[CostModel] Unify getArithmeticInstrCost

[samparker]

Summary

Add the remaining arithmetic opcodes into the generic implementation of getUserCost and then call this from getInstructionThroughput. Most of the backends have been modified to return the base implementation for cost kinds other RecipThroughput. The outlier here is AMDGPU which already uses getArithmeticInstrCost for all the cost kinds. This change means that most of the opcodes can be removed from that backends implementation of getUserCost.

[dfukalov]

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

429

Am I right that this reset and TCK_RecipThroughput override in the following call are needed to avoid tests failures? If so, would you please specify errors - I'll try to get a look.

[samparker]

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

429

No test failures actually, so shall I just remove this instead?

[samparker]

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

429

The RecipThroughput hack does cause a lot of failures, I was wondering whether hoisting some of the generic code here would be a better thing to do?

[samparker]

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

429

I tried hoisting from BasicTTI, but the v3 vector cost tests change like this:
-; FASTF64: estimated cost of 6 for {{.*}} fsub <3 x double>
-; SLOWF64: estimated cost of 9 for {{.*}} fsub <3 x double>
+; FASTF64: estimated cost of 8 for {{.*}} fsub <3 x double>
+; SLOWF64: estimated cost of 12 for {{.*}} fsub <3 x double>

[dfukalov]

AMDGPU part LGTM,
let's wait a couple of days to allow others to take a look

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

439

I see, this override is needed since scheme is changing. I guess I should review this place after you'll complete RFC

[uabelho]

Hi!

I wrote a PR about a crash that starts occuring with this patch:
https://bugs.llvm.org/show_bug.cgi?id=46276

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

public:
   unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
                                   TTI::TargetCostKind CostKind,
                                   TTI::OperandValueKind Opd1Info,
                                   TTI::OperandValueKind Opd2Info,
                                   TTI::OperandValueProperties Opd1PropInfo,
                                   TTI::OperandValueProperties Opd2PropInfo,
                                   ArrayRef<const Value *> Args,
                                   const Instruction *CxtI = nullptr) {
+    // FIXME: A number of transformation tests seem to require these values
+    // which seems a little odd for how arbitary there are.
+    switch (Opcode) {
+    default:
+      break;
+    case Instruction::FDiv:
+    case Instruction::FRem:
+    case Instruction::SDiv:
+    case Instruction::SRem:
+    case Instruction::UDiv:
+    case Instruction::URem:
+      // FIXME: Unlikely to be true for CodeSize.
+      return TTI::TCC_Expensive;
+    }
     return 1;
   }
 
   unsigned getShuffleCost(TTI::ShuffleKind Kind, VectorType *Ty, int Index,
                           VectorType *SubTp) {
     return 1;
   }
 
case Instruction::Alloca:
         return TTI::TCC_Free;
       break;
     case Instruction::GetElementPtr: {
       const GEPOperator *GEP = cast<GEPOperator>(U);
       return TargetTTI->getGEPCost(GEP->getSourceElementType(),
                                    GEP->getPointerOperand(),
                                    Operands.drop_front());
     }
-    case Instruction::FDiv:
-    case Instruction::FRem:
-    case Instruction::SDiv:
-    case Instruction::SRem:
+    case Instruction::Add:
+    case Instruction::FAdd:
+    case Instruction::Sub:
+    case Instruction::FSub:
+    case Instruction::Mul:
+    case Instruction::FMul:
     case Instruction::UDiv:
+    case Instruction::SDiv:
+    case Instruction::FDiv:
     case Instruction::URem:
-      // FIXME: Unlikely to be true for CodeSize.
-      return TTI::TCC_Expensive;
+    case Instruction::SRem:
+    case Instruction::FRem:
+    case Instruction::Shl:
+    case Instruction::LShr:
+    case Instruction::AShr:
+    case Instruction::And:
+    case Instruction::Or:
+    case Instruction::Xor: {
+      TargetTransformInfo::OperandValueKind Op1VK, Op2VK;
+      TargetTransformInfo::OperandValueProperties Op1VP, Op2VP;
+      Op1VK = TTI::getOperandInfo(U->getOperand(0), Op1VP);
+      Op2VK = TTI::getOperandInfo(U->getOperand(1), Op2VP);
+      SmallVector<const Value *, 2> Operands(U->operand_values());
+      return TargetTTI->getArithmeticInstrCost(Opcode, Ty, CostKind,
+                                               Op1VK, Op2VK,
+                                               Op1VP, Op2VP, Operands, I);
+    }
     case Instruction::IntToPtr:
     case Instruction::PtrToInt:
     case Instruction::SIToFP:
     case Instruction::UIToFP:
     case Instruction::FPToUI:
     case Instruction::FPToSI:
     case Instruction::Trunc:
     case Instruction::FPTrunc:

llvm/include/llvm/CodeGen/BasicTTIImpl.h

unsigned getArithmeticInstrCost(
       TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
       ArrayRef<const Value *> Args = ArrayRef<const Value *>(),
       const Instruction *CxtI = nullptr) {
     // Check if any of the operands are vector operands.
     const TargetLoweringBase *TLI = getTLI();
     int ISD = TLI->InstructionOpcodeToISD(Opcode);
     assert(ISD && "Invalid opcode");
 
