1 //===- ARMTargetTransformInfo.cpp - ARM specific TTI ----------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #include "ARMTargetTransformInfo.h" 11 #include "ARMSubtarget.h" 12 #include "MCTargetDesc/ARMAddressingModes.h" 13 #include "llvm/ADT/APInt.h" 14 #include "llvm/ADT/SmallVector.h" 15 #include "llvm/Analysis/LoopInfo.h" 16 #include "llvm/CodeGen/CostTable.h" 17 #include "llvm/CodeGen/ISDOpcodes.h" 18 #include "llvm/CodeGen/ValueTypes.h" 19 #include "llvm/IR/BasicBlock.h" 20 #include "llvm/IR/CallSite.h" 21 #include "llvm/IR/DataLayout.h" 22 #include "llvm/IR/DerivedTypes.h" 23 #include "llvm/IR/Instruction.h" 24 #include "llvm/IR/Instructions.h" 25 #include "llvm/IR/Type.h" 26 #include "llvm/MC/SubtargetFeature.h" 27 #include "llvm/Support/Casting.h" 28 #include "llvm/Support/MachineValueType.h" 29 #include "llvm/Target/TargetMachine.h" 30 #include <algorithm> 31 #include <cassert> 32 #include <cstdint> 33 #include <utility> 34 35 using namespace llvm; 36 37 #define DEBUG_TYPE "armtti" 38 39 bool ARMTTIImpl::areInlineCompatible(const Function *Caller, 40 const Function *Callee) const { 41 const TargetMachine &TM = getTLI()->getTargetMachine(); 42 const FeatureBitset &CallerBits = 43 TM.getSubtargetImpl(*Caller)->getFeatureBits(); 44 const FeatureBitset &CalleeBits = 45 TM.getSubtargetImpl(*Callee)->getFeatureBits(); 46 47 // To inline a callee, all features not in the whitelist must match exactly. 48 bool MatchExact = (CallerBits & ~InlineFeatureWhitelist) == 49 (CalleeBits & ~InlineFeatureWhitelist); 50 // For features in the whitelist, the callee's features must be a subset of 51 // the callers'. 52 bool MatchSubset = ((CallerBits & CalleeBits) & InlineFeatureWhitelist) == 53 (CalleeBits & InlineFeatureWhitelist); 54 return MatchExact && MatchSubset; 55 } 56 57 int ARMTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) { 58 assert(Ty->isIntegerTy()); 59 60 unsigned Bits = Ty->getPrimitiveSizeInBits(); 61 if (Bits == 0 || Imm.getActiveBits() >= 64) 62 return 4; 63 64 int64_t SImmVal = Imm.getSExtValue(); 65 uint64_t ZImmVal = Imm.getZExtValue(); 66 if (!ST->isThumb()) { 67 if ((SImmVal >= 0 && SImmVal < 65536) || 68 (ARM_AM::getSOImmVal(ZImmVal) != -1) || 69 (ARM_AM::getSOImmVal(~ZImmVal) != -1)) 70 return 1; 71 return ST->hasV6T2Ops() ? 2 : 3; 72 } 73 if (ST->isThumb2()) { 74 if ((SImmVal >= 0 && SImmVal < 65536) || 75 (ARM_AM::getT2SOImmVal(ZImmVal) != -1) || 76 (ARM_AM::getT2SOImmVal(~ZImmVal) != -1)) 77 return 1; 78 return ST->hasV6T2Ops() ? 2 : 3; 79 } 80 // Thumb1, any i8 imm cost 1. 81 if (Bits == 8 || (SImmVal >= 0 && SImmVal < 256)) 82 return 1; 83 if ((~SImmVal < 256) || ARM_AM::isThumbImmShiftedVal(ZImmVal)) 84 return 2; 85 // Load from constantpool. 86 return 3; 87 } 88 89 // Constants smaller than 256 fit in the immediate field of 90 // Thumb1 instructions so we return a zero cost and 1 otherwise. 91 int ARMTTIImpl::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 92 const APInt &Imm, Type *Ty) { 93 if (Imm.isNonNegative() && Imm.getLimitedValue() < 256) 94 return 0; 95 96 return 1; 97 } 98 99 int ARMTTIImpl::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm, 100 Type *Ty) { 101 // Division by a constant can be turned into multiplication, but only if we 102 // know it's constant. So it's not so much that the immediate is cheap (it's 103 // not), but that the alternative is worse. 104 // FIXME: this is probably unneeded with GlobalISel. 105 if ((Opcode == Instruction::SDiv || Opcode == Instruction::UDiv || 106 Opcode == Instruction::SRem || Opcode == Instruction::URem) && 107 Idx == 1) 108 return 0; 109 110 if (Opcode == Instruction::And) 111 // Conversion to BIC is free, and means we can use ~Imm instead. 