1 //===-- AMDGPUCodeGenPrepare.cpp ------------------------------------------===// 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 pass does misc. AMDGPU optimizations on IR before instruction 11 /// selection. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "AMDGPU.h" 16 #include "AMDGPUTargetMachine.h" 17 #include "llvm/Analysis/AssumptionCache.h" 18 #include "llvm/Analysis/ConstantFolding.h" 19 #include "llvm/Analysis/LegacyDivergenceAnalysis.h" 20 #include "llvm/Analysis/ValueTracking.h" 21 #include "llvm/CodeGen/TargetPassConfig.h" 22 #include "llvm/IR/Dominators.h" 23 #include "llvm/IR/InstVisitor.h" 24 #include "llvm/IR/IntrinsicsAMDGPU.h" 25 #include "llvm/IR/IRBuilder.h" 26 #include "llvm/InitializePasses.h" 27 #include "llvm/Pass.h" 28 #include "llvm/Support/KnownBits.h" 29 #include "llvm/Transforms/Utils/IntegerDivision.h" 30 31 #define DEBUG_TYPE "amdgpu-codegenprepare" 32 33 using namespace llvm; 34 35 namespace { 36 37 static cl::opt<bool> WidenLoads( 38 "amdgpu-codegenprepare-widen-constant-loads", 39 cl::desc("Widen sub-dword constant address space loads in AMDGPUCodeGenPrepare"), 40 cl::ReallyHidden, 41 cl::init(false)); 42 43 static cl::opt<bool> Widen16BitOps( 44 "amdgpu-codegenprepare-widen-16-bit-ops", 45 cl::desc("Widen uniform 16-bit instructions to 32-bit in AMDGPUCodeGenPrepare"), 46 cl::ReallyHidden, 47 cl::init(true)); 48 49 static cl::opt<bool> UseMul24Intrin( 50 "amdgpu-codegenprepare-mul24", 51 cl::desc("Introduce mul24 intrinsics in AMDGPUCodeGenPrepare"), 52 cl::ReallyHidden, 53 cl::init(true)); 54 55 // Legalize 64-bit division by using the generic IR expansion. 56 static cl::opt<bool> ExpandDiv64InIR( 57 "amdgpu-codegenprepare-expand-div64", 58 cl::desc("Expand 64-bit division in AMDGPUCodeGenPrepare"), 59 cl::ReallyHidden, 60 cl::init(false)); 61 62 // Leave all division operations as they are. This supersedes ExpandDiv64InIR 63 // and is used for testing the legalizer. 64 static cl::opt<bool> DisableIDivExpand( 65 "amdgpu-codegenprepare-disable-idiv-expansion", 66 cl::desc("Prevent expanding integer division in AMDGPUCodeGenPrepare"), 67 cl::ReallyHidden, 68 cl::init(false)); 69 70 class AMDGPUCodeGenPrepare : public FunctionPass, 71 public InstVisitor<AMDGPUCodeGenPrepare, bool> { 72 const GCNSubtarget *ST = nullptr; 73 AssumptionCache *AC = nullptr; 74 DominatorTree *DT = nullptr; 75 LegacyDivergenceAnalysis *DA = nullptr; 76 Module *Mod = nullptr; 77 const DataLayout *DL = nullptr; 78 bool HasUnsafeFPMath = false; 79 bool HasFP32Denormals = false; 80 81 /// Copies exact/nsw/nuw flags (if any) from binary operation \p I to 82 /// binary operation \p V. 83 /// 84 /// \returns Binary operation \p V. 85 /// \returns \p T's base element bit width. 86 unsigned getBaseElementBitWidth(const Type *T) const; 87 88 /// \returns Equivalent 32 bit integer type for given type \p T. For example, 89 /// if \p T is i7, then i32 is returned; if \p T is <3 x i12>, then <3 x i32> 90 /// is returned. 91 Type *getI32Ty(IRBuilder<> &B, const Type *T) const; 92 93 /// \returns True if binary operation \p I is a signed binary operation, false 94 /// otherwise. 95 bool isSigned(const BinaryOperator &I) const; 96 97 /// \returns True if the condition of 'select' operation \p I comes from a 98 /// signed 'icmp' operation, false otherwise. 99 bool isSigned(const SelectInst &I) const; 100 101 /// \returns True if type \p T needs to be promoted to 32 bit integer type, 102 /// false otherwise. 103 bool needsPromotionToI32(const Type *T) const; 104 105 /// Promotes uniform binary operation \p I to equivalent 32 bit binary 106 /// operation. 107 /// 108 /// \details \p I's base element bit width must be greater than 1 and less 109 /// than or equal 16. Promotion is done by sign or zero extending operands to 110 /// 32 bits, replacing \p I with equivalent 32 bit binary operation, and 111 /// truncating the result of 32 bit binary operation back to \p I's original 112 /// type. Division operation is not promoted. 113 /// 114 /// \returns True if \p I is promoted to equivalent 32 bit binary operation, 115 /// false otherwise. 116 bool promoteUniformOpToI32(BinaryOperator &I) const; 117 118 /// Promotes uniform 'icmp' operation \p I to 32 bit 'icmp' operation. 119 /// 120 /// \details \p I's base element bit width must be greater than 1 and less 121 /// than or equal 16. Promotion is done by sign or zero extending operands to 122 /// 32 bits, and replacing \p I with 32 bit 'icmp' operation. 123 /// 124 /// \returns True. 125 bool promoteUniformOpToI32(ICmpInst &I) const; 126 127 /// Promotes uniform 'select' operation \p I to 32 bit 'select' 128 /// operation. 129 /// 130 /// \details \p I's base element bit width must be greater than 1 and less 131 /// than or equal 16. Promotion is done by sign or zero extending operands to 132 /// 32 bits, replacing \p I with 32 bit 'select' operation, and truncating the 133 /// result of 32 bit 'select' operation back to \p I's original type. 134 /// 135 /// \returns True. 136 bool promoteUniformOpToI32(SelectInst &I) const; 137 138 /// Promotes uniform 'bitreverse' intrinsic \p I to 32 bit 'bitreverse' 139 /// intrinsic. 140 /// 141 /// \details \p I's base element bit width must be greater than 1 and less 142 /// than or equal 16. Promotion is done by zero extending the operand to 32 143 /// bits, replacing \p I with 32 bit 'bitreverse' intrinsic, shifting the 144 /// result of 32 bit 'bitreverse' intrinsic to the right with zero fill (the 145 /// shift amount is 32 minus \p I's base element bit width), and truncating 146 /// the result of the shift operation back to \p I's original type. 147 /// 148 /// \returns True. 149 bool promoteUniformBitreverseToI32(IntrinsicInst &I) const; 150 151 /// \returns The minimum number of bits needed to store the value of \Op as an 152 /// unsigned integer. Truncating to this size and then zero-extending to 153 /// the original will not change the value. 154 unsigned numBitsUnsigned(Value *Op) const; 155 156 /// \returns The minimum number of bits needed to store the value of \Op as a 157 /// signed integer. Truncating to this size and then sign-extending to 158 /// the original size will not change the value. 