1 //===-- X86FixupVectorConstants.cpp - optimize constant generation -------===// 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 // This file examines all full size vector constant pool loads and attempts to 10 // replace them with smaller constant pool entries, including: 11 // * Converting AVX512 memory-fold instructions to their broadcast-fold form 12 // * Broadcasting of full width loads. 13 // * TODO: Sign/Zero extension of full width loads. 14 // 15 //===----------------------------------------------------------------------===// 16 17 #include "X86.h" 18 #include "X86InstrFoldTables.h" 19 #include "X86InstrInfo.h" 20 #include "X86Subtarget.h" 21 #include "llvm/ADT/Statistic.h" 22 #include "llvm/CodeGen/MachineConstantPool.h" 23 24 using namespace llvm; 25 26 #define DEBUG_TYPE "x86-fixup-vector-constants" 27 28 STATISTIC(NumInstChanges, "Number of instructions changes"); 29 30 namespace { 31 class X86FixupVectorConstantsPass : public MachineFunctionPass { 32 public: 33 static char ID; 34 35 X86FixupVectorConstantsPass() : MachineFunctionPass(ID) {} 36 37 StringRef getPassName() const override { 38 return "X86 Fixup Vector Constants"; 39 } 40 41 bool runOnMachineFunction(MachineFunction &MF) override; 42 bool processInstruction(MachineFunction &MF, MachineBasicBlock &MBB, 43 MachineInstr &MI); 44 45 // This pass runs after regalloc and doesn't support VReg operands. 46 MachineFunctionProperties getRequiredProperties() const override { 47 return MachineFunctionProperties().set( 48 MachineFunctionProperties::Property::NoVRegs); 49 } 50 51 private: 52 const X86InstrInfo *TII = nullptr; 53 const X86Subtarget *ST = nullptr; 54 const MCSchedModel *SM = nullptr; 55 }; 56 } // end anonymous namespace 57 58 char X86FixupVectorConstantsPass::ID = 0; 59 60 INITIALIZE_PASS(X86FixupVectorConstantsPass, DEBUG_TYPE, DEBUG_TYPE, false, false) 61 62 FunctionPass *llvm::createX86FixupVectorConstants() { 63 return new X86FixupVectorConstantsPass(); 64 } 65 66 // Attempt to extract the full width of bits data from the constant. 67 static std::optional<APInt> extractConstantBits(const Constant *C) { 68 unsigned NumBits = C->getType()->getPrimitiveSizeInBits(); 69 70 if (auto *CInt = dyn_cast<ConstantInt>(C)) 71 return CInt->getValue(); 72 73 if (auto *CFP = dyn_cast<ConstantFP>(C)) 74 return CFP->getValue().bitcastToAPInt(); 75 76 if (auto *CV = dyn_cast<ConstantVector>(C)) { 77 if (auto *CVSplat = CV->getSplatValue(/*AllowUndefs*/ true)) { 78 if (std::optional<APInt> Bits = extractConstantBits(CVSplat)) { 79 assert((NumBits % Bits->getBitWidth()) == 0 && "Illegal splat"); 80 return APInt::getSplat(NumBits, *Bits); 81 } 82 } 83 } 84 85 if (auto *CDS = dyn_cast<ConstantDataSequential>(C)) { 86 bool IsInteger = CDS->getElementType()->isIntegerTy(); 87 bool IsFloat = CDS->getElementType()->isHalfTy() || 88 CDS->getElementType()->isBFloatTy() || 89 CDS->getElementType()->isFloatTy() || 90 CDS->getElementType()->isDoubleTy(); 91 if (IsInteger || IsFloat) { 92 APInt Bits = APInt::getZero(NumBits); 93 unsigned EltBits = CDS->getElementType()->getPrimitiveSizeInBits(); 94 for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) { 95 if (IsInteger) 96 Bits.insertBits(CDS->getElementAsAPInt(I), I * EltBits); 97 else 98 Bits.insertBits(CDS->getElementAsAPFloat(I).bitcastToAPInt(), 99 I * EltBits); 100 } 101 return Bits; 102 } 103 } 104 105 return std::nullopt; 106 } 107 108 // Attempt to compute the splat width of bits data by normalizing the splat to 109 // remove undefs. 110 static std::optional<APInt> getSplatableConstant(const Constant *C, 111 unsigned SplatBitWidth) { 112 const Type *Ty = C->getType(); 113 assert((Ty->getPrimitiveSizeInBits() % SplatBitWidth) == 0 && 114 "Illegal splat width"); 115 116 if (std::optional<APInt> Bits = extractConstantBits(C)) 117 if (Bits->isSplat(SplatBitWidth)) 118 return Bits->trunc(SplatBitWidth); 119 120 // Detect general splats with undefs. 121 // TODO: Do we need to handle NumEltsBits > SplatBitWidth splitting? 122 if (auto *CV = dyn_cast<ConstantVector>(C)) { 123 unsigned NumOps = CV->getNumOperands(); 124 unsigned NumEltsBits = Ty->getScalarSizeInBits(); 125 unsigned NumScaleOps = SplatBitWidth / NumEltsBits; 126 if ((SplatBitWidth % NumEltsBits) == 0) { 127 // Collect the elements and ensure that within the repeated splat sequence 128 // they either match or are undef. 129 SmallVector<Constant *, 16> Sequence(NumScaleOps, nullptr); 130 for (unsigned Idx = 0; Idx != NumOps; ++Idx) { 131 if (Constant *Elt = CV->getAggregateElement(Idx)) { 132 if (isa<UndefValue>(Elt)) 133 continue; 134 unsigned SplatIdx = Idx % NumScaleOps; 135 if (!Sequence[SplatIdx] || Sequence[SplatIdx] == Elt) { 136 Sequence[SplatIdx] = Elt; 137 continue; 138 } 139 } 140 return std::nullopt; 141 } 142 // Extract the constant bits forming the splat and insert into the bits 143 // data, leave undef as zero. 144 APInt SplatBits = APInt::getZero(SplatBitWidth); 145 for (unsigned I = 0; I != NumScaleOps; ++I) { 146 if (!Sequence[I]) 147 continue; 148 if (std::optional<APInt> Bits = extractConstantBits(Sequence[I])) { 149 SplatBits.insertBits(*Bits, I * Bits->getBitWidth()); 150 continue; 151 } 152 return std::nullopt; 153 } 154 return SplatBits; 155 } 156 } 157 158 return std::nullopt; 159 } 160 161 // Attempt to rebuild a normalized splat vector constant of the requested splat 162 // width, built up of potentially smaller scalar values. 163 // NOTE: We don't always bother converting to scalars if the vector length is 1. 164 static Constant *rebuildSplatableConstant(const Constant *C, 165 unsigned SplatBitWidth) { 166 std::optional<APInt> Splat = getSplatableConstant(C, SplatBitWidth); 167 if (!Splat) 168 return nullptr; 169 170 // Determine scalar size to use for the constant splat vector, clamping as we 171 // might have found a splat smaller than the original constant data. 172 const Type *OriginalType = C->getType(); 173 Type *SclTy = OriginalType->getScalarType(); 174 unsigned NumSclBits = SclTy->getPrimitiveSizeInBits(); 175 