112 return std::min(getIntImmCost(Imm, Ty), getIntImmCost(~Imm, Ty)); 113 114 if (Opcode == Instruction::Add) 115 // Conversion to SUB is free, and means we can use -Imm instead. 116 return std::min(getIntImmCost(Imm, Ty), getIntImmCost(-Imm, Ty)); 117 118 if (Opcode == Instruction::ICmp && Imm.isNegative() && 119 Ty->getIntegerBitWidth() == 32) { 120 int64_t NegImm = -Imm.getSExtValue(); 121 if (ST->isThumb2() && NegImm < 1<<12) 122 // icmp X, #-C -> cmn X, #C 123 return 0; 124 if (ST->isThumb() && NegImm < 1<<8) 125 // icmp X, #-C -> adds X, #C 126 return 0; 127 } 128 129 // xor a, -1 can always be folded to MVN 130 if (Opcode == Instruction::Xor && Imm.isAllOnesValue()) 131 return 0; 132 133 return getIntImmCost(Imm, Ty); 134 } 135 136 int ARMTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, 137 const Instruction *I) { 138 int ISD = TLI->InstructionOpcodeToISD(Opcode); 139 assert(ISD && "Invalid opcode"); 140 141 // Single to/from double precision conversions. 142 static const CostTblEntry NEONFltDblTbl[] = { 143 // Vector fptrunc/fpext conversions. 144 { ISD::FP_ROUND, MVT::v2f64, 2 }, 145 { ISD::FP_EXTEND, MVT::v2f32, 2 }, 146 { ISD::FP_EXTEND, MVT::v4f32, 4 } 147 }; 148 149 if (Src->isVectorTy() && ST->hasNEON() && (ISD == ISD::FP_ROUND || 150 ISD == ISD::FP_EXTEND)) { 151 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src); 152 if (const auto *Entry = CostTableLookup(NEONFltDblTbl, ISD, LT.second)) 153 return LT.first * Entry->Cost; 154 } 155 156 EVT SrcTy = TLI->getValueType(DL, Src); 157 EVT DstTy = TLI->getValueType(DL, Dst); 158 159 if (!SrcTy.isSimple() || !DstTy.isSimple()) 160 return BaseT::getCastInstrCost(Opcode, Dst, Src); 161 162 // Some arithmetic, load and store operations have specific instructions 163 // to cast up/down their types automatically at no extra cost. 164 // TODO: Get these tables to know at least what the related operations are. 165 static const TypeConversionCostTblEntry NEONVectorConversionTbl[] = { 166 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 0 }, 167 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 0 }, 168 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i32, 1 }, 169 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i32, 1 }, 170 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 0 }, 171 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 1 }, 172 173 // The number of vmovl instructions for the extension. 174 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3 }, 175 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 }, 176 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3 }, 177 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3 }, 178 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 7 }, 179 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 7 }, 180 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 6 }, 181 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 6 }, 182 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 6 }, 183 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 6 }, 184 185 // Operations that we legalize using splitting. 186 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 6 }, 187 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 3 }, 188 189 // Vector float <-> i32 conversions. 190 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 191 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 }, 192 193 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i8, 3 }, 194 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i8, 3 }, 195 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i16, 2 }, 196 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i16, 2 }, 197 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 }, 198 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 }, 199 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 }, 200 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 }, 201 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 }, 202 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 }, 203 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 }, 204 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 }, 205 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 }, 206 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 }, 207 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 }, 208 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 }, 209 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 8 }, 210 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i16, 8 }, 211 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 4 }, 212 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i32, 4 }, 213 214 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 1 }, 215 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 }, 216 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 3 }, 217 { ISD::FP_TO_UINT, MVT::v4i8, MVT::v4f32, 3 }, 218 { ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f32, 2 }, 219 { ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f32, 2 }, 220 221 // Vector double <-> i32 conversions. 222 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 }, 223 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 }, 224 225 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i8, 4 }, 226 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i8, 4 }, 227 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i16, 3 }, 228 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i16, 3 }, 229 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 }, 230 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 2 }, 231 232 { ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f64, 2 }, 233 { ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f64, 2 }, 234 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f32, 4 }, 235 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f32, 4 }, 236 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v16f32, 8 }, 237 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v16f32, 8 } 238 }; 239 240 if (SrcTy.isVector() && ST->hasNEON()) { 241 if (const auto *Entry = ConvertCostTableLookup(NEONVectorConversionTbl, ISD, 242 DstTy.getSimpleVT(), 243 SrcTy.getSimpleVT())) 244 return Entry->Cost; 245 } 246 247 // Scalar float to integer conversions. 248 static const TypeConversionCostTblEntry NEONFloatConversionTbl[] = { 249 { ISD::FP_TO_SINT, MVT::i1, MVT::f32, 2 }, 250 { ISD::FP_TO_UINT, MVT::i1, MVT::f32, 2 }, 251 { ISD::FP_TO_SINT, MVT::i1, MVT::f64, 2 }, 252 { ISD::FP_TO_UINT, MVT::i1, MVT::f64, 2 }, 253 { ISD::FP_TO_SINT, MVT::i8, MVT::f32, 2 }, 254 { ISD::FP_TO_UINT, MVT::i8, MVT::f32, 2 }, 255 { ISD::FP_TO_SINT, MVT::i8, MVT::f64, 2 }, 256 { ISD::FP_TO_UINT, MVT::i8, MVT::f64, 2 }, 257 { ISD::FP_TO_SINT, MVT::i16, MVT::f32, 2 }, 258 { ISD::FP_TO_UINT, MVT::i16, MVT::f32, 2 }, 259 { ISD::FP_TO_SINT, MVT::i16, MVT::f64, 2 }, 260 { ISD::FP_TO_UINT, MVT::i16, MVT::f64, 2 }, 261 { ISD::FP_TO_SINT, MVT::i32, MVT::f32, 2 }, 262 { ISD::FP_TO_UINT, MVT::i32, MVT::f32, 2 }, 263 { ISD::FP_TO_SINT, MVT::i32, MVT::f64, 2 }, 264 { ISD::FP_TO_UINT, MVT::i32, MVT::f64, 2 }, 265 { ISD::FP_TO_SINT, MVT::i64, MVT::f32, 10 }, 266 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 10 }, 267 { ISD::FP_TO_SINT, MVT::i64, MVT::f64, 10 }, 268 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 10 } 269 }; 270 if (SrcTy.isFloatingPoint() && ST->hasNEON()) { 271 if (const auto *Entry = ConvertCostTableLookup(NEONFloatConversionTbl, ISD, 272 DstTy.getSimpleVT(), 273 SrcTy.getSimpleVT())) 274 return Entry->Cost; 275 } 276 277 // Scalar integer to float conversions. 