64 5120 {3, MVT::v4i64, 5}, // (load 12i64 and) deinterleave into 3 x 4i64 5121 {3, MVT::v8i64, 10}, // (load 24i64 and) deinterleave into 3 x 8i64 5122 {3, MVT::v16i64, 20}, // (load 48i64 and) deinterleave into 3 x 16i64 5123 5124 {4, MVT::v2i8, 4}, // (load 8i8 and) deinterleave into 4 x 2i8 5125 {4, MVT::v4i8, 4}, // (load 16i8 and) deinterleave into 4 x 4i8 5126 {4, MVT::v8i8, 12}, // (load 32i8 and) deinterleave into 4 x 8i8 5127 {4, MVT::v16i8, 24}, // (load 64i8 and) deinterleave into 4 x 16i8 5128 {4, MVT::v32i8, 56}, // (load 128i8 and) deinterleave into 4 x 32i8 5129 5130 {4, MVT::v2i16, 6}, // (load 8i16 and) deinterleave into 4 x 2i16 5131 {4, MVT::v4i16, 17}, // (load 16i16 and) deinterleave into 4 x 4i16 5132 {4, MVT::v8i16, 33}, // (load 32i16 and) deinterleave into 4 x 8i16 5133 {4, MVT::v16i16, 75}, // (load 64i16 and) deinterleave into 4 x 16i16 5134 {4, MVT::v32i16, 150}, // (load 128i16 and) deinterleave into 4 x 32i16 5135 5136 {4, MVT::v2i32, 4}, // (load 8i32 and) deinterleave into 4 x 2i32 5137 {4, MVT::v4i32, 8}, // (load 16i32 and) deinterleave into 4 x 4i32 5138 {4, MVT::v8i32, 16}, // (load 32i32 and) deinterleave into 4 x 8i32 5139 {4, MVT::v16i32, 32}, // (load 64i32 and) deinterleave into 4 x 16i32 5140 5141 {4, MVT::v2i64, 6}, // (load 8i64 and) deinterleave into 4 x 2i64 5142 {4, MVT::v4i64, 8}, // (load 16i64 and) deinterleave into 4 x 4i64 5143 {4, MVT::v8i64, 20}, // (load 32i64 and) deinterleave into 4 x 8i64 5144 5145 {6, MVT::v2i8, 6}, // (load 12i8 and) deinterleave into 6 x 2i8 5146 {6, MVT::v4i8, 14}, // (load 24i8 and) deinterleave into 6 x 4i8 5147 {6, MVT::v8i8, 18}, // (load 48i8 and) deinterleave into 6 x 8i8 5148 {6, MVT::v16i8, 43}, // (load 96i8 and) deinterleave into 6 x 16i8 5149 {6, MVT::v32i8, 82}, // (load 192i8 and) deinterleave into 6 x 32i8 5150 5151 {6, MVT::v2i16, 13}, // (load 12i16 and) deinterleave into 6 x 2i16 5152 {6, MVT::v4i16, 9}, // (load 24i16 and) deinterleave into 6 x 4i16 5153 {6, MVT::v8i16, 39}, // (load 48i16 and) deinterleave into 6 x 8i16 5154 {6, MVT::v16i16, 106}, // (load 96i16 and) deinterleave into 6 x 16i16 5155 5156 {6, MVT::v2i32, 6}, // (load 12i32 and) deinterleave into 6 x 2i32 5157 {6, MVT::v4i32, 15}, // (load 24i32 and) deinterleave into 6 x 4i32 5158 {6, MVT::v8i32, 31}, // (load 48i32 and) deinterleave into 6 x 8i32 5159 {6, MVT::v16i32, 64}, // (load 96i32 and) deinterleave into 6 x 16i32 5160 5161 {6, MVT::v2i64, 6}, // (load 12i64 and) deinterleave into 6 x 2i64 5162 {6, MVT::v4i64, 18}, // (load 24i64 and) deinterleave into 6 x 4i64 5163 {6, MVT::v8i64, 36}, // (load 48i64 and) deinterleave into 6 x 8i64 5164 5165 {8, MVT::v8i32, 40} // (load 64i32 and) deinterleave into 8 x 8i32 5166 }; 5167 5168 static const CostTblEntry AVX2InterleavedStoreTbl[] = { 5169 {2, MVT::v2i8, 1}, // interleave 2 x 2i8 into 4i8 (and store) 5170 {2, MVT::v4i8, 1}, // interleave 2 x 4i8 into 8i8 (and store) 5171 {2, MVT::v8i8, 1}, // interleave 2 x 8i8 into 16i8 (and store) 5172 {2, MVT::v16i8, 3}, // interleave 2 x 16i8 into 32i8 (and store) 5173 {2, MVT::v32i8, 4}, // interleave 2 x 32i8 into 64i8 (and store) 5174 5175 {2, MVT::v2i16, 1}, // interleave 2 x 2i16 into 4i16 (and store) 5176 {2, MVT::v4i16, 1}, // interleave 2 x 4i16 into 8i16 (and store) 5177 {2, MVT::v8i16, 3}, // interleave 2 x 8i16 into 16i16 (and store) 5178 {2, MVT::v16i16, 4}, // interleave 2 x 16i16 into 32i16 (and store) 5179 {2, MVT::v32i16, 8}, // interleave 2 x 32i16 into 64i16 (and store) 5180 5181 {2, MVT::v2i32, 1}, // interleave 2 x 2i32 into 4i32 (and store) 5182 {2, MVT::v4i32, 2}, // interleave 2 x 4i32 into 8i32 (and store) 5183 {2, MVT::v8i32, 4}, // interleave 2 x 8i32 into 16i32 (and store) 5184 {2, MVT::v16i32, 8}, // interleave 2 x 16i32 into 32i32 (and store) 5185 {2, MVT::v32i32, 16}, // interleave 2 x 32i32 into 64i32 (and store) 