| 1 | //===-- WebAssemblyTargetTransformInfo.cpp - WebAssembly-specific TTI -----===// |
| 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 | /// |
| 9 | /// \file |
| 10 | /// This file defines the WebAssembly-specific TargetTransformInfo |
| 11 | /// implementation. |
| 12 | /// |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #include "WebAssemblyTargetTransformInfo.h" |
| 16 | #include "llvm/IR/IntrinsicInst.h" |
| 17 | #include "llvm/IR/IntrinsicsWebAssembly.h" |
| 18 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
| 19 | |
| 20 | #include "llvm/CodeGen/CostTable.h" |
| 21 | using namespace llvm; |
| 22 | |
| 23 | #define DEBUG_TYPE "wasmtti" |
| 24 | |
| 25 | TargetTransformInfo::PopcntSupportKind |
| 26 | WebAssemblyTTIImpl::getPopcntSupport(unsigned TyWidth) const { |
| 27 | assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2" ); |
| 28 | return TargetTransformInfo::PSK_FastHardware; |
| 29 | } |
| 30 | |
| 31 | unsigned WebAssemblyTTIImpl::getNumberOfRegisters(unsigned ClassID) const { |
| 32 | unsigned Result = BaseT::getNumberOfRegisters(ClassID); |
| 33 | |
| 34 | // For SIMD, use at least 16 registers, as a rough guess. |
| 35 | bool Vector = (ClassID == 1); |
| 36 | if (Vector) |
| 37 | Result = std::max(a: Result, b: 16u); |
| 38 | |
| 39 | return Result; |
| 40 | } |
| 41 | |
| 42 | TypeSize WebAssemblyTTIImpl::getRegisterBitWidth( |
| 43 | TargetTransformInfo::RegisterKind K) const { |
| 44 | switch (K) { |
| 45 | case TargetTransformInfo::RGK_Scalar: |
| 46 | return TypeSize::getFixed(ExactSize: 64); |
| 47 | case TargetTransformInfo::RGK_FixedWidthVector: |
| 48 | return TypeSize::getFixed(ExactSize: getST()->hasSIMD128() ? 128 : 64); |
| 49 | case TargetTransformInfo::RGK_ScalableVector: |
| 50 | return TypeSize::getScalable(MinimumSize: 0); |
| 51 | } |
| 52 | |
| 53 | llvm_unreachable("Unsupported register kind" ); |
| 54 | } |
| 55 | |
| 56 | InstructionCost WebAssemblyTTIImpl::getArithmeticInstrCost( |
| 57 | unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, |
| 58 | TTI::OperandValueInfo Op1Info, TTI::OperandValueInfo Op2Info, |
| 59 | ArrayRef<const Value *> Args, const Instruction *CxtI) const { |
| 60 | |
| 61 | if (ST->hasSIMD128()) { |
| 62 | static const CostTblEntry ArithCostTbl[]{ |
| 63 | // extmul + (maybe awkward) shuffle |
| 64 | {.ISD: ISD::MUL, .Type: MVT::v8i8, .Cost: 4}, |
| 65 | // 2x extmul + (okay) shuffle |
| 66 | {.ISD: ISD::MUL, .Type: MVT::v16i8, .Cost: 4}, |
| 67 | // extmul |
| 68 | {.ISD: ISD::MUL, .Type: MVT::v4i16, .Cost: 1}, |
| 69 | // extmul |
| 70 | {.ISD: ISD::MUL, .Type: MVT::v2i32, .Cost: 1}, |
| 71 | }; |
| 72 | EVT DstVT = TLI->getValueType(DL, Ty); |
| 73 | if (DstVT.isSimple()) { |
| 74 | int ISD = TLI->InstructionOpcodeToISD(Opcode); |
| 75 | if (const auto *Entry = |
| 76 | CostTableLookup(Table: ArithCostTbl, ISD, Ty: DstVT.getSimpleVT())) |
| 77 | return Entry->Cost; |
| 78 | } |
| 79 | } |
| 80 | |
| 81 | InstructionCost Cost = |
| 82 | BasicTTIImplBase<WebAssemblyTTIImpl>::getArithmeticInstrCost( |
| 83 | Opcode, Ty, CostKind, Opd1Info: Op1Info, Opd2Info: Op2Info); |
| 84 | |
| 85 | if (auto *VTy = dyn_cast<VectorType>(Val: Ty)) { |
| 86 | switch (Opcode) { |
| 87 | case Instruction::LShr: |
| 88 | case Instruction::AShr: |
| 89 | case Instruction::Shl: |
| 90 | // SIMD128's shifts currently only accept a scalar shift count. For each |