278 static const TypeConversionCostTblEntry NEONIntegerConversionTbl[] = { 279 { ISD::SINT_TO_FP, MVT::f32, MVT::i1, 2 }, 280 { ISD::UINT_TO_FP, MVT::f32, MVT::i1, 2 }, 281 { ISD::SINT_TO_FP, MVT::f64, MVT::i1, 2 }, 282 { ISD::UINT_TO_FP, MVT::f64, MVT::i1, 2 }, 283 { ISD::SINT_TO_FP, MVT::f32, MVT::i8, 2 }, 284 { ISD::UINT_TO_FP, MVT::f32, MVT::i8, 2 }, 285 { ISD::SINT_TO_FP, MVT::f64, MVT::i8, 2 }, 286 { ISD::UINT_TO_FP, MVT::f64, MVT::i8, 2 }, 287 { ISD::SINT_TO_FP, MVT::f32, MVT::i16, 2 }, 288 { ISD::UINT_TO_FP, MVT::f32, MVT::i16, 2 }, 289 { ISD::SINT_TO_FP, MVT::f64, MVT::i16, 2 }, 290 { ISD::UINT_TO_FP, MVT::f64, MVT::i16, 2 }, 291 { ISD::SINT_TO_FP, MVT::f32, MVT::i32, 2 }, 292 { ISD::UINT_TO_FP, MVT::f32, MVT::i32, 2 }, 293 { ISD::SINT_TO_FP, MVT::f64, MVT::i32, 2 }, 294 { ISD::UINT_TO_FP, MVT::f64, MVT::i32, 2 }, 295 { ISD::SINT_TO_FP, MVT::f32, MVT::i64, 10 }, 296 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 10 }, 297 { ISD::SINT_TO_FP, MVT::f64, MVT::i64, 10 }, 298 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 10 } 299 }; 300 301 if (SrcTy.isInteger() && ST->hasNEON()) { 302 if (const auto *Entry = ConvertCostTableLookup(NEONIntegerConversionTbl, 303 ISD, DstTy.getSimpleVT(), 304 SrcTy.getSimpleVT())) 305 return Entry->Cost; 306 } 307 308 // Scalar integer conversion costs. 309 static const TypeConversionCostTblEntry ARMIntegerConversionTbl[] = { 310 // i16 -> i64 requires two dependent operations. 311 { ISD::SIGN_EXTEND, MVT::i64, MVT::i16, 2 }, 312 313 // Truncates on i64 are assumed to be free. 314 { ISD::TRUNCATE, MVT::i32, MVT::i64, 0 }, 315 { ISD::TRUNCATE, MVT::i16, MVT::i64, 0 }, 316 { ISD::TRUNCATE, MVT::i8, MVT::i64, 0 }, 317 { ISD::TRUNCATE, MVT::i1, MVT::i64, 0 } 318 }; 319 320 if (SrcTy.isInteger()) { 321 if (const auto *Entry = ConvertCostTableLookup(ARMIntegerConversionTbl, ISD, 322 DstTy.getSimpleVT(), 323 SrcTy.getSimpleVT())) 324 return Entry->Cost; 325 } 326 327 return BaseT::getCastInstrCost(Opcode, Dst, Src); 328 } 329 330 int ARMTTIImpl::getVectorInstrCost(unsigned Opcode, Type *ValTy, 331 unsigned Index) { 332 // Penalize inserting into an D-subregister. We end up with a three times 333 // lower estimated throughput on swift. 334 if (ST->hasSlowLoadDSubregister() && Opcode == Instruction::InsertElement && 335 ValTy->isVectorTy() && ValTy->getScalarSizeInBits() <= 32) 336 return 3; 337 338 if ((Opcode == Instruction::InsertElement || 339 Opcode == Instruction::ExtractElement)) { 340 // Cross-class copies are expensive on many microarchitectures, 341 // so assume they are expensive by default. 342 if (ValTy->getVectorElementType()->isIntegerTy()) 343 return 3; 344 345 // Even if it's not a cross class copy, this likely leads to mixing 346 // of NEON and VFP code and should be therefore penalized. 347 if (ValTy->isVectorTy() && 348 ValTy->getScalarSizeInBits() <= 32) 349 return std::max(BaseT::getVectorInstrCost(Opcode, ValTy, Index), 2U); 350 } 351 352 return BaseT::getVectorInstrCost(Opcode, ValTy, Index); 353 } 354 355 int ARMTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, 356 const Instruction *I) { 357 int ISD = TLI->InstructionOpcodeToISD(Opcode); 358 // On NEON a vector select gets lowered to vbsl. 359 if (ST->hasNEON() && ValTy->isVectorTy() && ISD == ISD::SELECT) { 360 // Lowering of some vector selects is currently far from perfect. 