159 unsigned numBitsSigned(Value *Op) const; 160 161 /// Replace mul instructions with llvm.amdgcn.mul.u24 or llvm.amdgcn.mul.s24. 162 /// SelectionDAG has an issue where an and asserting the bits are known 163 bool replaceMulWithMul24(BinaryOperator &I) const; 164 165 /// Perform same function as equivalently named function in DAGCombiner. Since 166 /// we expand some divisions here, we need to perform this before obscuring. 167 bool foldBinOpIntoSelect(BinaryOperator &I) const; 168 169 bool divHasSpecialOptimization(BinaryOperator &I, 170 Value *Num, Value *Den) const; 171 int getDivNumBits(BinaryOperator &I, 172 Value *Num, Value *Den, 173 unsigned AtLeast, bool Signed) const; 174 175 /// Expands 24 bit div or rem. 176 Value* expandDivRem24(IRBuilder<> &Builder, BinaryOperator &I, 177 Value *Num, Value *Den, 178 bool IsDiv, bool IsSigned) const; 179 180 Value *expandDivRem24Impl(IRBuilder<> &Builder, BinaryOperator &I, 181 Value *Num, Value *Den, unsigned NumBits, 182 bool IsDiv, bool IsSigned) const; 183 184 /// Expands 32 bit div or rem. 185 Value* expandDivRem32(IRBuilder<> &Builder, BinaryOperator &I, 186 Value *Num, Value *Den) const; 187 188 Value *shrinkDivRem64(IRBuilder<> &Builder, BinaryOperator &I, 189 Value *Num, Value *Den) const; 190 void expandDivRem64(BinaryOperator &I) const; 191 192 /// Widen a scalar load. 193 /// 194 /// \details \p Widen scalar load for uniform, small type loads from constant 195 // memory / to a full 32-bits and then truncate the input to allow a scalar 196 // load instead of a vector load. 197 // 198 /// \returns True. 199 200 bool canWidenScalarExtLoad(LoadInst &I) const; 201 202 public: 203 static char ID; 204 205 AMDGPUCodeGenPrepare() : FunctionPass(ID) {} 206 207 bool visitFDiv(BinaryOperator &I); 208 bool visitXor(BinaryOperator &I); 209 210 bool visitInstruction(Instruction &I) { return false; } 211 bool visitBinaryOperator(BinaryOperator &I); 212 bool visitLoadInst(LoadInst &I); 213 bool visitICmpInst(ICmpInst &I); 214 bool visitSelectInst(SelectInst &I); 215 216 bool visitIntrinsicInst(IntrinsicInst &I); 217 bool visitBitreverseIntrinsicInst(IntrinsicInst &I); 218 219 bool doInitialization(Module &M) override; 220 bool runOnFunction(Function &F) override; 221 222 StringRef getPassName() const override { return "AMDGPU IR optimizations"; } 223 224 void getAnalysisUsage(AnalysisUsage &AU) const override { 225 AU.addRequired<AssumptionCacheTracker>(); 226 AU.addRequired<LegacyDivergenceAnalysis>(); 227 228 // FIXME: Division expansion needs to preserve the dominator tree. 229 if (!ExpandDiv64InIR) 230 AU.setPreservesAll(); 231 } 232 }; 233 234 } // end anonymous namespace 235 236 unsigned AMDGPUCodeGenPrepare::getBaseElementBitWidth(const Type *T) const { 237 assert(needsPromotionToI32(T) && "T does not need promotion to i32"); 238 239 if (T->isIntegerTy()) 240 return T->getIntegerBitWidth(); 241 return cast<VectorType>(T)->getElementType()->getIntegerBitWidth(); 242 } 243 244 Type *AMDGPUCodeGenPrepare::getI32Ty(IRBuilder<> &B, const Type *T) const { 245 assert(needsPromotionToI32(T) && "T does not need promotion to i32"); 246 247 if (T->isIntegerTy()) 248 return B.getInt32Ty(); 249 return FixedVectorType::get(B.getInt32Ty(), cast<FixedVectorType>(T)); 250 } 251 252 bool AMDGPUCodeGenPrepare::isSigned(const BinaryOperator &I) const { 253 return I.getOpcode() == Instruction::AShr || 254 I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::SRem; 255 } 256 257 bool AMDGPUCodeGenPrepare::isSigned(const SelectInst &I) const { 258 return isa<ICmpInst>(I.getOperand(0)) ? 259 cast<ICmpInst>(I.getOperand(0))->isSigned() : false; 260 } 261 262 bool AMDGPUCodeGenPrepare::needsPromotionToI32(const Type *T) const { 263 if (!Widen16BitOps) 264 return false; 265 266 const IntegerType *IntTy = dyn_cast<IntegerType>(T); 267 if (IntTy && IntTy->getBitWidth() > 1 && IntTy->getBitWidth() <= 16) 268 return true; 269 270 if (const VectorType *VT = dyn_cast<VectorType>(T)) { 271 // TODO: The set of packed operations is more limited, so may want to 272 // promote some anyway. 273 if (ST->hasVOP3PInsts()) 274 return false; 275 276 return needsPromotionToI32(VT->getElementType()); 277 } 278 279 return false; 280 } 281 282 // Return true if the op promoted to i32 should have nsw set. 283 static bool promotedOpIsNSW(const Instruction &I) { 284 switch (I.getOpcode()) { 285 case Instruction::Shl: 286 case Instruction::Add: 287 case Instruction::Sub: 288 return true; 289 case Instruction::Mul: 290 return I.hasNoUnsignedWrap(); 291 default: 292 return false; 293 } 294 } 295 296 // Return true if the op promoted to i32 should have nuw set. 297 static bool promotedOpIsNUW(const Instruction &I) { 298 switch (I.getOpcode()) { 299 case Instruction::Shl: 300 case Instruction::Add: 301 case Instruction::Mul: 302 return true; 303 case Instruction::Sub: 304 return I.hasNoUnsignedWrap(); 305 default: 306 return false; 307 } 308 } 309 310 bool AMDGPUCodeGenPrepare::canWidenScalarExtLoad(LoadInst &I) const { 311 Type *Ty = I.getType(); 312 const DataLayout &DL = Mod->getDataLayout(); 313 int TySize = DL.getTypeSizeInBits(Ty); 314 Align Alignment = DL.getValueOrABITypeAlignment(I.getAlign(), Ty); 315 316 return I.isSimple() && TySize < 32 && Alignment >= 4 && DA->isUniform(&I); 317 } 318 319 bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(BinaryOperator &I) const { 320 assert(needsPromotionToI32(I.getType()) && 321 "I does not need promotion to i32"); 322 323 if (I.getOpcode() == Instruction::SDiv || 324 I.getOpcode() == Instruction::UDiv || 325 I.getOpcode() == Instruction::SRem || 326 I.getOpcode() == Instruction::URem) 327 return false; 328 329 IRBuilder<> Builder(&I); 330 Builder.SetCurrentDebugLocation(I.getDebugLoc()); 331 332 Type *I32Ty = getI32Ty(Builder, I.getType()); 333 Value *ExtOp0 = nullptr; 334 Value *ExtOp1 = nullptr; 335 Value *ExtRes = nullptr; 336 Value *TruncRes = nullptr; 337 338 if (isSigned(I)) { 339 ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty); 340 ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty); 341 } else { 342 ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty); 343 ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty); 344 } 345 346 ExtRes = Builder.CreateBinOp(I.getOpcode(), ExtOp0, ExtOp1); 347 if (Instruction *Inst = dyn_cast<Instruction>(ExtRes)) { 348 if (promotedOpIsNSW(cast<Instruction>(I))) 349 Inst->setHasNoSignedWrap(); 350 351 if (promotedOpIsNUW(cast<Instruction>(I))) 352 Inst->setHasNoUnsignedWrap(); 353 354 if (const auto *ExactOp = dyn_cast<PossiblyExactOperator>(&I)) 355 Inst->setIsExact(ExactOp->isExact()); 356 } 357 358 TruncRes = Builder.CreateTrunc(ExtRes, I.getType()); 359 360 I.replaceAllUsesWith(TruncRes); 361 I.eraseFromParent(); 362 363 return true; 364 } 365 366 bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(ICmpInst &I) const { 367 assert(needsPromotionToI32(I.getOperand(0)->getType()) && 368 "I does not need promotion to i32"); 369 370 IRBuilder<> Builder(&I); 371 Builder.SetCurrentDebugLocation(I.getDebugLoc()); 372 373 Type *I32Ty = getI32Ty(Builder, I.getOperand(0)->getType()); 374 Value *ExtOp0 = nullptr; 375 Value *ExtOp1 = nullptr; 376 Value *NewICmp = nullptr; 377 378 if (I.isSigned()) { 379 ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty); 380 ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty); 381 } else { 382 ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty); 383 ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty); 384 } 385 NewICmp = Builder.CreateICmp(I.getPredicate(), ExtOp0, ExtOp1); 386 387 I.replaceAllUsesWith(NewICmp); 388 I.eraseFromParent(); 389 390 return true; 391 } 392 393 bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(SelectInst &I) const { 394 assert(needsPromotionToI32(I.getType()) && 395 "I does not need promotion to i32"); 396 397 IRBuilder<> Builder(&I); 398 Builder.SetCurrentDebugLocation(I.getDebugLoc()); 399 400 Type *I32Ty = getI32Ty(Builder, I.getType()); 401 Value *ExtOp1 = nullptr; 402 Value *ExtOp2 = nullptr; 403 Value *ExtRes = nullptr; 404 Value *TruncRes = nullptr; 405 406 if (isSigned(I)) { 407 ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty); 408 ExtOp2 = Builder.CreateSExt(I.getOperand(2), I32Ty); 409 } else { 410 ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty); 411 ExtOp2 = Builder.CreateZExt(I.getOperand(2), I32Ty); 412 } 413 ExtRes = Builder.CreateSelect(I.getOperand(0), ExtOp1, ExtOp2); 414 TruncRes = Builder.CreateTrunc(ExtRes, I.getType()); 415 416 I.replaceAllUsesWith(TruncRes); 417 I.eraseFromParent(); 418 419 return true; 420 } 421 422 bool AMDGPUCodeGenPrepare::promoteUniformBitreverseToI32( 423 IntrinsicInst &I) const { 424 assert(I.getIntrinsicID() == Intrinsic::bitreverse && 425 "I must be bitreverse intrinsic"); 426 assert(needsPromotionToI32(I.getType()) && 427 "I does not need promotion to i32"); 428 429 IRBuilder<> Builder(&I); 430 Builder.SetCurrentDebugLocation(I.getDebugLoc()); 431 432 Type *I32Ty = getI32Ty(Builder, I.getType()); 433 Function *I32 = 434 Intrinsic::getDeclaration(Mod, Intrinsic::bitreverse, { I32Ty }); 435 Value *ExtOp = Builder.CreateZExt(I.getOperand(0), I32Ty); 436 Value *ExtRes = Builder.CreateCall(I32, { ExtOp }); 437 Value *LShrOp = 438 Builder.CreateLShr(ExtRes, 32 - getBaseElementBitWidth(I.getType())); 439 Value *TruncRes = 440 Builder.CreateTrunc(LShrOp, I.getType()); 441 442 I.replaceAllUsesWith(TruncRes); 443 I.eraseFromParent(); 444 445 return true; 446 } 447 448 unsigned AMDGPUCodeGenPrepare::numBitsUnsigned(Value *Op) const { 449 return computeKnownBits(Op, *DL, 0, AC).countMaxActiveBits(); 450 } 451 452 unsigned AMDGPUCodeGenPrepare::numBitsSigned(Value *Op) const { 453 return ComputeMaxSignificantBits(Op, *DL, 0, AC); 454 } 455 456 static void extractValues(IRBuilder<> &Builder, 457 SmallVectorImpl<Value *> &Values, Value *V) { 458 auto *VT = dyn_cast<FixedVectorType>(V->getType()); 459 if (!VT) { 460 Values.push_back(V); 461 return; 462 } 463 464 for (int I = 0, E = VT->getNumElements(); I != E; ++I) 465 Values.push_back(Builder.CreateExtractElement(V, I)); 466 } 467 468 static Value *insertValues(IRBuilder<> &Builder, 469 Type *Ty, 470 SmallVectorImpl<Value *> &Values) { 471 if (Values.size() == 1) 472 return Values[0]; 473 474 Value *NewVal = UndefValue::get(Ty); 475 for (int I = 0, E = Values.size(); I != E; ++I) 476 NewVal = Builder.CreateInsertElement(NewVal, Values[I], I); 477 478 return NewVal; 479 } 480 481 // Returns 24-bit or 48-bit (as per `NumBits` and `Size`) mul of `LHS` and 482 // `RHS`. `NumBits` is the number of KnownBits of the result and `Size` is the 483 // width of the original destination. 484 static Value *getMul24(IRBuilder<> &Builder, Value *LHS, Value *RHS, 485 unsigned Size, unsigned NumBits, bool IsSigned) { 486 if (Size <= 32 || NumBits <= 32) { 487 Intrinsic::ID ID = 488 IsSigned ? Intrinsic::amdgcn_mul_i24 : Intrinsic::amdgcn_mul_u24; 489 return Builder.CreateIntrinsic(ID, {}, {LHS, RHS}); 490 } 491 492 assert(NumBits <= 48); 493 494 Intrinsic::ID LoID = 495 IsSigned ? Intrinsic::amdgcn_mul_i24 : Intrinsic::amdgcn_mul_u24; 496 Intrinsic::ID HiID = 497 IsSigned ? Intrinsic::amdgcn_mulhi_i24 : Intrinsic::amdgcn_mulhi_u24; 498 499 Value *Lo = Builder.CreateIntrinsic(LoID, {}, {LHS, RHS}); 500 Value *Hi = Builder.CreateIntrinsic(HiID, {}, {LHS, RHS}); 501 502 IntegerType *I64Ty = Builder.getInt64Ty(); 503 Lo = Builder.CreateZExtOrTrunc(Lo, I64Ty); 504 Hi = Builder.CreateZExtOrTrunc(Hi, I64Ty); 505 506 return Builder.CreateOr(Lo, Builder.CreateShl(Hi, 32)); 507 } 508 509 bool AMDGPUCodeGenPrepare::replaceMulWithMul24(BinaryOperator &I) const { 510 if (I.getOpcode() != Instruction::Mul) 511 return false; 512 513 Type *Ty = I.getType(); 514 unsigned Size = Ty->getScalarSizeInBits(); 515 if (Size <= 16 && ST->has16BitInsts()) 516 return false; 517 518 // Prefer scalar if this could be s_mul_i32 519 if (DA->isUniform(&I)) 520 return false; 521 522 Value *LHS = I.getOperand(0); 523 Value *RHS = I.getOperand(1); 524 IRBuilder<> Builder(&I); 525 Builder.SetCurrentDebugLocation(I.getDebugLoc()); 526 527 unsigned LHSBits = 0, RHSBits = 0; 528 bool IsSigned = false; 529 530 if (ST->hasMulU24() && (LHSBits = numBitsUnsigned(LHS)) <= 24 && 531 (RHSBits = numBitsUnsigned(RHS)) <= 24) { 532 IsSigned = false; 533 534 } else if (ST->hasMulI24() && (LHSBits = numBitsSigned(LHS)) <= 24 && 535 (RHSBits = numBitsSigned(RHS)) <= 24) { 536 IsSigned = true; 537 538 } else 539 return false; 540 541 SmallVector<Value *, 4> LHSVals; 542 SmallVector<Value *, 4> RHSVals; 543 SmallVector<Value *, 4> ResultVals; 544 extractValues(Builder, LHSVals, LHS); 545 extractValues(Builder, RHSVals, RHS); 546 547 IntegerType *I32Ty = Builder.getInt32Ty(); 548 for (int I = 0, E = LHSVals.size(); I != E; ++I) { 549 Value *LHS, *RHS; 550 if (IsSigned) { 551 LHS = Builder.CreateSExtOrTrunc(LHSVals[I], I32Ty); 552 RHS = Builder.CreateSExtOrTrunc(RHSVals[I], I32Ty); 553 } else { 554 LHS = Builder.CreateZExtOrTrunc(LHSVals[I], I32Ty); 555 RHS = Builder.CreateZExtOrTrunc(RHSVals[I], I32Ty); 556 } 557 558 Value *Result = 559 getMul24(Builder, LHS, RHS, Size, LHSBits + RHSBits, IsSigned); 560 561 if (IsSigned) { 562 ResultVals.push_back( 563 Builder.CreateSExtOrTrunc(Result, LHSVals[I]->getType())); 564 } else { 565 ResultVals.push_back( 566 Builder.CreateZExtOrTrunc(Result, LHSVals[I]->getType())); 567 } 568 } 569 570 Value *NewVal = insertValues(Builder, Ty, ResultVals); 571 NewVal->takeName(&I); 572 I.replaceAllUsesWith(NewVal); 573 I.eraseFromParent(); 574 575 return true; 576 } 577 578 // Find a select instruction, which may have been casted. This is mostly to deal 579 // with cases where i16 selects were promoted here to i32. 580 static SelectInst *findSelectThroughCast(Value *V, CastInst *&Cast) { 581 Cast = nullptr; 582 if (SelectInst *Sel = dyn_cast<SelectInst>(V)) 583 return Sel; 584 585 if ((Cast = dyn_cast<CastInst>(V))) { 586 if (SelectInst *Sel = dyn_cast<SelectInst>(Cast->getOperand(0))) 587 return Sel; 588 } 589 590 return nullptr; 591 } 592 593 bool AMDGPUCodeGenPrepare::foldBinOpIntoSelect(BinaryOperator &BO) const { 594 // Don't do this unless the old select is going away. We want to eliminate the 595 // binary operator, not replace a binop with a select. 596 int SelOpNo = 0; 597 598 CastInst *CastOp; 599 600 // TODO: Should probably try to handle some cases with multiple 601 // users. Duplicating the select may be profitable for division. 602 SelectInst *Sel = findSelectThroughCast(BO.getOperand(0), CastOp); 603 if (!Sel || !Sel->hasOneUse()) { 604 SelOpNo = 1; 605 Sel = findSelectThroughCast(BO.getOperand(1), CastOp); 606 } 607 608 if (!Sel || !Sel->hasOneUse()) 609 return false; 610 611 Constant *CT = dyn_cast<Constant>(Sel->getTrueValue()); 612 Constant *CF = dyn_cast<Constant>(Sel->getFalseValue()); 613 Constant *CBO = dyn_cast<Constant>(BO.getOperand(SelOpNo ^ 1)); 614 if (!CBO || !CT || !CF) 615 return false; 616 617 if (CastOp) { 618 if (!CastOp->hasOneUse()) 619 return false; 620 CT = ConstantFoldCastOperand(CastOp->getOpcode(), CT, BO.getType(), *DL); 621 CF = ConstantFoldCastOperand(CastOp->getOpcode(), CF, BO.getType(), *DL); 622 } 623 624 // TODO: Handle special 0/-1 cases DAG combine does, although we only really 625 // need to handle divisions here. 626 Constant *FoldedT = SelOpNo ? 627 ConstantFoldBinaryOpOperands(BO.getOpcode(), CBO, CT, *DL) : 628 ConstantFoldBinaryOpOperands(BO.getOpcode(), CT, CBO, *DL); 629 if (isa<ConstantExpr>(FoldedT)) 630 return false; 631 632 Constant *FoldedF = SelOpNo ? 633 ConstantFoldBinaryOpOperands(BO.getOpcode(), CBO, CF, *DL) : 634 ConstantFoldBinaryOpOperands(BO.getOpcode(), CF, CBO, *DL); 635 if (isa<ConstantExpr>(FoldedF)) 636 return false; 637 638 IRBuilder<> Builder(&BO); 639 Builder.SetCurrentDebugLocation(BO.getDebugLoc()); 640 if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(&BO)) 641 Builder.setFastMathFlags(FPOp->getFastMathFlags()); 642 643 Value *NewSelect = Builder.CreateSelect(Sel->getCondition(), 644 FoldedT, FoldedF); 645 NewSelect->takeName(&BO); 646 BO.replaceAllUsesWith(NewSelect); 647 BO.eraseFromParent(); 648 if (CastOp) 649 CastOp->eraseFromParent(); 650 Sel->eraseFromParent(); 651 return true; 652 } 653 654 // Optimize fdiv with rcp: 655 // 656 // 1/x -> rcp(x) when rcp is sufficiently accurate or inaccurate rcp is 657 // allowed with unsafe-fp-math or afn. 658 // 659 // a/b -> a*rcp(b) when inaccurate rcp is allowed with unsafe-fp-math or afn. 660 static Value *optimizeWithRcp(Value *Num, Value *Den, bool AllowInaccurateRcp, 661 bool RcpIsAccurate, IRBuilder<> &Builder, 662 Module *Mod) { 663 664 if (!AllowInaccurateRcp && !RcpIsAccurate) 665 return nullptr; 666 667 Type *Ty = Den->getType(); 668 if (const ConstantFP *CLHS = dyn_cast<ConstantFP>(Num)) { 669 if (AllowInaccurateRcp || RcpIsAccurate) { 670 if (CLHS->isExactlyValue(1.0)) { 671 Function *Decl = Intrinsic::getDeclaration( 672 Mod, Intrinsic::amdgcn_rcp, Ty); 673 674 // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to 675 // the CI documentation has a worst case error of 1 ulp. 676 // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to 677 // use it as long as we aren't trying to use denormals. 678 // 679 // v_rcp_f16 and v_rsq_f16 DO support denormals. 680 681 // NOTE: v_sqrt and v_rcp will be combined to v_rsq later. So we don't 682 // insert rsq intrinsic here. 683 684 // 1.0 / x -> rcp(x) 685 return Builder.CreateCall(Decl, { Den }); 686 } 687 688 // Same as for 1.0, but expand the sign out of the constant. 689 if (CLHS->isExactlyValue(-1.0)) { 690 Function *Decl = Intrinsic::getDeclaration( 691 Mod, Intrinsic::amdgcn_rcp, Ty); 692 693 // -1.0 / x -> rcp (fneg x) 694 Value *FNeg = Builder.CreateFNeg(Den); 695 return Builder.CreateCall(Decl, { FNeg }); 696 } 697 } 698 } 699 700 if (AllowInaccurateRcp) { 701 Function *Decl = Intrinsic::getDeclaration( 702 Mod, Intrinsic::amdgcn_rcp, Ty); 703 704 // Turn into multiply by the reciprocal. 705 // x / y -> x * (1.0 / y) 706 Value *Recip = Builder.CreateCall(Decl, { Den }); 707 return Builder.CreateFMul(Num, Recip); 708 } 709 return nullptr; 710 } 711 712 // optimize with fdiv.fast: 713 // 714 // a/b -> fdiv.fast(a, b) when !fpmath >= 2.5ulp with denormals flushed. 715 // 716 // 1/x -> fdiv.fast(1,x) when !fpmath >= 2.5ulp. 717 // 718 // NOTE: optimizeWithRcp should be tried first because rcp is the preference. 719 static Value *optimizeWithFDivFast(Value *Num, Value *Den, float ReqdAccuracy, 720 bool HasDenormals, IRBuilder<> &Builder, 721 Module *Mod) { 722 // fdiv.fast can achieve 2.5 ULP accuracy. 