| 91 | // element, we'll need to extract, op, insert. The following is a rough |
| 92 | // approximation. |
| 93 | if (!Op2Info.isUniform()) |
| 94 | Cost = |
| 95 | cast<FixedVectorType>(Val: VTy)->getNumElements() * |
| 96 | (TargetTransformInfo::TCC_Basic + |
| 97 | getArithmeticInstrCost(Opcode, Ty: VTy->getElementType(), CostKind) + |
| 98 | TargetTransformInfo::TCC_Basic); |
| 99 | break; |
| 100 | } |
| 101 | } |
| 102 | return Cost; |
| 103 | } |
| 104 | |
| 105 | InstructionCost WebAssemblyTTIImpl::getCastInstrCost( |
| 106 | unsigned Opcode, Type *Dst, Type *Src, TTI::CastContextHint CCH, |
| 107 | TTI::TargetCostKind CostKind, const Instruction *I) const { |
| 108 | int ISD = TLI->InstructionOpcodeToISD(Opcode); |
| 109 | auto SrcTy = TLI->getValueType(DL, Ty: Src); |
| 110 | auto DstTy = TLI->getValueType(DL, Ty: Dst); |
| 111 | |
| 112 | if (!SrcTy.isSimple() || !DstTy.isSimple()) { |
| 113 | return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I); |
| 114 | } |
| 115 | |
| 116 | if (!ST->hasSIMD128()) { |
| 117 | return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I); |
| 118 | } |
| 119 | |
| 120 | auto DstVT = DstTy.getSimpleVT(); |
| 121 | auto SrcVT = SrcTy.getSimpleVT(); |
| 122 | |
| 123 | if (I && I->hasOneUser()) { |
| 124 | auto *SingleUser = cast<Instruction>(Val: *I->user_begin()); |
| 125 | int UserISD = TLI->InstructionOpcodeToISD(Opcode: SingleUser->getOpcode()); |
| 126 | |
| 127 | // extmul_low support |
| 128 | if (UserISD == ISD::MUL && |
| 129 | (ISD == ISD::ZERO_EXTEND || ISD == ISD::SIGN_EXTEND)) { |
| 130 | // Free low extensions. |
| 131 | if ((SrcVT == MVT::v8i8 && DstVT == MVT::v8i16) || |
| 132 | (SrcVT == MVT::v4i16 && DstVT == MVT::v4i32) || |
| 133 | (SrcVT == MVT::v2i32 && DstVT == MVT::v2i64)) { |
| 134 | return 0; |
| 135 | } |
| 136 | // Will require an additional extlow operation for the intermediate |
| 137 | // i16/i32 value. |
| 138 | if ((SrcVT == MVT::v4i8 && DstVT == MVT::v4i32) || |
| 139 | (SrcVT == MVT::v2i16 && DstVT == MVT::v2i64)) { |
| 140 | return 1; |
| 141 | } |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | static constexpr TypeConversionCostTblEntry ConversionTbl[] = { |
| 146 | // extend_low |
| 147 | {.ISD: ISD::SIGN_EXTEND, .Dst: MVT::v2i64, .Src: MVT::v2i32, .Cost: 1}, |
| 148 | {.ISD: ISD::ZERO_EXTEND, .Dst: MVT::v2i64, .Src: MVT::v2i32, .Cost: 1}, |
| 149 | {.ISD: ISD::SIGN_EXTEND, .Dst: MVT::v4i32, .Src: MVT::v4i16, .Cost: 1}, |
| 150 | {.ISD: ISD::ZERO_EXTEND, .Dst: MVT::v4i32, .Src: MVT::v4i16, .Cost: 1}, |
| 151 | {.ISD: ISD::SIGN_EXTEND, .Dst: MVT::v8i16, .Src: MVT::v8i8, .Cost: 1}, |
| 152 | {.ISD: ISD::ZERO_EXTEND, .Dst: MVT::v8i16, .Src: MVT::v8i8, .Cost: 1}, |
| 153 | // 2 x extend_low |
| 154 | {.ISD: ISD::SIGN_EXTEND, .Dst: MVT::v2i64, .Src: MVT::v2i16, .Cost: 2}, |
| 155 | {.ISD: ISD::ZERO_EXTEND, .Dst: MVT::v2i64, .Src: MVT::v2i16, .Cost: 2}, |
| 156 | {.ISD: ISD::SIGN_EXTEND, .Dst: MVT::v4i32, .Src: MVT::v4i8, .Cost: 2}, |
| 157 | {.ISD: ISD::ZERO_EXTEND, .Dst: MVT::v4i32, .Src: MVT::v4i8, .Cost: 2}, |
| 158 | // extend_low, extend_high |
| 159 | {.ISD: ISD::SIGN_EXTEND, .Dst: MVT::v4i64, .Src: MVT::v4i32, .Cost: 2}, |
| 160 | {.ISD: ISD::ZERO_EXTEND, .Dst: MVT::v4i64, .Src: MVT::v4i32, .Cost: 2}, |
| 161 | {.ISD: ISD::SIGN_EXTEND, .Dst: MVT::v8i32, .Src: MVT::v8i16, .Cost: 2}, |
| 162 | {.ISD: ISD::ZERO_EXTEND, .Dst: MVT::v8i32, .Src: MVT::v8i16, .Cost: 2}, |