5186 5187 {2, MVT::v2i64, 2}, // interleave 2 x 2i64 into 4i64 (and store) 5188 {2, MVT::v4i64, 4}, // interleave 2 x 4i64 into 8i64 (and store) 5189 {2, MVT::v8i64, 8}, // interleave 2 x 8i64 into 16i64 (and store) 5190 {2, MVT::v16i64, 16}, // interleave 2 x 16i64 into 32i64 (and store) 5191 5192 {3, MVT::v2i8, 4}, // interleave 3 x 2i8 into 6i8 (and store) 5193 {3, MVT::v4i8, 4}, // interleave 3 x 4i8 into 12i8 (and store) 5194 {3, MVT::v8i8, 6}, // interleave 3 x 8i8 into 24i8 (and store) 5195 {3, MVT::v16i8, 11}, // interleave 3 x 16i8 into 48i8 (and store) 5196 {3, MVT::v32i8, 13}, // interleave 3 x 32i8 into 96i8 (and store) 5197 5198 {3, MVT::v2i16, 4}, // interleave 3 x 2i16 into 6i16 (and store) 5199 {3, MVT::v4i16, 6}, // interleave 3 x 4i16 into 12i16 (and store) 5200 {3, MVT::v8i16, 12}, // interleave 3 x 8i16 into 24i16 (and store) 5201 {3, MVT::v16i16, 27}, // interleave 3 x 16i16 into 48i16 (and store) 5202 {3, MVT::v32i16, 54}, // interleave 3 x 32i16 into 96i16 (and store) 5203 5204 {3, MVT::v2i32, 4}, // interleave 3 x 2i32 into 6i32 (and store) 5205 {3, MVT::v4i32, 5}, // interleave 3 x 4i32 into 12i32 (and store) 5206 {3, MVT::v8i32, 11}, // interleave 3 x 8i32 into 24i32 (and store) 5207 {3, MVT::v16i32, 22}, // interleave 3 x 16i32 into 48i32 (and store) 5208 5209 {3, MVT::v2i64, 4}, // interleave 3 x 2i64 into 6i64 (and store) 5210 {3, MVT::v4i64, 6}, // interleave 3 x 4i64 into 12i64 (and store) 5211 {3, MVT::v8i64, 12}, // interleave 3 x 8i64 into 24i64 (and store) 5212 {3, MVT::v16i64, 24}, // interleave 3 x 16i64 into 48i64 (and store) 5213 5214 {4, MVT::v2i8, 4}, // interleave 4 x 2i8 into 8i8 (and store) 5215 {4, MVT::v4i8, 4}, // interleave 4 x 4i8 into 16i8 (and store) 5216 {4, MVT::v8i8, 4}, // interleave 4 x 8i8 into 32i8 (and store) 5217 {4, MVT::v16i8, 8}, // interleave 4 x 16i8 into 64i8 (and store) 5218 {4, MVT::v32i8, 12}, // interleave 4 x 32i8 into 128i8 (and store) 5219 5220 {4, MVT::v2i16, 2}, // interleave 4 x 2i16 into 8i16 (and store) 5221 {4, MVT::v4i16, 6}, // interleave 4 x 4i16 into 16i16 (and store) 5222 {4, MVT::v8i16, 10}, // interleave 4 x 8i16 into 32i16 (and store) 5223 {4, MVT::v16i16, 32}, // interleave 4 x 16i16 into 64i16 (and store) 5224 {4, MVT::v32i16, 64}, // interleave 4 x 32i16 into 128i16 (and store) 5225 5226 {4, MVT::v2i32, 5}, // interleave 4 x 2i32 into 8i32 (and store) 5227 {4, MVT::v4i32, 6}, // interleave 4 x 4i32 into 16i32 (and store) 5228 {4, MVT::v8i32, 16}, // interleave 4 x 8i32 into 32i32 (and store) 5229 {4, MVT::v16i32, 32}, // interleave 4 x 16i32 into 64i32 (and store) 5230 5231 {4, MVT::v2i64, 6}, // interleave 4 x 2i64 into 8i64 (and store) 5232 {4, MVT::v4i64, 8}, // interleave 4 x 4i64 into 16i64 (and store) 5233 {4, MVT::v8i64, 20}, // interleave 4 x 8i64 into 32i64 (and store) 5234 5235 {6, MVT::v2i8, 7}, // interleave 6 x 2i8 into 12i8 (and store) 5236 {6, MVT::v4i8, 9}, // interleave 6 x 4i8 into 24i8 (and store) 5237 {6, MVT::v8i8, 16}, // interleave 6 x 8i8 into 48i8 (and store) 5238 {6, MVT::v16i8, 27}, // interleave 6 x 16i8 into 96i8 (and store) 5239 {6, MVT::v32i8, 90}, // interleave 6 x 32i8 into 192i8 (and store) 5240 5241 {6, MVT::v2i16, 10}, // interleave 6 x 2i16 into 12i16 (and store) 5242 {6, MVT::v4i16, 15}, // interleave 6 x 4i16 into 24i16 (and store) 5243 {6, MVT::v8i16, 21}, // interleave 6 x 8i16 into 48i16 (and store) 5244 {6, MVT::v16i16, 58}, // interleave 6 x 16i16 into 96i16 (and store) 5245 5246 {6, MVT::v2i32, 9}, // interleave 6 x 2i32 into 12i32 (and store) 5247 {6, MVT::v4i32, 12}, // interleave 6 x 4i32 into 24i32 (and store) 5248 {6, MVT::v8i32, 33}, // interleave 6 x 8i32 into 48i32 (and store) 5249 {6, MVT::v16i32, 66}, // interleave 6 x 16i32 into 96i32 (and store) 5250 5251 {6, MVT::v2i64, 8}, // interleave 6 x 2i64 into 12i64 (and store) 5252 {6, MVT::v4i64, 