723 if (ReqdAccuracy < 2.5f) 724 return nullptr; 725 726 // Only have fdiv.fast for f32. 727 Type *Ty = Den->getType(); 728 if (!Ty->isFloatTy()) 729 return nullptr; 730 731 bool NumIsOne = false; 732 if (const ConstantFP *CNum = dyn_cast<ConstantFP>(Num)) { 733 if (CNum->isExactlyValue(+1.0) || CNum->isExactlyValue(-1.0)) 734 NumIsOne = true; 735 } 736 737 // fdiv does not support denormals. But 1.0/x is always fine to use it. 738 if (HasDenormals && !NumIsOne) 739 return nullptr; 740 741 Function *Decl = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_fdiv_fast); 742 return Builder.CreateCall(Decl, { Num, Den }); 743 } 744 745 // Optimizations is performed based on fpmath, fast math flags as well as 746 // denormals to optimize fdiv with either rcp or fdiv.fast. 747 // 748 // With rcp: 749 // 1/x -> rcp(x) when rcp is sufficiently accurate or inaccurate rcp is 750 // allowed with unsafe-fp-math or afn. 751 // 752 // a/b -> a*rcp(b) when inaccurate rcp is allowed with unsafe-fp-math or afn. 753 // 754 // With fdiv.fast: 755 // a/b -> fdiv.fast(a, b) when !fpmath >= 2.5ulp with denormals flushed. 756 // 757 // 1/x -> fdiv.fast(1,x) when !fpmath >= 2.5ulp. 758 // 759 // NOTE: rcp is the preference in cases that both are legal. 760 bool AMDGPUCodeGenPrepare::visitFDiv(BinaryOperator &FDiv) { 761 762 Type *Ty = FDiv.getType()->getScalarType(); 763 764 // The f64 rcp/rsq approximations are pretty inaccurate. We can do an 765 // expansion around them in codegen. 766 if (Ty->isDoubleTy()) 767 return false; 768 769 // No intrinsic for fdiv16 if target does not support f16. 770 if (Ty->isHalfTy() && !ST->has16BitInsts()) 771 return false; 772 773 const FPMathOperator *FPOp = cast<const FPMathOperator>(&FDiv); 774 const float ReqdAccuracy = FPOp->getFPAccuracy(); 775 776 // Inaccurate rcp is allowed with unsafe-fp-math or afn. 777 FastMathFlags FMF = FPOp->getFastMathFlags(); 778 const bool AllowInaccurateRcp = HasUnsafeFPMath || FMF.approxFunc(); 779 780 // rcp_f16 is accurate for !fpmath >= 1.0ulp. 781 // rcp_f32 is accurate for !fpmath >= 1.0ulp and denormals are flushed. 782 // rcp_f64 is never accurate. 783 const bool RcpIsAccurate = (Ty->isHalfTy() && ReqdAccuracy >= 1.0f) || 784 (Ty->isFloatTy() && !HasFP32Denormals && ReqdAccuracy >= 1.0f); 785 786 IRBuilder<> Builder(FDiv.getParent(), std::next(FDiv.getIterator())); 787 Builder.setFastMathFlags(FMF); 788 Builder.SetCurrentDebugLocation(FDiv.getDebugLoc()); 789 790 Value *Num = FDiv.getOperand(0); 791 Value *Den = FDiv.getOperand(1); 792 793 Value *NewFDiv = nullptr; 794 if (auto *VT = dyn_cast<FixedVectorType>(FDiv.getType())) { 795 NewFDiv = UndefValue::get(VT); 796 797 // FIXME: Doesn't do the right thing for cases where the vector is partially 798 // constant. This works when the scalarizer pass is run first. 799 for (unsigned I = 0, E = VT->getNumElements(); I != E; ++I) { 800 Value *NumEltI = Builder.CreateExtractElement(Num, I); 801 Value *DenEltI = Builder.CreateExtractElement(Den, I); 802 // Try rcp first. 803 Value *NewElt = optimizeWithRcp(NumEltI, DenEltI, AllowInaccurateRcp, 804 RcpIsAccurate, Builder, Mod); 805 if (!NewElt) // Try fdiv.fast. 806 NewElt = optimizeWithFDivFast(NumEltI, DenEltI, ReqdAccuracy, 807 HasFP32Denormals, Builder, Mod); 808 if (!NewElt) // Keep the original. 809 NewElt = Builder.CreateFDiv(NumEltI, DenEltI); 810 811 NewFDiv = Builder.CreateInsertElement(NewFDiv, NewElt, I); 812 } 813 } else { // Scalar FDiv. 814 // Try rcp first. 815 NewFDiv = optimizeWithRcp(Num, Den, AllowInaccurateRcp, RcpIsAccurate, 816 Builder, Mod); 817 if (!NewFDiv) { // Try fdiv.fast. 818 NewFDiv = optimizeWithFDivFast(Num, Den, ReqdAccuracy, HasFP32Denormals, 819 Builder, Mod); 820 } 821 } 822 823 if (NewFDiv) { 824 FDiv.replaceAllUsesWith(NewFDiv); 825 NewFDiv->takeName(&FDiv); 826 FDiv.eraseFromParent(); 827 } 828 829 return !!NewFDiv; 830 } 831 832 bool AMDGPUCodeGenPrepare::visitXor(BinaryOperator &I) { 833 // Match the Xor instruction, its type and its operands 834 IntrinsicInst *IntrinsicCall = dyn_cast<IntrinsicInst>(I.getOperand(0)); 835 ConstantInt *RHS = dyn_cast<ConstantInt>(I.getOperand(1)); 836 if (!RHS || !IntrinsicCall || RHS->getSExtValue() != -1) 837 return visitBinaryOperator(I); 838 839 // Check if the Call is an intrinsic instruction to amdgcn_class intrinsic 840 // has only one use 841 if (IntrinsicCall->getIntrinsicID() != Intrinsic::amdgcn_class || 842 !IntrinsicCall->hasOneUse()) 843 return visitBinaryOperator(I); 844 845 // "Not" the second argument of the intrinsic call 846 ConstantInt *Arg = dyn_cast<ConstantInt>(IntrinsicCall->getOperand(1)); 847 if (!Arg) 848 return visitBinaryOperator(I); 849 850 IntrinsicCall->setOperand( 851 1, ConstantInt::get(Arg->getType(), Arg->getZExtValue() ^ 0x3ff)); 852 I.replaceAllUsesWith(IntrinsicCall); 853 I.eraseFromParent(); 854 return true; 855 } 856 857 static bool hasUnsafeFPMath(const Function &F) { 858 Attribute Attr = F.getFnAttribute("unsafe-fp-math"); 859 return Attr.getValueAsBool(); 860 } 861 862 static std::pair<Value*, Value*> getMul64(IRBuilder<> &Builder, 863 Value *LHS, Value *RHS) { 864 Type *I32Ty = Builder.getInt32Ty(); 865 Type *I64Ty = Builder.getInt64Ty(); 866 867 Value *LHS_EXT64 = Builder.CreateZExt(LHS, I64Ty); 868 Value *RHS_EXT64 = Builder.CreateZExt(RHS, I64Ty); 869 Value *MUL64 = Builder.CreateMul(LHS_EXT64, RHS_EXT64); 870 Value *Lo = Builder.CreateTrunc(MUL64, I32Ty); 871 Value *Hi = Builder.CreateLShr(MUL64, Builder.getInt64(32)); 872 Hi = Builder.CreateTrunc(Hi, I32Ty); 873 return std::make_pair(Lo, Hi); 874 } 875 876 static Value* getMulHu(IRBuilder<> &Builder, Value *LHS, Value *RHS) { 877 return getMul64(Builder, LHS, RHS).second; 878 } 879 880 /// Figure out how many bits are really needed for this ddivision. \p AtLeast is 881 /// an optimization hint to bypass the second ComputeNumSignBits call if we the 882 /// first one is insufficient. Returns -1 on failure. 