| 163 | {.ISD: ISD::SIGN_EXTEND, .Dst: MVT::v16i16, .Src: MVT::v16i8, .Cost: 2}, |
| 164 | {.ISD: ISD::ZERO_EXTEND, .Dst: MVT::v16i16, .Src: MVT::v16i8, .Cost: 2}, |
| 165 | // 2x extend_low, extend_high |
| 166 | {.ISD: ISD::SIGN_EXTEND, .Dst: MVT::v8i64, .Src: MVT::v8i32, .Cost: 4}, |
| 167 | {.ISD: ISD::ZERO_EXTEND, .Dst: MVT::v8i64, .Src: MVT::v8i32, .Cost: 4}, |
| 168 | {.ISD: ISD::SIGN_EXTEND, .Dst: MVT::v16i32, .Src: MVT::v16i16, .Cost: 4}, |
| 169 | {.ISD: ISD::ZERO_EXTEND, .Dst: MVT::v16i32, .Src: MVT::v16i16, .Cost: 4}, |
| 170 | // shuffle |
| 171 | {.ISD: ISD::TRUNCATE, .Dst: MVT::v2i16, .Src: MVT::v2i32, .Cost: 2}, |
| 172 | {.ISD: ISD::TRUNCATE, .Dst: MVT::v2i8, .Src: MVT::v2i32, .Cost: 4}, |
| 173 | {.ISD: ISD::TRUNCATE, .Dst: MVT::v4i16, .Src: MVT::v4i32, .Cost: 2}, |
| 174 | {.ISD: ISD::TRUNCATE, .Dst: MVT::v4i8, .Src: MVT::v4i32, .Cost: 4}, |
| 175 | // narrow, and |
| 176 | {.ISD: ISD::TRUNCATE, .Dst: MVT::v8i16, .Src: MVT::v8i32, .Cost: 2}, |
| 177 | {.ISD: ISD::TRUNCATE, .Dst: MVT::v8i8, .Src: MVT::v8i16, .Cost: 2}, |
| 178 | // narrow, 2x and |
| 179 | {.ISD: ISD::TRUNCATE, .Dst: MVT::v16i8, .Src: MVT::v16i16, .Cost: 3}, |
| 180 | // 3x narrow, 4x and |
| 181 | {.ISD: ISD::TRUNCATE, .Dst: MVT::v8i16, .Src: MVT::v8i64, .Cost: 7}, |
| 182 | {.ISD: ISD::TRUNCATE, .Dst: MVT::v16i8, .Src: MVT::v16i32, .Cost: 7}, |
| 183 | // 7x narrow, 8x and |
| 184 | {.ISD: ISD::TRUNCATE, .Dst: MVT::v16i8, .Src: MVT::v16i64, .Cost: 15}, |
| 185 | // convert_i32x4 |
| 186 | {.ISD: ISD::SINT_TO_FP, .Dst: MVT::v2f32, .Src: MVT::v2i32, .Cost: 1}, |
| 187 | {.ISD: ISD::UINT_TO_FP, .Dst: MVT::v2f32, .Src: MVT::v2i32, .Cost: 1}, |
| 188 | {.ISD: ISD::SINT_TO_FP, .Dst: MVT::v4f32, .Src: MVT::v4i32, .Cost: 1}, |
| 189 | {.ISD: ISD::UINT_TO_FP, .Dst: MVT::v4f32, .Src: MVT::v4i32, .Cost: 1}, |
| 190 | // extend_low, convert |
| 191 | {.ISD: ISD::SINT_TO_FP, .Dst: MVT::v2f32, .Src: MVT::v2i16, .Cost: 2}, |
| 192 | {.ISD: ISD::UINT_TO_FP, .Dst: MVT::v2f32, .Src: MVT::v2i16, .Cost: 2}, |
| 193 | {.ISD: ISD::SINT_TO_FP, .Dst: MVT::v4f32, .Src: MVT::v4i16, .Cost: 2}, |
| 194 | {.ISD: ISD::UINT_TO_FP, .Dst: MVT::v4f32, .Src: MVT::v4i16, .Cost: 2}, |
| 195 | // extend_low x 2, convert |
| 196 | {.ISD: ISD::SINT_TO_FP, .Dst: MVT::v2f32, .Src: MVT::v2i8, .Cost: 3}, |
| 197 | {.ISD: ISD::UINT_TO_FP, .Dst: MVT::v2f32, .Src: MVT::v2i8, .Cost: 3}, |
| 198 | {.ISD: ISD::SINT_TO_FP, .Dst: MVT::v4f32, .Src: MVT::v4i8, .Cost: 3}, |
| 199 | {.ISD: ISD::UINT_TO_FP, .Dst: MVT::v4f32, .Src: MVT::v4i8, .Cost: 3}, |
| 200 | // several shuffles |
| 201 | {.ISD: ISD::SINT_TO_FP, .Dst: MVT::v8f32, .Src: MVT::v8i8, .Cost: 10}, |
| 202 | {.ISD: ISD::UINT_TO_FP, .Dst: MVT::v8f32, .Src: MVT::v8i8, .Cost: 10}, |
| 203 | {.ISD: ISD::SINT_TO_FP, .Dst: MVT::v8f32, .Src: MVT::v8i16, .Cost: 10}, |
| 204 | {.ISD: ISD::UINT_TO_FP, .Dst: MVT::v8f32, .Src: MVT::v8i8, .Cost: 10}, |
| 205 | /// trunc_sat, const, and, 3x narrow |
| 206 | {.ISD: ISD::FP_TO_SINT, .Dst: MVT::v2i8, .Src: MVT::v2f32, .Cost: 6}, |
| 207 | {.ISD: ISD::FP_TO_UINT, .Dst: MVT::v2i8, .Src: MVT::v2f32, .Cost: 6}, |
| 208 | {.ISD: ISD::FP_TO_SINT, .Dst: MVT::v4i8, .Src: MVT::v4f32, .Cost: 6}, |
| 209 | {.ISD: ISD::FP_TO_UINT, .Dst: MVT::v4i8, .Src: MVT::v4f32, .Cost: 6}, |
| 210 | /// trunc_sat, const, and, narrow |
| 211 | {.ISD: ISD::FP_TO_UINT, .Dst: MVT::v2i16, .Src: MVT::v2f32, .Cost: 4}, |
| 212 | {.ISD: ISD::FP_TO_SINT, .Dst: MVT::v2i16, .Src: MVT::v2f32, .Cost: 4}, |