+    // TODO: Handle more cost kinds.
+    if (CostKind != TTI::TCK_RecipThroughput)
+      return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind,
+                                           Opd1Info, Opd2Info,
+                                           Opd1PropInfo, Opd2PropInfo,
+                                           Args, CxtI);
+
     std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
 
     bool IsFloat = Ty->isFPOrFPVectorTy();
     // Assume that floating point arithmetic operations cost twice as much as
     // integer operations.
     unsigned OpCost = (IsFloat ? 2 : 1);
 
     if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {

llvm/lib/Analysis/TargetTransformInfo.cpp

   case Instruction::URem:
   case Instruction::SRem:
   case Instruction::FRem:
   case Instruction::Shl:
   case Instruction::LShr:
   case Instruction::AShr:
   case Instruction::And:
   case Instruction::Or:
-  case Instruction::Xor: {
-    TargetTransformInfo::OperandValueKind Op1VK, Op2VK;
-    TargetTransformInfo::OperandValueProperties Op1VP, Op2VP;
-    Op1VK = getOperandInfo(I->getOperand(0), Op1VP);
-    Op2VK = getOperandInfo(I->getOperand(1), Op2VP);
-    SmallVector<const Value *, 2> Operands(I->operand_values());
-    return getArithmeticInstrCost(I->getOpcode(), I->getType(), CostKind,
-                                  Op1VK, Op2VK,
-                                  Op1VP, Op2VP, Operands, I);
-  }
+  case Instruction::Xor:
+    return getUserCost(I, CostKind);
   case Instruction::FNeg: {
     TargetTransformInfo::OperandValueKind Op1VK, Op2VK;
     TargetTransformInfo::OperandValueProperties Op1VP, Op2VP;
     Op1VK = getOperandInfo(I->getOperand(0), Op1VP);
     Op2VK = OK_AnyValue;
     Op2VP = OP_None;
     SmallVector<const Value *, 2> Operands(I->operand_values());
     return getArithmeticInstrCost(I->getOpcode(), I->getType(), CostKind,

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

 }
 
 int AArch64TTIImpl::getArithmeticInstrCost(
     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
     TTI::OperandValueKind Opd1Info,
     TTI::OperandValueKind Opd2Info, TTI::OperandValueProperties Opd1PropInfo,
     TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
     const Instruction *CxtI) {
+  // TODO: Handle more cost kinds.
+  if (CostKind != TTI::TCK_RecipThroughput)
+    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info,
+                                         Opd2Info, Opd1PropInfo,
+                                         Opd2PropInfo, Args, CxtI);
+
   // Legalize the type.
   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
 
   // If the instruction is a widening instruction (e.g., uaddl, saddw, etc.),
   // add in the widening overhead specified by the sub-target. Since the
   // extends feeding widening instructions are performed automatically, they
   // aren't present in the generated code and have a zero cost. By adding a
   // widening overhead here, we attach the total cost of the combined operation

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

int GCNTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
                                        TTI::TargetCostKind CostKind,
                                        TTI::OperandValueKind Opd1Info,
                                        TTI::OperandValueKind Opd2Info,
                                        TTI::OperandValueProperties Opd1PropInfo,
                                        TTI::OperandValueProperties Opd2PropInfo,
                                        ArrayRef<const Value *> Args,
                                        const Instruction *CxtI) {
   EVT OrigTy = TLI->getValueType(DL, Ty);
+  // TODO: Is this reset of these values necessary?
+  if (CostKind != TTI::TCK_RecipThroughput) {
+    Opd1Info = TTI::OK_AnyValue;
+    Opd2Info = TTI::OK_AnyValue;
+    Opd1PropInfo = TTI::OP_None;
+    Opd2PropInfo = TTI::OP_None;
+  }
   if (!OrigTy.isSimple()) {
-    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info,
-                                         Opd2Info,
+    // FIXME: We're having to query the throughput cost so that the basic
+    // implementation tries to generate legalize and scalarization costs.
+    return BaseT::getArithmeticInstrCost(Opcode, Ty, TTI::TCK_RecipThroughput,
+                                         Opd1Info, Opd2Info,
                                          Opd1PropInfo, Opd2PropInfo);
   }
 