883 int AMDGPUCodeGenPrepare::getDivNumBits(BinaryOperator &I, 884 Value *Num, Value *Den, 885 unsigned AtLeast, bool IsSigned) const { 886 const DataLayout &DL = Mod->getDataLayout(); 887 unsigned LHSSignBits = ComputeNumSignBits(Num, DL, 0, AC, &I); 888 if (LHSSignBits < AtLeast) 889 return -1; 890 891 unsigned RHSSignBits = ComputeNumSignBits(Den, DL, 0, AC, &I); 892 if (RHSSignBits < AtLeast) 893 return -1; 894 895 unsigned SignBits = std::min(LHSSignBits, RHSSignBits); 896 unsigned DivBits = Num->getType()->getScalarSizeInBits() - SignBits; 897 if (IsSigned) 898 ++DivBits; 899 return DivBits; 900 } 901 902 // The fractional part of a float is enough to accurately represent up to 903 // a 24-bit signed integer. 904 Value *AMDGPUCodeGenPrepare::expandDivRem24(IRBuilder<> &Builder, 905 BinaryOperator &I, 906 Value *Num, Value *Den, 907 bool IsDiv, bool IsSigned) const { 908 int DivBits = getDivNumBits(I, Num, Den, 9, IsSigned); 909 if (DivBits == -1) 910 return nullptr; 911 return expandDivRem24Impl(Builder, I, Num, Den, DivBits, IsDiv, IsSigned); 912 } 913 914 Value *AMDGPUCodeGenPrepare::expandDivRem24Impl(IRBuilder<> &Builder, 915 BinaryOperator &I, 916 Value *Num, Value *Den, 917 unsigned DivBits, 918 bool IsDiv, bool IsSigned) const { 919 Type *I32Ty = Builder.getInt32Ty(); 920 Num = Builder.CreateTrunc(Num, I32Ty); 921 Den = Builder.CreateTrunc(Den, I32Ty); 922 923 Type *F32Ty = Builder.getFloatTy(); 924 ConstantInt *One = Builder.getInt32(1); 925 Value *JQ = One; 926 927 if (IsSigned) { 928 // char|short jq = ia ^ ib; 929 JQ = Builder.CreateXor(Num, Den); 930 931 // jq = jq >> (bitsize - 2) 932 JQ = Builder.CreateAShr(JQ, Builder.getInt32(30)); 933 934 // jq = jq | 0x1 935 JQ = Builder.CreateOr(JQ, One); 936 } 937 938 // int ia = (int)LHS; 939 Value *IA = Num; 940 941 // int ib, (int)RHS; 942 Value *IB = Den; 943 944 // float fa = (float)ia; 945 Value *FA = IsSigned ? Builder.CreateSIToFP(IA, F32Ty) 946 : Builder.CreateUIToFP(IA, F32Ty); 947 948 // float fb = (float)ib; 949 Value *FB = IsSigned ? Builder.CreateSIToFP(IB,F32Ty) 950 : Builder.CreateUIToFP(IB,F32Ty); 951 952 Function *RcpDecl = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_rcp, 953 Builder.getFloatTy()); 954 Value *RCP = Builder.CreateCall(RcpDecl, { FB }); 955 Value *FQM = Builder.CreateFMul(FA, RCP); 956 957 // fq = trunc(fqm); 958 CallInst *FQ = Builder.CreateUnaryIntrinsic(Intrinsic::trunc, FQM); 959 FQ->copyFastMathFlags(Builder.getFastMathFlags()); 960 961 // float fqneg = -fq; 962 Value *FQNeg = Builder.CreateFNeg(FQ); 963 964 // float fr = mad(fqneg, fb, fa); 965 auto FMAD = !ST->hasMadMacF32Insts() 966 ? Intrinsic::fma 967 : (Intrinsic::ID)Intrinsic::amdgcn_fmad_ftz; 968 Value *FR = Builder.CreateIntrinsic(FMAD, 969 {FQNeg->getType()}, {FQNeg, FB, FA}, FQ); 970 971 // int iq = (int)fq; 972 Value *IQ = IsSigned ? Builder.CreateFPToSI(FQ, I32Ty) 973 : Builder.CreateFPToUI(FQ, I32Ty); 974 975 // fr = fabs(fr); 976 FR = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FR, FQ); 977 978 // fb = fabs(fb); 979 FB = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FB, FQ); 980 981 // int cv = fr >= fb; 982 Value *CV = Builder.CreateFCmpOGE(FR, FB); 983 984 // jq = (cv ? jq : 0); 985 JQ = Builder.CreateSelect(CV, JQ, Builder.getInt32(0)); 986 987 // dst = iq + jq; 988 Value *Div = Builder.CreateAdd(IQ, JQ); 989 990 Value *Res = Div; 991 if (!IsDiv) { 992 // Rem needs compensation, it's easier to recompute it 993 Value *Rem = Builder.CreateMul(Div, Den); 994 Res = Builder.CreateSub(Num, Rem); 995 } 996 997 if (DivBits != 0 && DivBits < 32) { 998 // Extend in register from the number of bits this divide really is. 999 if (IsSigned) { 1000 int InRegBits = 32 - DivBits; 1001 1002 Res = Builder.CreateShl(Res, InRegBits); 1003 Res = Builder.CreateAShr(Res, InRegBits); 1004 } else { 1005 ConstantInt *TruncMask 1006 = Builder.getInt32((UINT64_C(1) << DivBits) - 1); 1007 Res = Builder.CreateAnd(Res, TruncMask); 1008 } 1009 } 1010 1011 return Res; 1012 } 1013 1014 // Try to recognize special cases the DAG will emit special, better expansions 1015 // than the general expansion we do here. 1016 1017 // TODO: It would be better to just directly handle those optimizations here. 1018 bool AMDGPUCodeGenPrepare::divHasSpecialOptimization( 1019 BinaryOperator &I, Value *Num, Value *Den) const { 1020 if (Constant *C = dyn_cast<Constant>(Den)) { 1021 // Arbitrary constants get a better expansion as long as a wider mulhi is 1022 // legal. 1023 if (C->getType()->getScalarSizeInBits() <= 32) 1024 return true; 1025 1026 // TODO: Sdiv check for not exact for some reason. 1027 1028 // If there's no wider mulhi, there's only a better expansion for powers of 1029 // two. 1030 // TODO: Should really know for each vector element. 1031 if (isKnownToBeAPowerOfTwo(C, *DL, true, 0, AC, &I, DT)) 1032 return true; 1033 1034 return false; 1035 } 1036 1037 if (BinaryOperator *BinOpDen = dyn_cast<BinaryOperator>(Den)) { 1038 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2 1039 if (BinOpDen->getOpcode() == Instruction::Shl && 1040 isa<Constant>(BinOpDen->getOperand(0)) && 1041 isKnownToBeAPowerOfTwo(BinOpDen->getOperand(0), *DL, true, 1042 0, AC, &I, DT)) { 1043 return true; 1044 } 1045 } 1046 1047 return false; 1048 } 1049 1050 static Value *getSign32(Value *V, IRBuilder<> &Builder, const DataLayout *DL) { 1051 // Check whether the sign can be determined statically. 1052 KnownBits Known = computeKnownBits(V, *DL); 1053 if (Known.isNegative()) 1054 return Constant::getAllOnesValue(V->getType()); 1055 if (Known.isNonNegative()) 1056 return Constant::getNullValue(V->getType()); 1057 return Builder.CreateAShr(V, Builder.getInt32(31)); 1058 } 1059 1060 Value *AMDGPUCodeGenPrepare::expandDivRem32(IRBuilder<> &Builder, 1061 BinaryOperator &I, Value *X, 1062 Value *Y) const { 1063 Instruction::BinaryOps Opc = I.getOpcode(); 1064 assert(Opc == Instruction::URem || Opc == Instruction::UDiv || 1065 Opc == Instruction::SRem || Opc == Instruction::SDiv); 1066 1067 FastMathFlags FMF; 1068 FMF.setFast(); 1069 Builder.setFastMathFlags(FMF); 1070 1071 if (divHasSpecialOptimization(I, X, Y)) 1072 return nullptr; // Keep it for later optimization. 