| 213 | {.ISD: ISD::FP_TO_SINT, .Dst: MVT::v4i16, .Src: MVT::v4f32, .Cost: 4}, |
| 214 | {.ISD: ISD::FP_TO_UINT, .Dst: MVT::v4i16, .Src: MVT::v4f32, .Cost: 4}, |
| 215 | // 2x trunc_sat, const, 2x and, 3x narrow |
| 216 | {.ISD: ISD::FP_TO_SINT, .Dst: MVT::v8i8, .Src: MVT::v8f32, .Cost: 8}, |
| 217 | {.ISD: ISD::FP_TO_UINT, .Dst: MVT::v8i8, .Src: MVT::v8f32, .Cost: 8}, |
| 218 | // 2x trunc_sat, const, 2x and, narrow |
| 219 | {.ISD: ISD::FP_TO_SINT, .Dst: MVT::v8i16, .Src: MVT::v8f32, .Cost: 6}, |
| 220 | {.ISD: ISD::FP_TO_UINT, .Dst: MVT::v8i16, .Src: MVT::v8f32, .Cost: 6}, |
| 221 | }; |
| 222 | |
| 223 | if (const auto *Entry = |
| 224 | ConvertCostTableLookup(Table: ConversionTbl, ISD, Dst: DstVT, Src: SrcVT)) { |
| 225 | return Entry->Cost; |
| 226 | } |
| 227 | |
| 228 | return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I); |
| 229 | } |
| 230 | |
| 231 | WebAssemblyTTIImpl::TTI::MemCmpExpansionOptions |
| 232 | WebAssemblyTTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { |
| 233 | TTI::MemCmpExpansionOptions Options; |
| 234 | |
| 235 | Options.AllowOverlappingLoads = true; |
| 236 | |
| 237 | if (ST->hasSIMD128()) |
| 238 | Options.LoadSizes.push_back(Elt: 16); |
| 239 | |
| 240 | Options.LoadSizes.append(IL: {8, 4, 2, 1}); |
| 241 | Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize); |
| 242 | Options.NumLoadsPerBlock = Options.MaxNumLoads; |
| 243 | |
| 244 | return Options; |
| 245 | } |
| 246 | |
| 247 | InstructionCost WebAssemblyTTIImpl::getMemoryOpCost( |
| 248 | unsigned Opcode, Type *Ty, Align Alignment, unsigned AddressSpace, |
| 249 | TTI::TargetCostKind CostKind, TTI::OperandValueInfo OpInfo, |
| 250 | const Instruction *I) const { |
| 251 | if (!ST->hasSIMD128() || !isa<FixedVectorType>(Val: Ty)) { |
| 252 | return BaseT::getMemoryOpCost(Opcode, Src: Ty, Alignment, AddressSpace, |
| 253 | CostKind); |
| 254 | } |
| 255 | |
| 256 | EVT VT = TLI->getValueType(DL, Ty, AllowUnknown: true); |
| 257 | // Type legalization can't handle structs |
| 258 | if (VT == MVT::Other) |
| 259 | return BaseT::getMemoryOpCost(Opcode, Src: Ty, Alignment, AddressSpace, |
| 260 | CostKind); |
| 261 | |
| 262 | auto LT = getTypeLegalizationCost(Ty); |
| 263 | if (!LT.first.isValid()) |
| 264 | return InstructionCost::getInvalid(); |
| 265 | |
| 266 | int ISD = TLI->InstructionOpcodeToISD(Opcode); |
| 267 | unsigned width = VT.getSizeInBits(); |
| 268 | if (ISD == ISD::LOAD) { |
| 269 | // 128-bit loads are a single instruction. 32-bit and 64-bit vector loads |
| 270 | // can be lowered to load32_zero and load64_zero respectively. Assume SIMD |
| 271 | // loads are twice as expensive as scalar. |
| 272 | switch (width) { |
| 273 | default: |
| 274 | break; |
| 275 | case 32: |
| 276 | case 64: |
| 277 | case 128: |
| 278 | return 2; |
| 279 | } |
| 280 | } else if (ISD == ISD::STORE) { |
| 281 | // For stores, we can use store lane operations. |
| 282 | switch (width) { |
| 283 | default: |
| 284 | break; |
| 285 | case 8: |
| 286 | case 16: |
| 287 | case 32: |
| 288 | case 64: |
| 289 | case 128: |
| 290 | return 2; |
| 291 | } |
| 292 | } |
| 293 | |
| 294 | return BaseT::getMemoryOpCost(Opcode, Src: Ty, Alignment, AddressSpace, CostKind); |
| 295 | } |
| 296 | |
| 297 | InstructionCost WebAssemblyTTIImpl::getInterleavedMemoryOpCost( |
| 298 | unsigned Opcode, Type *Ty, unsigned Factor, ArrayRef<unsigned> Indices, |
| 299 | Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, |