NumSclBits = std::min<unsigned>(NumSclBits, SplatBitWidth); 176 LLVMContext &Ctx = OriginalType->getContext(); 177 178 if (NumSclBits == 8) { 179 SmallVector<uint8_t> RawBits; 180 for (unsigned I = 0; I != SplatBitWidth; I += 8) 181 RawBits.push_back(Splat->extractBits(8, I).getZExtValue()); 182 return ConstantDataVector::get(Ctx, RawBits); 183 } 184 185 if (NumSclBits == 16) { 186 SmallVector<uint16_t> RawBits; 187 for (unsigned I = 0; I != SplatBitWidth; I += 16) 188 RawBits.push_back(Splat->extractBits(16, I).getZExtValue()); 189 if (SclTy->is16bitFPTy()) 190 return ConstantDataVector::getFP(SclTy, RawBits); 191 return ConstantDataVector::get(Ctx, RawBits); 192 } 193 194 if (NumSclBits == 32) { 195 SmallVector<uint32_t> RawBits; 196 for (unsigned I = 0; I != SplatBitWidth; I += 32) 197 RawBits.push_back(Splat->extractBits(32, I).getZExtValue()); 198 if (SclTy->isFloatTy()) 199 return ConstantDataVector::getFP(SclTy, RawBits); 200 return ConstantDataVector::get(Ctx, RawBits); 201 } 202 203 // Fallback to i64 / double. 204 SmallVector<uint64_t> RawBits; 205 for (unsigned I = 0; I != SplatBitWidth; I += 64) 206 RawBits.push_back(Splat->extractBits(64, I).getZExtValue()); 207 if (SclTy->isDoubleTy()) 208 return ConstantDataVector::getFP(SclTy, RawBits); 209 return ConstantDataVector::get(Ctx, RawBits); 210 } 211 212 bool X86FixupVectorConstantsPass::processInstruction(MachineFunction &MF, 213 MachineBasicBlock &MBB, 214 MachineInstr &MI) { 215 unsigned Opc = MI.getOpcode(); 216 MachineConstantPool *CP = MI.getParent()->getParent()->getConstantPool(); 217 bool HasAVX2 = ST->hasAVX2(); 218 bool HasDQI = ST->hasDQI(); 219 bool HasBWI = ST->hasBWI(); 220 bool HasVLX = ST->hasVLX(); 221 222 auto ConvertToBroadcast = [&](unsigned OpBcst256, unsigned OpBcst128, 223 unsigned OpBcst64, unsigned OpBcst32, 224 unsigned OpBcst16, unsigned OpBcst8, 225 unsigned OperandNo) { 226 assert(MI.getNumOperands() >= (OperandNo + X86::AddrNumOperands) && 227 "Unexpected number of operands!"); 228 229 MachineOperand &CstOp = MI.getOperand(OperandNo + X86::AddrDisp); 230 if (auto *C = X86::getConstantFromPool(MI, CstOp)) { 231 // Attempt to detect a suitable splat from increasing splat widths. 232 std::pair<unsigned, unsigned> Broadcasts[] = { 233 {8, OpBcst8}, {16, OpBcst16}, {32, OpBcst32}, 234 {64, OpBcst64}, {128, OpBcst128}, {256, OpBcst256}, 235 }; 236 for (auto [BitWidth, OpBcst] : Broadcasts) { 237 if (OpBcst) { 238 // Construct a suitable splat constant and adjust the MI to 239 // use the new constant pool entry. 240 if (Constant *NewCst = rebuildSplatableConstant(C, BitWidth)) { 241 unsigned NewCPI = 242 CP->getConstantPoolIndex(NewCst, Align(BitWidth / 8)); 243 MI.setDesc(TII->get(OpBcst)); 244 CstOp.setIndex(NewCPI); 245 return true; 246 } 247 } 248 } 249 } 250 return false; 251 }; 252 253 // Attempt to convert full width vector loads into broadcast loads. 