361 static const TypeConversionCostTblEntry NEONVectorSelectTbl[] = { 362 { ISD::SELECT, MVT::v4i1, MVT::v4i64, 4*4 + 1*2 + 1 }, 363 { ISD::SELECT, MVT::v8i1, MVT::v8i64, 50 }, 364 { ISD::SELECT, MVT::v16i1, MVT::v16i64, 100 } 365 }; 366 367 EVT SelCondTy = TLI->getValueType(DL, CondTy); 368 EVT SelValTy = TLI->getValueType(DL, ValTy); 369 if (SelCondTy.isSimple() && SelValTy.isSimple()) { 370 if (const auto *Entry = ConvertCostTableLookup(NEONVectorSelectTbl, ISD, 371 SelCondTy.getSimpleVT(), 372 SelValTy.getSimpleVT())) 373 return Entry->Cost; 374 } 375 376 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy); 377 return LT.first; 378 } 379 380 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 381 } 382 383 int ARMTTIImpl::getAddressComputationCost(Type *Ty, ScalarEvolution *SE, 384 const SCEV *Ptr) { 385 // Address computations in vectorized code with non-consecutive addresses will 386 // likely result in more instructions compared to scalar code where the 387 // computation can more often be merged into the index mode. The resulting 388 // extra micro-ops can significantly decrease throughput. 389 unsigned NumVectorInstToHideOverhead = 10; 390 int MaxMergeDistance = 64; 391 392 if (Ty->isVectorTy() && SE && 393 !BaseT::isConstantStridedAccessLessThan(SE, Ptr, MaxMergeDistance + 1)) 394 return NumVectorInstToHideOverhead; 395 396 // In many cases the address computation is not merged into the instruction 397 // addressing mode. 398 return 1; 399 } 400 401 int ARMTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index, 402 Type *SubTp) { 403 if (Kind == TTI::SK_Broadcast) { 404 static const CostTblEntry NEONDupTbl[] = { 405 // VDUP handles these cases. 406 {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1}, 407 {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1}, 408 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, 409 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, 410 {ISD::VECTOR_SHUFFLE, MVT::v4i16, 1}, 411 {ISD::VECTOR_SHUFFLE, MVT::v8i8, 1}, 412 413 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1}, 414 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1}, 415 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1}, 416 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 1}}; 417 418 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp); 419 420 if (const auto *Entry = CostTableLookup(NEONDupTbl, ISD::VECTOR_SHUFFLE, 421 LT.second)) 422 return LT.first * Entry->Cost; 423 424 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp); 425 } 426 if (Kind == TTI::SK_Reverse) { 427 static const CostTblEntry NEONShuffleTbl[] = { 428 // Reverse shuffle cost one instruction if we are shuffling within a 429 // double word (vrev) or two if we shuffle a quad word (vrev, vext). 430 {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1}, 431 {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1}, 432 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, 433 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, 434 {ISD::VECTOR_SHUFFLE, MVT::v4i16, 1}, 435 {ISD::VECTOR_SHUFFLE, MVT::v8i8, 1}, 436 437 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2}, 438 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2}, 439 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 2}, 440 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 2}}; 441 442 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp); 443 444 if (const auto *Entry = CostTableLookup(NEONShuffleTbl, ISD::VECTOR_SHUFFLE, 445 LT.second)) 446 return LT.first * Entry->Cost; 447 448 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp); 449 } 450 if (Kind == TTI::SK_Select) { 451 static const CostTblEntry NEONSelShuffleTbl[] = { 452 // Select shuffle cost table for ARM. Cost is the number of instructions 453 // required to create the shuffled vector. 