| 300 | bool UseMaskForCond, bool UseMaskForGaps) const { |
| 301 | assert(Factor >= 2 && "Invalid interleave factor" ); |
| 302 | |
| 303 | auto *VecTy = cast<VectorType>(Val: Ty); |
| 304 | if (!ST->hasSIMD128() || !isa<FixedVectorType>(Val: VecTy)) { |
| 305 | return InstructionCost::getInvalid(); |
| 306 | } |
| 307 | |
| 308 | if (UseMaskForCond || UseMaskForGaps) |
| 309 | return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy: Ty, Factor, Indices, |
| 310 | Alignment, AddressSpace, CostKind, |
| 311 | UseMaskForCond, UseMaskForGaps); |
| 312 | |
| 313 | constexpr unsigned MaxInterleaveFactor = 4; |
| 314 | if (Factor <= MaxInterleaveFactor) { |
| 315 | unsigned MinElts = VecTy->getElementCount().getKnownMinValue(); |
| 316 | // Ensure the number of vector elements is greater than 1. |
| 317 | if (MinElts < 2 || MinElts % Factor != 0) |
| 318 | return InstructionCost::getInvalid(); |
| 319 | |
| 320 | unsigned ElSize = DL.getTypeSizeInBits(Ty: VecTy->getElementType()); |
| 321 | // Ensure the element type is legal. |
| 322 | if (ElSize != 8 && ElSize != 16 && ElSize != 32 && ElSize != 64) |
| 323 | return InstructionCost::getInvalid(); |
| 324 | |
| 325 | if (Factor != 2 && Factor != 4) |
| 326 | return InstructionCost::getInvalid(); |
| 327 | |
| 328 | auto *SubVecTy = |
| 329 | VectorType::get(ElementType: VecTy->getElementType(), |
| 330 | EC: VecTy->getElementCount().divideCoefficientBy(RHS: Factor)); |
| 331 | InstructionCost MemCost = |
| 332 | getMemoryOpCost(Opcode, Ty: SubVecTy, Alignment, AddressSpace, CostKind); |
| 333 | |
| 334 | unsigned VecSize = DL.getTypeSizeInBits(Ty: SubVecTy); |
| 335 | unsigned MaxVecSize = 128; |
| 336 | unsigned NumAccesses = |
| 337 | std::max<unsigned>(a: 1, b: (MinElts * ElSize + MaxVecSize - 1) / VecSize); |
| 338 | |
| 339 | // A stride of two is commonly supported via dedicated instructions, so it |
| 340 | // should be relatively cheap for all element sizes. A stride of four is |
| 341 | // more expensive as it will likely require more shuffles. Using two |
| 342 | // simd128 inputs is considered more expensive and we mainly account for |
| 343 | // shuffling two inputs (32 bytes), but we do model 4 x v4i32 to enable |
| 344 | // arithmetic kernels. |
| 345 | static const CostTblEntry ShuffleCostTbl[] = { |
| 346 | // One reg. |
| 347 | {.ISD: 2, .Type: MVT::v2i8, .Cost: 1}, // interleave 2 x 2i8 into 4i8 |
| 348 | {.ISD: 2, .Type: MVT::v4i8, .Cost: 1}, // interleave 2 x 4i8 into 8i8 |
| 349 | {.ISD: 2, .Type: MVT::v8i8, .Cost: 1}, // interleave 2 x 8i8 into 16i8 |
| 350 | {.ISD: 2, .Type: MVT::v2i16, .Cost: 1}, // interleave 2 x 2i16 into 4i16 |
| 351 | {.ISD: 2, .Type: MVT::v4i16, .Cost: 1}, // interleave 2 x 4i16 into 8i16 |
| 352 | {.ISD: 2, .Type: MVT::v2i32, .Cost: 1}, // interleave 2 x 2i32 into 4i32 |
| 353 | |
| 354 | // Two regs. |
| 355 | {.ISD: 2, .Type: MVT::v16i8, .Cost: 2}, // interleave 2 x 16i8 into 32i8 |
| 356 | {.ISD: 2, .Type: MVT::v8i16, .Cost: 2}, // interleave 2 x 8i16 into 16i16 |
| 357 | {.ISD: 2, .Type: MVT::v4i32, .Cost: 2}, // interleave 2 x 4i32 into 8i32 |
| 358 | |
| 359 | // One reg. |
| 360 | {.ISD: 4, .Type: MVT::v2i8, .Cost: 4}, // interleave 4 x 2i8 into 8i8 |
| 361 | {.ISD: 4, .Type: MVT::v4i8, .Cost: 4}, // interleave 4 x 4i8 into 16i8 |
| 362 | {.ISD: 4, .Type: MVT::v2i16, .Cost: 4}, // interleave 4 x 2i16 into 8i16 |
| 363 | |
| 364 | // Two regs. |