254 switch (Opc) { 255 /* FP Loads */ 256 case X86::MOVAPDrm: 257 case X86::MOVAPSrm: 258 case X86::MOVUPDrm: 259 case X86::MOVUPSrm: 260 // TODO: SSE3 MOVDDUP Handling 261 return false; 262 case X86::VMOVAPDrm: 263 case X86::VMOVAPSrm: 264 case X86::VMOVUPDrm: 265 case X86::VMOVUPSrm: 266 return ConvertToBroadcast(0, 0, X86::VMOVDDUPrm, X86::VBROADCASTSSrm, 0, 0, 267 1); 268 case X86::VMOVAPDYrm: 269 case X86::VMOVAPSYrm: 270 case X86::VMOVUPDYrm: 271 case X86::VMOVUPSYrm: 272 return ConvertToBroadcast(0, X86::VBROADCASTF128rm, X86::VBROADCASTSDYrm, 273 X86::VBROADCASTSSYrm, 0, 0, 1); 274 case X86::VMOVAPDZ128rm: 275 case X86::VMOVAPSZ128rm: 276 case X86::VMOVUPDZ128rm: 277 case X86::VMOVUPSZ128rm: 278 return ConvertToBroadcast(0, 0, X86::VMOVDDUPZ128rm, 279 X86::VBROADCASTSSZ128rm, 0, 0, 1); 280 case X86::VMOVAPDZ256rm: 281 case X86::VMOVAPSZ256rm: 282 case X86::VMOVUPDZ256rm: 283 case X86::VMOVUPSZ256rm: 284 return ConvertToBroadcast(0, X86::VBROADCASTF32X4Z256rm, 285 X86::VBROADCASTSDZ256rm, X86::VBROADCASTSSZ256rm, 286 0, 0, 1); 287 case X86::VMOVAPDZrm: 288 case X86::VMOVAPSZrm: 289 case X86::VMOVUPDZrm: 290 case X86::VMOVUPSZrm: 291 return ConvertToBroadcast(X86::VBROADCASTF64X4rm, X86::VBROADCASTF32X4rm, 292 X86::VBROADCASTSDZrm, X86::VBROADCASTSSZrm, 0, 0, 293 1); 294 /* Integer Loads */ 295 case X86::VMOVDQArm: 296 case X86::VMOVDQUrm: 297 return ConvertToBroadcast( 298 0, 0, HasAVX2 ? X86::VPBROADCASTQrm : X86::VMOVDDUPrm, 299 HasAVX2 ? X86::VPBROADCASTDrm : X86::VBROADCASTSSrm, 300 HasAVX2 ? X86::VPBROADCASTWrm : 0, HasAVX2 ? X86::VPBROADCASTBrm : 0, 301 1); 302 case X86::VMOVDQAYrm: 303 case X86::VMOVDQUYrm: 304 return ConvertToBroadcast( 305 0, HasAVX2 ? X86::VBROADCASTI128rm : X86::VBROADCASTF128rm, 306 HasAVX2 ? X86::VPBROADCASTQYrm : X86::VBROADCASTSDYrm, 307 HasAVX2 ? X86::VPBROADCASTDYrm : X86::VBROADCASTSSYrm, 308 HasAVX2 ? X86::VPBROADCASTWYrm : 0, HasAVX2 ? X86::VPBROADCASTBYrm : 0, 309 1); 310 case X86::VMOVDQA32Z128rm: 311 case X86::VMOVDQA64Z128rm: 312 case X86::VMOVDQU32Z128rm: 313 case X86::VMOVDQU64Z128rm: 314 return ConvertToBroadcast(0, 0, X86::VPBROADCASTQZ128rm, 315 X86::VPBROADCASTDZ128rm, 316 HasBWI ? X86::VPBROADCASTWZ128rm : 0, 317 HasBWI ? X86::VPBROADCASTBZ128rm : 0, 1); 318 case X86::VMOVDQA32Z256rm: 319 case X86::VMOVDQA64Z256rm: 320 case X86::VMOVDQU32Z256rm: 321 case X86::VMOVDQU64Z256rm: 322 return ConvertToBroadcast(0, X86::VBROADCASTI32X4Z256rm, 323 X86::VPBROADCASTQZ256rm, X86::VPBROADCASTDZ256rm, 324 HasBWI ? X86::VPBROADCASTWZ256rm : 0, 325 HasBWI ? X86::VPBROADCASTBZ256rm : 0, 1); 326 case X86::VMOVDQA32Zrm: 327 case X86::VMOVDQA64Zrm: 328 case X86::VMOVDQU32Zrm: 329 case X86::VMOVDQU64Zrm: 330 return ConvertToBroadcast(X86::VBROADCASTI64X4rm, X86::VBROADCASTI32X4rm, 331 X86::VPBROADCASTQZrm, X86::VPBROADCASTDZrm, 332 HasBWI ? X86::VPBROADCASTWZrm : 0, 333 HasBWI ? X86::VPBROADCASTBZrm : 0, 1); 334 } 335 336 auto ConvertToBroadcastAVX512 = [&](unsigned OpSrc32, unsigned OpSrc64) { 337 unsigned OpBcst32 = 0, OpBcst64 = 0; 338 unsigned OpNoBcst32 = 0, OpNoBcst64 = 0; 339 if (OpSrc32) { 340 if (const X86FoldTableEntry *Mem2Bcst = 341 