15}, // interleave 6 x 4i64 into 24i64 (and store) 5253 {6, MVT::v8i64, 30}, // interleave 6 x 8i64 into 48i64 (and store) 5254 }; 5255 5256 if (Opcode == Instruction::Load) { 5257 if (const auto *Entry = 5258 CostTableLookup(AVX2InterleavedLoadTbl, Factor, ETy.getSimpleVT())) 5259 return MemOpCosts + Entry->Cost; 5260 } else { 5261 assert(Opcode == Instruction::Store && 5262 "Expected Store Instruction at this point"); 5263 if (const auto *Entry = 5264 CostTableLookup(AVX2InterleavedStoreTbl, Factor, ETy.getSimpleVT())) 5265 return MemOpCosts + Entry->Cost; 5266 } 5267 5268 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5269 Alignment, AddressSpace, CostKind); 5270 } 5271 5272 // Get estimation for interleaved load/store operations and strided load. 5273 // \p Indices contains indices for strided load. 5274 // \p Factor - the factor of interleaving. 5275 // AVX-512 provides 3-src shuffles that significantly reduces the cost. 5276 InstructionCost X86TTIImpl::getInterleavedMemoryOpCostAVX512( 5277 unsigned Opcode, FixedVectorType *VecTy, unsigned Factor, 5278 ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace, 5279 TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) { 5280 5281 if (UseMaskForCond || UseMaskForGaps) 5282 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5283 Alignment, AddressSpace, CostKind, 5284 UseMaskForCond, UseMaskForGaps); 5285 5286 // VecTy for interleave memop is <VF*Factor x Elt>. 5287 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have 5288 // VecTy = <12 x i32>. 5289 5290 // Calculate the number of memory operations (NumOfMemOps), required 5291 // for load/store the VecTy. 5292 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second; 5293 unsigned VecTySize = DL.getTypeStoreSize(VecTy); 5294 unsigned LegalVTSize = LegalVT.getStoreSize(); 5295 unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize; 5296 5297 // Get the cost of one memory operation. 5298 auto *SingleMemOpTy = FixedVectorType::get(VecTy->getElementType(), 5299 LegalVT.getVectorNumElements()); 5300 InstructionCost MemOpCost = getMemoryOpCost( 5301 Opcode, SingleMemOpTy, MaybeAlign(Alignment), AddressSpace, CostKind); 5302 5303 unsigned VF = VecTy->getNumElements() / Factor; 5304 MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF); 5305 5306 if (Opcode == Instruction::Load) { 5307 // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl) 5308 // contain the cost of the optimized shuffle sequence that the 5309 // X86InterleavedAccess pass will generate. 5310 // The cost of loads and stores are computed separately from the table. 5311 5312 // X86InterleavedAccess support only the following interleaved-access group. 5313 static const CostTblEntry AVX512InterleavedLoadTbl[] = { 5314 {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8 5315 {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8 5316 {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8 5317 }; 5318 5319 if (const auto *Entry = 5320 CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT)) 5321 return NumOfMemOps * MemOpCost + Entry->Cost; 5322 //If an entry does not exist, fallback to the default implementation. 5323 5324 // Kind of shuffle depends on number of loaded values. 5325 // If we load the entire data in one register, we can use a 1-src shuffle. 5326 // Otherwise, we'll merge 2 sources in each operation. 5327 TTI::ShuffleKind ShuffleKind = 5328 (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc; 5329 5330 InstructionCost ShuffleCost = 5331 getShuffleCost(ShuffleKind, SingleMemOpTy, None, 0, nullptr); 5332 5333 unsigned NumOfLoadsInInterleaveGrp = 5334 Indices.size() ? Indices.size() : Factor; 5335 auto *ResultTy = FixedVectorType::get(VecTy->getElementType(), 5336 VecTy->getNumElements() / Factor); 5337 InstructionCost NumOfResults = 5338 getTLI()->getTypeLegalizationCost(DL, ResultTy).first * 5339 NumOfLoadsInInterleaveGrp; 5340 5341 // About a half of the loads may be folded in shuffles when we have only 5342 // one result. If we have more than one result, we do not fold loads at all. 5343 unsigned NumOfUnfoldedLoads = 5344 NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2; 5345 5346 // Get a number of shuffle operations per result. 