| 365 | {.ISD: 4, .Type: MVT::v8i8, .Cost: 16}, // interleave 4 x 8i8 into 32i8 |
| 366 | {.ISD: 4, .Type: MVT::v4i16, .Cost: 8}, // interleave 4 x 4i16 into 16i16 |
| 367 | {.ISD: 4, .Type: MVT::v2i32, .Cost: 4}, // interleave 4 x 2i32 into 8i32 |
| 368 | |
| 369 | // Four regs. |
| 370 | {.ISD: 4, .Type: MVT::v4i32, .Cost: 16}, // interleave 4 x 4i32 into 16i32 |
| 371 | }; |
| 372 | |
| 373 | EVT ETy = TLI->getValueType(DL, Ty: SubVecTy); |
| 374 | if (const auto *Entry = |
| 375 | CostTableLookup(Table: ShuffleCostTbl, ISD: Factor, Ty: ETy.getSimpleVT())) |
| 376 | return Entry->Cost + (NumAccesses * MemCost); |
| 377 | } |
| 378 | |
| 379 | return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, |
| 380 | Alignment, AddressSpace, CostKind, |
| 381 | UseMaskForCond, UseMaskForGaps); |
| 382 | } |
| 383 | |
| 384 | InstructionCost WebAssemblyTTIImpl::getVectorInstrCost( |
| 385 | unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index, |
| 386 | const Value *Op0, const Value *Op1, TTI::VectorInstrContext VIC) const { |
| 387 | InstructionCost Cost = BasicTTIImplBase::getVectorInstrCost( |
| 388 | Opcode, Val, CostKind, Index, Op0, Op1, VIC); |
| 389 | |
| 390 | // SIMD128's insert/extract currently only take constant indices. |
| 391 | if (Index == -1u) |
| 392 | return Cost + 25 * TargetTransformInfo::TCC_Expensive; |
| 393 | |
| 394 | return Cost; |
| 395 | } |
| 396 | |
| 397 | InstructionCost WebAssemblyTTIImpl::getPartialReductionCost( |
| 398 | unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType, |
| 399 | ElementCount VF, TTI::PartialReductionExtendKind OpAExtend, |
| 400 | TTI::PartialReductionExtendKind OpBExtend, std::optional<unsigned> BinOp, |
| 401 | TTI::TargetCostKind CostKind, std::optional<FastMathFlags> FMF) const { |
| 402 | InstructionCost Invalid = InstructionCost::getInvalid(); |
| 403 | if (!VF.isFixed() || !ST->hasSIMD128()) |
| 404 | return Invalid; |
| 405 | |
| 406 | if (CostKind != TTI::TCK_RecipThroughput) |
| 407 | return Invalid; |
| 408 | |
| 409 | if (Opcode != Instruction::Add) |
| 410 | return Invalid; |
| 411 | |
| 412 | EVT AccumEVT = EVT::getEVT(Ty: AccumType); |
| 413 | // TODO: Add i64 accumulator. |
| 414 | if (AccumEVT != MVT::i32) |
| 415 | return Invalid; |
| 416 | |
| 417 | // Possible options: |
| 418 | // - i16x8.extadd_pairwise_i8x16_sx |
| 419 | // - i32x4.extadd_pairwise_i16x8_sx |
| 420 | // - i32x4.dot_i16x8_s |
| 421 | // Only try to support dot, for now. |
| 422 | |
| 423 | EVT InputEVT = EVT::getEVT(Ty: InputTypeA); |
| 424 | if (!((InputEVT == MVT::i16 && VF.getFixedValue() == 8) || |
| 425 | (InputEVT == MVT::i8 && VF.getFixedValue() == 16))) { |
| 426 | return Invalid; |
| 427 | } |
| 428 | |
| 429 | if (OpAExtend == TTI::PR_None) |
| 430 | return Invalid; |
| 431 | |
| 432 | InstructionCost Cost(TTI::TCC_Basic); |
| 433 | if (!BinOp) |
| 434 | return Cost; |
| 435 | |
| 436 | if (OpAExtend != OpBExtend) |
| 437 | return Invalid; |
| 438 | |
| 439 | if (*BinOp != Instruction::Mul) |
| 440 | return Invalid; |
| 441 | |
| 442 | if (InputTypeA != InputTypeB) |
| 443 | return Invalid; |
| 444 | |
| 445 | // Signed inputs can lower to dot |
| 446 | if (InputEVT == MVT::i16 && VF.getFixedValue() == 8) |
| 447 | return OpAExtend == TTI::PR_SignExtend ? Cost : Cost * 2; |
| 448 | |
| 449 | // Double the size of the lowered sequence. |
| 450 | if (InputEVT == MVT::i8 && VF.getFixedValue() == 16) |