llvm::lookupBroadcastFoldTable(OpSrc32, 32)) { 342 OpBcst32 = Mem2Bcst->DstOp; 343 OpNoBcst32 = Mem2Bcst->Flags & TB_INDEX_MASK; 344 } 345 } 346 if (OpSrc64) { 347 if (const X86FoldTableEntry *Mem2Bcst = 348 llvm::lookupBroadcastFoldTable(OpSrc64, 64)) { 349 OpBcst64 = Mem2Bcst->DstOp; 350 OpNoBcst64 = Mem2Bcst->Flags & TB_INDEX_MASK; 351 } 352 } 353 assert(((OpBcst32 == 0) || (OpBcst64 == 0) || (OpNoBcst32 == OpNoBcst64)) && 354 "OperandNo mismatch"); 355 356 if (OpBcst32 || OpBcst64) { 357 unsigned OpNo = OpBcst32 == 0 ? OpNoBcst64 : OpNoBcst32; 358 return ConvertToBroadcast(0, 0, OpBcst64, OpBcst32, 0, 0, OpNo); 359 } 360 return false; 361 }; 362 363 // Attempt to find a AVX512 mapping from a full width memory-fold instruction 364 // to a broadcast-fold instruction variant. 365 if ((MI.getDesc().TSFlags & X86II::EncodingMask) == X86II::EVEX) 366 return ConvertToBroadcastAVX512(Opc, Opc); 367 368 // Reverse the X86InstrInfo::setExecutionDomainCustom EVEX->VEX logic 369 // conversion to see if we can convert to a broadcasted (integer) logic op. 370 if (HasVLX && !HasDQI) { 371 unsigned OpSrc32 = 0, OpSrc64 = 0; 372 switch (Opc) { 373 case X86::VANDPDrm: 374 case X86::VANDPSrm: 375 case X86::VPANDrm: 376 OpSrc32 = X86 ::VPANDDZ128rm; 377 OpSrc64 = X86 ::VPANDQZ128rm; 378 break; 379 case X86::VANDPDYrm: 380 case X86::VANDPSYrm: 381 case X86::VPANDYrm: 382 OpSrc32 = X86 ::VPANDDZ256rm; 383 OpSrc64 = X86 ::VPANDQZ256rm; 384 break; 385 case X86::VANDNPDrm: 386 case X86::VANDNPSrm: 387 case X86::VPANDNrm: 388 OpSrc32 = X86 ::VPANDNDZ128rm; 389 OpSrc64 = X86 ::VPANDNQZ128rm; 390 break; 391 case X86::VANDNPDYrm: 392 case X86::VANDNPSYrm: 393 case X86::VPANDNYrm: 394 OpSrc32 = X86 ::VPANDNDZ256rm; 395 OpSrc64 = X86 ::VPANDNQZ256rm; 396 break; 397 case X86::VORPDrm: 398 case X86::VORPSrm: 399 case X86::VPORrm: 400 OpSrc32 = X86 ::VPORDZ128rm; 401 OpSrc64 = X86 ::VPORQZ128rm; 402 break; 403 case X86::VORPDYrm: 404 case X86::VORPSYrm: 405 case X86::VPORYrm: 406 OpSrc32 = X86 ::VPORDZ256rm; 407 OpSrc64 = X86 ::VPORQZ256rm; 408 break; 409 case X86::VXORPDrm: 410 case X86::VXORPSrm: 411 case X86::VPXORrm: 412 OpSrc32 = X86 ::VPXORDZ128rm; 413 OpSrc64 = X86 ::VPXORQZ128rm; 414 break; 415 case X86::VXORPDYrm: 416 case X86::VXORPSYrm: 417 case X86::VPXORYrm: 418 OpSrc32 = X86 ::VPXORDZ256rm; 419 OpSrc64 = X86 ::VPXORQZ256rm; 420 break; 421 } 422 if (OpSrc32 || OpSrc64) 423 return ConvertToBroadcastAVX512(OpSrc32, OpSrc64); 424 } 425 426 return false; 427 } 428 429 bool X86FixupVectorConstantsPass::runOnMachineFunction(MachineFunction &MF) { 430 LLVM_DEBUG(dbgs() << "Start X86FixupVectorConstants\n";); 431 bool Changed = false; 432 ST = &MF.getSubtarget<X86Subtarget>(); 433 TII = ST->getInstrInfo(); 434 SM = &ST->getSchedModel(); 435 436 for (MachineBasicBlock &MBB : MF) { 437 for (MachineInstr &MI : MBB) { 438 if (processInstruction(MF, MBB, MI)) { 439 ++NumInstChanges; 440 Changed = true; 441 } 442 } 443 } 444 LLVM_DEBUG(dbgs() << "End X86FixupVectorConstants\n";); 445 return Changed; 446 } 447