5347 unsigned NumOfShufflesPerResult = 5348 std::max((unsigned)1, (unsigned)(NumOfMemOps - 1)); 5349 5350 // The SK_MergeTwoSrc shuffle clobbers one of src operands. 5351 // When we have more than one destination, we need additional instructions 5352 // to keep sources. 5353 InstructionCost NumOfMoves = 0; 5354 if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc) 5355 NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2; 5356 5357 InstructionCost Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost + 5358 NumOfUnfoldedLoads * MemOpCost + NumOfMoves; 5359 5360 return Cost; 5361 } 5362 5363 // Store. 5364 assert(Opcode == Instruction::Store && 5365 "Expected Store Instruction at this point"); 5366 // X86InterleavedAccess support only the following interleaved-access group. 5367 static const CostTblEntry AVX512InterleavedStoreTbl[] = { 5368 {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store) 5369 {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store) 5370 {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store) 5371 5372 {4, MVT::v8i8, 10}, // interleave 4 x 8i8 into 32i8 (and store) 5373 {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8 (and store) 5374 {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store) 5375 {4, MVT::v64i8, 24} // interleave 4 x 32i8 into 256i8 (and store) 5376 }; 5377 5378 if (const auto *Entry = 5379 CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT)) 5380 return NumOfMemOps * MemOpCost + Entry->Cost; 5381 //If an entry does not exist, fallback to the default implementation. 5382 5383 // There is no strided stores meanwhile. And store can't be folded in 5384 // shuffle. 5385 unsigned NumOfSources = Factor; // The number of values to be merged. 5386 InstructionCost ShuffleCost = 5387 getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, None, 0, nullptr); 5388 unsigned NumOfShufflesPerStore = NumOfSources - 1; 5389 5390 // The SK_MergeTwoSrc shuffle clobbers one of src operands. 5391 // We need additional instructions to keep sources. 5392 unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2; 5393 InstructionCost Cost = 5394 NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) + 5395 NumOfMoves; 5396 return Cost; 5397 } 5398 5399 InstructionCost X86TTIImpl::getInterleavedMemoryOpCost( 5400 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 5401 Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, 5402 bool UseMaskForCond, bool UseMaskForGaps) { 5403 auto isSupportedOnAVX512 = [&](Type *VecTy, bool HasBW) { 5404 Type *EltTy = cast<VectorType>(VecTy)->getElementType(); 5405 if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) || 5406 EltTy->isIntegerTy(32) || EltTy->isPointerTy()) 5407 return true; 5408 if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8) || 5409 (!ST->useSoftFloat() && ST->hasFP16() && EltTy->isHalfTy())) 5410 return HasBW; 5411 return false; 5412 }; 5413 if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI())) 5414 return getInterleavedMemoryOpCostAVX512( 5415 Opcode, cast<FixedVectorType>(VecTy), Factor, Indices, Alignment, 5416 AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps); 5417 if (ST->hasAVX2()) 5418 return getInterleavedMemoryOpCostAVX2( 5419 Opcode, cast<FixedVectorType>(VecTy), Factor, Indices, Alignment, 5420 AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps); 5421 5422 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 5423 Alignment, AddressSpace, CostKind, 5424 UseMaskForCond, UseMaskForGaps); 5425 } 5426