| 451 | return OpAExtend == TTI::PR_SignExtend ? Cost * 2 : Cost * 4; |
| 452 | |
| 453 | return Invalid; |
| 454 | } |
| 455 | |
| 456 | InstructionCost |
| 457 | WebAssemblyTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, |
| 458 | TTI::TargetCostKind CostKind) const { |
| 459 | switch (ICA.getID()) { |
| 460 | case Intrinsic::experimental_vector_extract_last_active: |
| 461 | // TODO: Remove once the intrinsic can be lowered without crashes. |
| 462 | return InstructionCost::getInvalid(); |
| 463 | default: |
| 464 | break; |
| 465 | } |
| 466 | return BaseT::getIntrinsicInstrCost(ICA, CostKind); |
| 467 | } |
| 468 | |
| 469 | TTI::ReductionShuffle WebAssemblyTTIImpl::getPreferredExpandedReductionShuffle( |
| 470 | const IntrinsicInst *II) const { |
| 471 | |
| 472 | switch (II->getIntrinsicID()) { |
| 473 | default: |
| 474 | break; |
| 475 | case Intrinsic::vector_reduce_fadd: |
| 476 | return TTI::ReductionShuffle::Pairwise; |
| 477 | } |
| 478 | return TTI::ReductionShuffle::SplitHalf; |
| 479 | } |
| 480 | |
| 481 | void WebAssemblyTTIImpl::( |
| 482 | Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP, |
| 483 | OptimizationRemarkEmitter *ORE) const { |
| 484 | // Scan the loop: don't unroll loops with calls. This is a standard approach |
| 485 | // for most (all?) targets. |
| 486 | for (BasicBlock *BB : L->blocks()) |
| 487 | for (Instruction &I : *BB) |
| 488 | if (isa<CallInst>(Val: I) || isa<InvokeInst>(Val: I)) |
| 489 | if (const Function *F = cast<CallBase>(Val&: I).getCalledFunction()) |
| 490 | if (isLoweredToCall(F)) |
| 491 | return; |
| 492 | |
| 493 | // The chosen threshold is within the range of 'LoopMicroOpBufferSize' of |
| 494 | // the various microarchitectures that use the BasicTTI implementation and |
| 495 | // has been selected through heuristics across multiple cores and runtimes. |
| 496 | UP.Partial = UP.Runtime = UP.UpperBound = true; |
| 497 | UP.PartialThreshold = 30; |
| 498 | |
| 499 | // Avoid unrolling when optimizing for size. |
| 500 | UP.OptSizeThreshold = 0; |
| 501 | UP.PartialOptSizeThreshold = 0; |
| 502 | |
| 503 | // Set number of instructions optimized when "back edge" |
| 504 | // becomes "fall through" to default value of 2. |
| 505 | UP.BEInsns = 2; |
| 506 | } |
| 507 | |
| 508 | bool WebAssemblyTTIImpl::supportsTailCalls() const { |
| 509 | return getST()->hasTailCall(); |
| 510 | } |
| 511 | |
| 512 | bool WebAssemblyTTIImpl::isProfitableToSinkOperands( |
| 513 | Instruction *I, SmallVectorImpl<Use *> &Ops) const { |
| 514 | using namespace llvm::PatternMatch; |
| 515 | |
| 516 | if (!I->getType()->isVectorTy() || !I->isShift()) |
| 517 | return false; |
| 518 | |
| 519 | Value *V = I->getOperand(i: 1); |
| 520 | // We dont need to sink constant splat. |
| 521 | if (isa<Constant>(Val: V)) |
| 522 | return false; |
| 523 | |
| 524 | if (match(V, P: m_Shuffle(v1: m_InsertElt(Val: m_Value(), Elt: m_Value(), Idx: m_ZeroInt()), |
| 525 | v2: m_Value(), mask: m_ZeroMask()))) { |
| 526 | // Sink insert |
| 527 | Ops.push_back(Elt: &cast<Instruction>(Val: V)->getOperandUse(i: 0)); |
| 528 | // Sink shuffle |
| 529 | Ops.push_back(Elt: &I->getOperandUse(i: 1)); |
| 530 | return true; |
| 531 | } |
| 532 | |
| 533 | return false; |
| 534 | } |
| 535 | |
| 536 | /// Attempt to convert [relaxed_]swizzle to shufflevector if the mask is |
| 537 | /// constant. |
| 538 | static Value *simplifyWasmSwizzle(const IntrinsicInst &II, |