454 455 {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1}, 456 {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1}, 457 {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1}, 458 {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1}, 459 460 {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2}, 461 {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2}, 462 {ISD::VECTOR_SHUFFLE, MVT::v4i16, 2}, 463 464 {ISD::VECTOR_SHUFFLE, MVT::v8i16, 16}, 465 466 {ISD::VECTOR_SHUFFLE, MVT::v16i8, 32}}; 467 468 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp); 469 if (const auto *Entry = CostTableLookup(NEONSelShuffleTbl, 470 ISD::VECTOR_SHUFFLE, LT.second)) 471 return LT.first * Entry->Cost; 472 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp); 473 } 474 return BaseT::getShuffleCost(Kind, Tp, Index, SubTp); 475 } 476 477 int ARMTTIImpl::getArithmeticInstrCost( 478 unsigned Opcode, Type *Ty, TTI::OperandValueKind Op1Info, 479 TTI::OperandValueKind Op2Info, TTI::OperandValueProperties Opd1PropInfo, 480 TTI::OperandValueProperties Opd2PropInfo, 481 ArrayRef<const Value *> Args) { 482 int ISDOpcode = TLI->InstructionOpcodeToISD(Opcode); 483 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty); 484 485 const unsigned FunctionCallDivCost = 20; 486 const unsigned ReciprocalDivCost = 10; 487 static const CostTblEntry CostTbl[] = { 488 // Division. 489 // These costs are somewhat random. Choose a cost of 20 to indicate that 490 // vectorizing devision (added function call) is going to be very expensive. 491 // Double registers types. 492 { ISD::SDIV, MVT::v1i64, 1 * FunctionCallDivCost}, 493 { ISD::UDIV, MVT::v1i64, 1 * FunctionCallDivCost}, 494 { ISD::SREM, MVT::v1i64, 1 * FunctionCallDivCost}, 495 { ISD::UREM, MVT::v1i64, 1 * FunctionCallDivCost}, 496 { ISD::SDIV, MVT::v2i32, 2 * FunctionCallDivCost}, 497 { ISD::UDIV, MVT::v2i32, 2 * FunctionCallDivCost}, 498 { ISD::SREM, MVT::v2i32, 2 * FunctionCallDivCost}, 499 { ISD::UREM, MVT::v2i32, 2 * FunctionCallDivCost}, 500 { ISD::SDIV, MVT::v4i16, ReciprocalDivCost}, 501 { ISD::UDIV, MVT::v4i16, ReciprocalDivCost}, 502 { ISD::SREM, MVT::v4i16, 4 * FunctionCallDivCost}, 503 { ISD::UREM, MVT::v4i16, 4 * FunctionCallDivCost}, 504 { ISD::SDIV, MVT::v8i8, ReciprocalDivCost}, 505 { ISD::UDIV, MVT::v8i8, ReciprocalDivCost}, 506 { ISD::SREM, MVT::v8i8, 8 * FunctionCallDivCost}, 507 { ISD::UREM, MVT::v8i8, 8 * FunctionCallDivCost}, 508 // Quad register types. 509 { ISD::SDIV, MVT::v2i64, 2 * FunctionCallDivCost}, 510 { ISD::UDIV, MVT::v2i64, 2 * FunctionCallDivCost}, 511 { ISD::SREM, MVT::v2i64, 2 * FunctionCallDivCost}, 512 { ISD::UREM, MVT::v2i64, 2 * FunctionCallDivCost}, 513 { ISD::SDIV, MVT::v4i32, 4 * FunctionCallDivCost}, 514 { ISD::UDIV, MVT::v4i32, 4 * FunctionCallDivCost}, 515 { ISD::SREM, MVT::v4i32, 4 * FunctionCallDivCost}, 516 { ISD::UREM, MVT::v4i32, 4 * FunctionCallDivCost}, 517 { ISD::SDIV, MVT::v8i16, 8 * FunctionCallDivCost}, 518 { ISD::UDIV, MVT::v8i16, 8 * FunctionCallDivCost}, 519 { ISD::SREM, MVT::v8i16, 8 * FunctionCallDivCost}, 520 { ISD::UREM, MVT::v8i16, 8 * FunctionCallDivCost}, 521 { ISD::SDIV, MVT::v16i8, 16 * FunctionCallDivCost}, 522 { ISD::UDIV, MVT::v16i8, 16 * FunctionCallDivCost}, 523 { ISD::SREM, MVT::v16i8, 16 * FunctionCallDivCost}, 524 { ISD::UREM, MVT::v16i8, 16 * FunctionCallDivCost}, 525 // Multiplication. 526 }; 527 528 if (ST->hasNEON()) 529 if (const auto *Entry = CostTableLookup(CostTbl, ISDOpcode, LT.second)) 530 return LT.first * Entry->Cost; 531 532 int Cost = BaseT::getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info, 533 Opd1PropInfo, Opd2PropInfo); 534 535 // This is somewhat of a hack. The problem that we are facing is that SROA 536 // creates a sequence of shift, and, or instructions to construct values. 537 // These sequences are recognized by the ISel and have zero-cost. Not so for 538 // the vectorized code. Because we have support for v2i64 but not i64 those 539 // sequences look particularly beneficial to vectorize. 540 // To work around this we increase the cost of v2i64 operations to make them 541 // seem less beneficial. 