1073 1074 bool IsDiv = Opc == Instruction::UDiv || Opc == Instruction::SDiv; 1075 bool IsSigned = Opc == Instruction::SRem || Opc == Instruction::SDiv; 1076 1077 Type *Ty = X->getType(); 1078 Type *I32Ty = Builder.getInt32Ty(); 1079 Type *F32Ty = Builder.getFloatTy(); 1080 1081 if (Ty->getScalarSizeInBits() < 32) { 1082 if (IsSigned) { 1083 X = Builder.CreateSExt(X, I32Ty); 1084 Y = Builder.CreateSExt(Y, I32Ty); 1085 } else { 1086 X = Builder.CreateZExt(X, I32Ty); 1087 Y = Builder.CreateZExt(Y, I32Ty); 1088 } 1089 } 1090 1091 if (Value *Res = expandDivRem24(Builder, I, X, Y, IsDiv, IsSigned)) { 1092 return IsSigned ? Builder.CreateSExtOrTrunc(Res, Ty) : 1093 Builder.CreateZExtOrTrunc(Res, Ty); 1094 } 1095 1096 ConstantInt *Zero = Builder.getInt32(0); 1097 ConstantInt *One = Builder.getInt32(1); 1098 1099 Value *Sign = nullptr; 1100 if (IsSigned) { 1101 Value *SignX = getSign32(X, Builder, DL); 1102 Value *SignY = getSign32(Y, Builder, DL); 1103 // Remainder sign is the same as LHS 1104 Sign = IsDiv ? Builder.CreateXor(SignX, SignY) : SignX; 1105 1106 X = Builder.CreateAdd(X, SignX); 1107 Y = Builder.CreateAdd(Y, SignY); 1108 1109 X = Builder.CreateXor(X, SignX); 1110 Y = Builder.CreateXor(Y, SignY); 1111 } 1112 1113 // The algorithm here is based on ideas from "Software Integer Division", Tom 1114 // Rodeheffer, August 2008. 1115 // 1116 // unsigned udiv(unsigned x, unsigned y) { 1117 // // Initial estimate of inv(y). The constant is less than 2^32 to ensure 1118 // // that this is a lower bound on inv(y), even if some of the calculations 1119 // // round up. 1120 // unsigned z = (unsigned)((4294967296.0 - 512.0) * v_rcp_f32((float)y)); 1121 // 1122 // // One round of UNR (Unsigned integer Newton-Raphson) to improve z. 1123 // // Empirically this is guaranteed to give a "two-y" lower bound on 1124 // // inv(y). 1125 // z += umulh(z, -y * z); 1126 // 1127 // // Quotient/remainder estimate. 1128 // unsigned q = umulh(x, z); 1129 // unsigned r = x - q * y; 1130 // 1131 // // Two rounds of quotient/remainder refinement. 1132 // if (r >= y) { 1133 // ++q; 1134 // r -= y; 1135 // } 1136 // if (r >= y) { 1137 // ++q; 1138 // r -= y; 1139 // } 1140 // 1141 // return q; 1142 // } 1143 1144 // Initial estimate of inv(y). 1145 Value *FloatY = Builder.CreateUIToFP(Y, F32Ty); 1146 Function *Rcp = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_rcp, F32Ty); 1147 Value *RcpY = Builder.CreateCall(Rcp, {FloatY}); 1148 Constant *Scale = ConstantFP::get(F32Ty, BitsToFloat(0x4F7FFFFE)); 1149 Value *ScaledY = Builder.CreateFMul(RcpY, Scale); 1150 Value *Z = Builder.CreateFPToUI(ScaledY, I32Ty); 1151 1152 // One round of UNR. 1153 Value *NegY = Builder.CreateSub(Zero, Y); 1154 Value *NegYZ = Builder.CreateMul(NegY, Z); 1155 Z = Builder.CreateAdd(Z, getMulHu(Builder, Z, NegYZ)); 1156 1157 // Quotient/remainder estimate. 1158 Value *Q = getMulHu(Builder, X, Z); 1159 Value *R = Builder.CreateSub(X, Builder.CreateMul(Q, Y)); 1160 1161 // First quotient/remainder refinement. 1162 Value *Cond = Builder.CreateICmpUGE(R, Y); 1163 if (IsDiv) 1164 Q = Builder.CreateSelect(Cond, Builder.CreateAdd(Q, One), Q); 1165 R = Builder.CreateSelect(Cond, Builder.CreateSub(R, Y), R); 1166 1167 // Second quotient/remainder refinement. 1168 Cond = Builder.CreateICmpUGE(R, Y); 1169 Value *Res; 1170 if (IsDiv) 1171 Res = Builder.CreateSelect(Cond, Builder.CreateAdd(Q, One), Q); 1172 else 1173 Res = Builder.CreateSelect(Cond, Builder.CreateSub(R, Y), R); 1174 1175 if (IsSigned) { 1176 Res = Builder.CreateXor(Res, Sign); 1177 Res = Builder.CreateSub(Res, Sign); 1178 } 1179 1180 Res = Builder.CreateTrunc(Res, Ty); 1181 1182 return Res; 1183 } 1184 1185 Value *AMDGPUCodeGenPrepare::shrinkDivRem64(IRBuilder<> &Builder, 1186 BinaryOperator &I, 1187 Value *Num, Value *Den) const { 1188 if (!ExpandDiv64InIR && divHasSpecialOptimization(I, Num, Den)) 1189 return nullptr; // Keep it for later optimization. 1190 1191 Instruction::BinaryOps Opc = I.getOpcode(); 1192 1193 bool IsDiv = Opc == Instruction::SDiv || Opc == Instruction::UDiv; 1194 bool IsSigned = Opc == Instruction::SDiv || Opc == Instruction::SRem; 1195 1196 int NumDivBits = getDivNumBits(I, Num, Den, 32, IsSigned); 1197 if (NumDivBits == -1) 1198 return nullptr; 1199 1200 Value *Narrowed = nullptr; 1201 if (NumDivBits <= 24) { 1202 Narrowed = expandDivRem24Impl(Builder, I, Num, Den, NumDivBits, 1203 IsDiv, IsSigned); 1204 } else if (NumDivBits <= 32) { 1205 Narrowed = expandDivRem32(Builder, I, Num, Den); 1206 } 1207 1208 if (Narrowed) { 1209 return IsSigned ? Builder.CreateSExt(Narrowed, Num->getType()) : 1210 Builder.CreateZExt(Narrowed, Num->getType()); 1211 } 1212 1213 return nullptr; 1214 } 1215 1216 void AMDGPUCodeGenPrepare::expandDivRem64(BinaryOperator &I) const { 1217 Instruction::BinaryOps Opc = I.getOpcode(); 1218 // Do the general expansion. 1219 if (Opc == Instruction::UDiv || Opc == Instruction::SDiv) { 1220 expandDivisionUpTo64Bits(&I); 1221 return; 1222 } 1223 1224 if (Opc == Instruction::URem || Opc == Instruction::SRem) { 1225 expandRemainderUpTo64Bits(&I); 1226 return; 1227 } 1228 1229 llvm_unreachable("not a division"); 1230 } 1231 1232 bool AMDGPUCodeGenPrepare::visitBinaryOperator(BinaryOperator &I) { 1233 if (foldBinOpIntoSelect(I)) 1234 return true; 1235 1236 if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) && 1237 DA->isUniform(&I) && promoteUniformOpToI32(I)) 1238 return true; 1239 1240 if (UseMul24Intrin && replaceMulWithMul24(I)) 1241 return true; 1242 1243 bool Changed = false; 1244 Instruction::BinaryOps Opc = I.getOpcode(); 1245 Type *Ty = I.getType(); 1246 Value *NewDiv = nullptr; 1247 unsigned ScalarSize = Ty->getScalarSizeInBits(); 1248 1249 SmallVector<BinaryOperator *, 8> Div64ToExpand; 1250 1251 if ((Opc == Instruction::URem || Opc == Instruction::UDiv || 1252 Opc == Instruction::SRem || Opc == Instruction::SDiv) && 1253 ScalarSize <= 64 && 1254 !DisableIDivExpand) { 1255 Value *Num = I.getOperand(0); 1256 Value *Den = I.getOperand(1); 1257 IRBuilder<> Builder(&I); 1258 Builder.SetCurrentDebugLocation(I.getDebugLoc()); 1259 1260 if (auto *VT = dyn_cast<FixedVectorType>(Ty)) { 1261 NewDiv = UndefValue::get(VT); 1262 1263 for (unsigned N = 0, E = VT->getNumElements(); N != E; ++N) { 1264 Value *NumEltN = Builder.CreateExtractElement(Num, N); 1265 Value *DenEltN = Builder.CreateExtractElement(Den, N); 1266 1267 Value *NewElt; 1268 if (ScalarSize <= 32) { 1269 NewElt = expandDivRem32(Builder, I, NumEltN, DenEltN); 1270 if (!NewElt) 1271 NewElt = Builder.CreateBinOp(Opc, NumEltN, DenEltN); 1272 } else { 1273 // See if this 64-bit division can be shrunk to 32/24-bits before 1274 // producing the general expansion. 