| 539 | InstCombiner::BuilderTy &Builder, |
| 540 | bool IsRelaxed) { |
| 541 | auto *V = dyn_cast<Constant>(Val: II.getArgOperand(i: 1)); |
| 542 | if (!V) |
| 543 | return nullptr; |
| 544 | |
| 545 | auto *VecTy = cast<FixedVectorType>(Val: II.getType()); |
| 546 | unsigned NumElts = VecTy->getNumElements(); |
| 547 | assert(NumElts == 16); |
| 548 | |
| 549 | // Construct a shuffle mask from constant integers or UNDEFs. |
| 550 | int Indexes[16]; |
| 551 | bool AnyOutOfBounds = false; |
| 552 | |
| 553 | for (unsigned I = 0; I < NumElts; ++I) { |
| 554 | Constant *COp = V->getAggregateElement(Elt: I); |
| 555 | if (!COp || (!isa<UndefValue>(Val: COp) && !isa<ConstantInt>(Val: COp))) |
| 556 | return nullptr; |
| 557 | |
| 558 | if (isa<UndefValue>(Val: COp)) { |
| 559 | Indexes[I] = -1; |
| 560 | continue; |
| 561 | } |
| 562 | |
| 563 | if (IsRelaxed && cast<ConstantInt>(Val: COp)->getSExtValue() >= NumElts) { |
| 564 | // The relaxed_swizzle operation always returns 0 if the lane index is |
| 565 | // less than 0 when interpreted as a signed value. For lane indices above |
| 566 | // 15, however, it can choose between returning 0 or the lane at `Index % |
| 567 | // 16`. However, the choice must be made consistently. As the WebAssembly |
| 568 | // spec states: |
| 569 | // |
| 570 | // "The result of relaxed operators are implementation-dependent, because |
| 571 | // the set of possible results may depend on properties of the host |
| 572 | // environment, such as its hardware. Technically, their behaviour is |
| 573 | // controlled by a set of global parameters to the semantics that an |
| 574 | // implementation can instantiate in different ways. These choices are |
| 575 | // fixed, that is, parameters are constant during the execution of any |
| 576 | // given program." |
| 577 | // |
| 578 | // The WebAssembly runtime may choose differently from us, so we can't |
| 579 | // optimize a relaxed swizzle with lane indices above 15. |
| 580 | return nullptr; |
| 581 | } |
| 582 | |
| 583 | uint64_t Index = cast<ConstantInt>(Val: COp)->getZExtValue(); |
| 584 | if (Index >= NumElts) { |
| 585 | AnyOutOfBounds = true; |
| 586 | // If there are out-of-bounds indices, the swizzle instruction returns |
| 587 | // zeroes in those lanes. We'll provide an all-zeroes vector as the |
| 588 | // second argument to shufflevector and read the first element from it. |
| 589 | Indexes[I] = NumElts; |
| 590 | continue; |
| 591 | } |
| 592 | |
| 593 | Indexes[I] = Index; |
| 594 | } |
| 595 | |
| 596 | auto *V1 = II.getArgOperand(i: 0); |
| 597 | auto *V2 = |
| 598 | AnyOutOfBounds ? Constant::getNullValue(Ty: VecTy) : PoisonValue::get(T: VecTy); |
| 599 | |
| 600 | return Builder.CreateShuffleVector(V1, V2, Mask: ArrayRef(Indexes, NumElts)); |
| 601 | } |
| 602 | |
| 603 | std::optional<Instruction *> |
| 604 | WebAssemblyTTIImpl::instCombineIntrinsic(InstCombiner &IC, |
| 605 | IntrinsicInst &II) const { |
| 606 | Intrinsic::ID IID = II.getIntrinsicID(); |
| 607 | switch (IID) { |
| 608 | case Intrinsic::wasm_swizzle: |
| 609 | case Intrinsic::wasm_relaxed_swizzle: |
| 610 | if (Value *V = simplifyWasmSwizzle( |
| 611 | II, Builder&: IC.Builder, IsRelaxed: IID == Intrinsic::wasm_relaxed_swizzle)) { |
| 612 | return IC.replaceInstUsesWith(I&: II, V); |
| 613 | } |
| 614 | break; |
| 615 | } |
| 616 | |
| 617 | return std::nullopt; |
| 618 | } |
| 619 | |