542 if (LT.second == MVT::v2i64 && 543 Op2Info == TargetTransformInfo::OK_UniformConstantValue) 544 Cost += 4; 545 546 return Cost; 547 } 548 549 int ARMTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, 550 unsigned AddressSpace, const Instruction *I) { 551 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src); 552 553 if (Src->isVectorTy() && Alignment != 16 && 554 Src->getVectorElementType()->isDoubleTy()) { 555 // Unaligned loads/stores are extremely inefficient. 556 // We need 4 uops for vst.1/vld.1 vs 1uop for vldr/vstr. 557 return LT.first * 4; 558 } 559 return LT.first; 560 } 561 562 int ARMTTIImpl::getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, 563 unsigned Factor, 564 ArrayRef<unsigned> Indices, 565 unsigned Alignment, 566 unsigned AddressSpace, 567 bool UseMaskForCond, 568 bool UseMaskForGaps) { 569 assert(Factor >= 2 && "Invalid interleave factor"); 570 assert(isa<VectorType>(VecTy) && "Expect a vector type"); 571 572 // vldN/vstN doesn't support vector types of i64/f64 element. 573 bool EltIs64Bits = DL.getTypeSizeInBits(VecTy->getScalarType()) == 64; 574 575 if (Factor <= TLI->getMaxSupportedInterleaveFactor() && !EltIs64Bits && 576 !UseMaskForCond && !UseMaskForGaps) { 577 unsigned NumElts = VecTy->getVectorNumElements(); 578 auto *SubVecTy = VectorType::get(VecTy->getScalarType(), NumElts / Factor); 579 580 // vldN/vstN only support legal vector types of size 64 or 128 in bits. 581 // Accesses having vector types that are a multiple of 128 bits can be 582 // matched to more than one vldN/vstN instruction. 583 if (NumElts % Factor == 0 && 584 TLI->isLegalInterleavedAccessType(SubVecTy, DL)) 585 return Factor * TLI->getNumInterleavedAccesses(SubVecTy, DL); 586 } 587 588 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 589 Alignment, AddressSpace, 590 UseMaskForCond, UseMaskForGaps); 591 } 592 593 void ARMTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE, 594 TTI::UnrollingPreferences &UP) { 595 // Only currently enable these preferences for M-Class cores. 596 if (!ST->isMClass()) 597 return BasicTTIImplBase::getUnrollingPreferences(L, SE, UP); 598 599 // Disable loop unrolling for Oz and Os. 600 UP.OptSizeThreshold = 0; 601 UP.PartialOptSizeThreshold = 0; 602 if (L->getHeader()->getParent()->optForSize()) 603 return; 604 605 // Only enable on Thumb-2 targets. 606 if (!ST->isThumb2()) 607 return; 608 609 SmallVector<BasicBlock*, 4> ExitingBlocks; 610 L->getExitingBlocks(ExitingBlocks); 611 LLVM_DEBUG(dbgs() << "Loop has:\n" 612 << "Blocks: " << L->getNumBlocks() << "\n" 613 << "Exit blocks: " << ExitingBlocks.size() << "\n"); 614 615 // Only allow another exit other than the latch. This acts as an early exit 616 // as it mirrors the profitability calculation of the runtime unroller. 617 if (ExitingBlocks.size() > 2) 618 return; 619 620 // Limit the CFG of the loop body for targets with a branch predictor. 621 // Allowing 4 blocks permits if-then-else diamonds in the body. 622 if (ST->hasBranchPredictor() && L->getNumBlocks() > 4) 623 return; 624 625 // Scan the loop: don't unroll loops with calls as this could prevent 626 // inlining. 627 unsigned Cost = 0; 628 for (auto *BB : L->getBlocks()) { 629 for (auto &I : *BB) { 630 if (isa<CallInst>(I) || isa<InvokeInst>(I)) { 631 ImmutableCallSite CS(&I); 632 if (const Function *F = CS.getCalledFunction()) { 633 if (!isLoweredToCall(F)) 634 continue; 635 } 636 return; 637 } 638 SmallVector<const Value*, 4> Operands(I.value_op_begin(), 639 I.value_op_end()); 640 Cost += getUserCost(&I, Operands); 641 } 642 } 643 644 LLVM_DEBUG(dbgs() << "Cost of loop: " << Cost << "\n"); 645 646 UP.Partial = true; 647 UP.Runtime = true; 648 UP.UnrollRemainder = true; 649 UP.DefaultUnrollRuntimeCount = 4; 650 UP.UnrollAndJam = true; 651 UP.UnrollAndJamInnerLoopThreshold = 60; 652 653 // Force unrolling small loops can be very useful because of the branch 654 // taken cost of the backedge. 655 if (Cost < 12) 656 UP.Force = true; 657 } 658