1275 NewElt = shrinkDivRem64(Builder, I, NumEltN, DenEltN); 1276 if (!NewElt) { 1277 // The general 64-bit expansion introduces control flow and doesn't 1278 // return the new value. Just insert a scalar copy and defer 1279 // expanding it. 1280 NewElt = Builder.CreateBinOp(Opc, NumEltN, DenEltN); 1281 Div64ToExpand.push_back(cast<BinaryOperator>(NewElt)); 1282 } 1283 } 1284 1285 NewDiv = Builder.CreateInsertElement(NewDiv, NewElt, N); 1286 } 1287 } else { 1288 if (ScalarSize <= 32) 1289 NewDiv = expandDivRem32(Builder, I, Num, Den); 1290 else { 1291 NewDiv = shrinkDivRem64(Builder, I, Num, Den); 1292 if (!NewDiv) 1293 Div64ToExpand.push_back(&I); 1294 } 1295 } 1296 1297 if (NewDiv) { 1298 I.replaceAllUsesWith(NewDiv); 1299 I.eraseFromParent(); 1300 Changed = true; 1301 } 1302 } 1303 1304 if (ExpandDiv64InIR) { 1305 // TODO: We get much worse code in specially handled constant cases. 1306 for (BinaryOperator *Div : Div64ToExpand) { 1307 expandDivRem64(*Div); 1308 Changed = true; 1309 } 1310 } 1311 1312 return Changed; 1313 } 1314 1315 bool AMDGPUCodeGenPrepare::visitLoadInst(LoadInst &I) { 1316 if (!WidenLoads) 1317 return false; 1318 1319 if ((I.getPointerAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS || 1320 I.getPointerAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT) && 1321 canWidenScalarExtLoad(I)) { 1322 IRBuilder<> Builder(&I); 1323 Builder.SetCurrentDebugLocation(I.getDebugLoc()); 1324 1325 Type *I32Ty = Builder.getInt32Ty(); 1326 Type *PT = PointerType::get(I32Ty, I.getPointerAddressSpace()); 1327 Value *BitCast= Builder.CreateBitCast(I.getPointerOperand(), PT); 1328 LoadInst *WidenLoad = Builder.CreateLoad(I32Ty, BitCast); 1329 WidenLoad->copyMetadata(I); 1330 1331 // If we have range metadata, we need to convert the type, and not make 1332 // assumptions about the high bits. 1333 if (auto *Range = WidenLoad->getMetadata(LLVMContext::MD_range)) { 1334 ConstantInt *Lower = 1335 mdconst::extract<ConstantInt>(Range->getOperand(0)); 1336 1337 if (Lower->isNullValue()) { 1338 WidenLoad->setMetadata(LLVMContext::MD_range, nullptr); 1339 } else { 1340 Metadata *LowAndHigh[] = { 1341 ConstantAsMetadata::get(ConstantInt::get(I32Ty, Lower->getValue().zext(32))), 1342 // Don't make assumptions about the high bits. 1343 ConstantAsMetadata::get(ConstantInt::get(I32Ty, 0)) 1344 }; 1345 1346 WidenLoad->setMetadata(LLVMContext::MD_range, 1347 MDNode::get(Mod->getContext(), LowAndHigh)); 1348 } 1349 } 1350 1351 int TySize = Mod->getDataLayout().getTypeSizeInBits(I.getType()); 1352 Type *IntNTy = Builder.getIntNTy(TySize); 1353 Value *ValTrunc = Builder.CreateTrunc(WidenLoad, IntNTy); 1354 Value *ValOrig = Builder.CreateBitCast(ValTrunc, I.getType()); 1355 I.replaceAllUsesWith(ValOrig); 1356 I.eraseFromParent(); 1357 return true; 1358 } 1359 1360 return false; 1361 } 1362 1363 bool AMDGPUCodeGenPrepare::visitICmpInst(ICmpInst &I) { 1364 bool Changed = false; 1365 1366 if (ST->has16BitInsts() && needsPromotionToI32(I.getOperand(0)->getType()) && 1367 DA->isUniform(&I)) 1368 Changed |= promoteUniformOpToI32(I); 1369 1370 return Changed; 1371 } 1372 1373 bool AMDGPUCodeGenPrepare::visitSelectInst(SelectInst &I) { 1374 bool Changed = false; 1375 1376 if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) && 1377 DA->isUniform(&I)) 1378 Changed |= promoteUniformOpToI32(I); 1379 1380 return Changed; 1381 } 1382 1383 bool AMDGPUCodeGenPrepare::visitIntrinsicInst(IntrinsicInst &I) { 1384 switch (I.getIntrinsicID()) { 1385 case Intrinsic::bitreverse: 1386 return visitBitreverseIntrinsicInst(I); 1387 default: 1388 return false; 1389 } 1390 } 1391 1392 bool AMDGPUCodeGenPrepare::visitBitreverseIntrinsicInst(IntrinsicInst &I) { 1393 bool Changed = false; 1394 1395 if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) && 1396 DA->isUniform(&I)) 1397 Changed |= promoteUniformBitreverseToI32(I); 1398 1399 return Changed; 1400 } 1401 1402 bool AMDGPUCodeGenPrepare::doInitialization(Module &M) { 1403 Mod = &M; 1404 DL = &Mod->getDataLayout(); 1405 return false; 1406 } 1407 1408 bool AMDGPUCodeGenPrepare::runOnFunction(Function &F) { 1409 if (skipFunction(F)) 1410 return false; 1411 1412 auto *TPC = getAnalysisIfAvailable<TargetPassConfig>(); 1413 if (!TPC) 1414 return false; 1415 1416 const AMDGPUTargetMachine &TM = TPC->getTM<AMDGPUTargetMachine>(); 1417 ST = &TM.getSubtarget<GCNSubtarget>(F); 1418 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 1419 DA = &getAnalysis<LegacyDivergenceAnalysis>(); 1420 1421 auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>(); 1422 DT = DTWP ? &DTWP->getDomTree() : nullptr; 1423 1424 HasUnsafeFPMath = hasUnsafeFPMath(F); 1425 1426 AMDGPU::SIModeRegisterDefaults Mode(F); 1427 HasFP32Denormals = Mode.allFP32Denormals(); 1428 1429 bool MadeChange = false; 1430 1431 Function::iterator NextBB; 1432 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; FI = NextBB) { 1433 BasicBlock *BB = &*FI; 1434 NextBB = std::next(FI); 1435 1436 BasicBlock::iterator Next; 1437 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; I = Next) { 1438 Next = std::next(I); 1439 1440 MadeChange |= visit(*I); 1441 1442 if (Next != E) { // Control flow changed 1443 BasicBlock *NextInstBB = Next->getParent(); 1444 if (NextInstBB != BB) { 1445 BB = NextInstBB; 1446 E = BB->end(); 1447 FE = F.end(); 1448 } 1449 } 1450 } 1451 } 1452 1453 return MadeChange; 1454 } 1455 1456 INITIALIZE_PASS_BEGIN(AMDGPUCodeGenPrepare, DEBUG_TYPE, 1457 "AMDGPU IR optimizations", false, false) 1458 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 1459 INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis) 1460 INITIALIZE_PASS_END(AMDGPUCodeGenPrepare, DEBUG_TYPE, "AMDGPU IR optimizations", 1461 false, false) 1462 1463 char AMDGPUCodeGenPrepare::ID = 0; 1464 1465 FunctionPass *llvm::createAMDGPUCodeGenPreparePass() { 1466 return new AMDGPUCodeGenPrepare(); 1467 } 1468