   // Legalize the type.
   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
   int ISD = TLI->InstructionOpcodeToISD(Opcode);
 
   // Because we don't have any legal vector operations, but the legal types, we
GCNTTIImpl::getUserCost(const User *U, ArrayRef<const Value *> Operands,
   case Instruction::UIToFP:
   case Instruction::Trunc:
   case Instruction::FPTrunc:
   case Instruction::BitCast:
   case Instruction::AddrSpaceCast: {
     return getCastInstrCost(I->getOpcode(), I->getType(),
                             I->getOperand(0)->getType(), CostKind, I);
   }
-  case Instruction::Add:
-  case Instruction::FAdd:
-  case Instruction::Sub:
-  case Instruction::FSub:
-  case Instruction::Mul:
-  case Instruction::FMul:
-  case Instruction::UDiv:
-  case Instruction::SDiv:
-  case Instruction::FDiv:
-  case Instruction::URem:
-  case Instruction::SRem:
-  case Instruction::FRem:
-  case Instruction::Shl:
-  case Instruction::LShr:
-  case Instruction::AShr:
-  case Instruction::And:
-  case Instruction::Or:
-  case Instruction::Xor:
-  case Instruction::FNeg: {
+  case Instruction::FNeg:
     return getArithmeticInstrCost(I->getOpcode(), I->getType(), CostKind,
                                   TTI::OK_AnyValue, TTI::OK_AnyValue,
                                   TTI::OP_None, TTI::OP_None, Operands, I);
-  }
   default:
     break;
   }
 
   return BaseT::getUserCost(U, Operands, CostKind);
 }
 
 unsigned R600TTIImpl::getHardwareNumberOfRegisters(bool Vec) const {

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

 int ARMTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
                                        TTI::TargetCostKind CostKind,
                                        TTI::OperandValueKind Op1Info,
                                        TTI::OperandValueKind Op2Info,
                                        TTI::OperandValueProperties Opd1PropInfo,
                                        TTI::OperandValueProperties Opd2PropInfo,
                                        ArrayRef<const Value *> Args,
                                        const Instruction *CxtI) {
+  // TODO: Handle more cost kinds.
+  if (CostKind != TTI::TCK_RecipThroughput)
+    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
+                                         Op2Info, Opd1PropInfo,
+                                         Opd2PropInfo, Args, CxtI);
+
   int ISDOpcode = TLI->InstructionOpcodeToISD(Opcode);
   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
 
   if (ST->hasNEON()) {
     const unsigned FunctionCallDivCost = 20;
     const unsigned ReciprocalDivCost = 10;
     static const CostTblEntry CostTbl[] = {
       // Division.

llvm/lib/Target/Hexagon/HexagonTargetTransformInfo.cpp

 }
 
 unsigned HexagonTTIImpl::getArithmeticInstrCost(
     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
     TTI::OperandValueKind Opd1Info,
     TTI::OperandValueKind Opd2Info, TTI::OperandValueProperties Opd1PropInfo,
     TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
     const Instruction *CxtI) {
+  // TODO: Handle more cost kinds.
+  if (CostKind != TTI::TCK_RecipThroughput)
+    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info,
+                                         Opd2Info, Opd1PropInfo,
+                                         Opd2PropInfo, Args, CxtI);
+
   if (Ty->isVectorTy()) {
     std::pair<int, MVT> LT = TLI.getTypeLegalizationCost(DL, Ty);
     if (LT.second.isFloatingPoint())
       return LT.first + FloatFactor * getTypeNumElements(Ty);
   }
   return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info, Opd2Info,
                                        Opd1PropInfo, Opd2PropInfo, Args, CxtI);
 }

llvm/lib/Target/PowerPC/PPCTargetTransformInfo.cpp

int PPCTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
                                        TTI::TargetCostKind CostKind,
                                        TTI::OperandValueKind Op1Info,
                                        TTI::OperandValueKind Op2Info,
                                        TTI::OperandValueProperties Opd1PropInfo,
                                        TTI::OperandValueProperties Opd2PropInfo,
                                        ArrayRef<const Value *> Args,
                                        const Instruction *CxtI) {
   assert(TLI->InstructionOpcodeToISD(Opcode) && "Invalid opcode");
+  // TODO: Handle more cost kinds.
+  if (CostKind != TTI::TCK_RecipThroughput)
+    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
+                                         Op2Info, Opd1PropInfo,
+                                         Opd2PropInfo, Args, CxtI);
 
   // Fallback to the default implementation.
   int Cost = BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
                                            Op2Info,
                                            Opd1PropInfo, Opd2PropInfo);
   return vectorCostAdjustment(Cost, Opcode, Ty, nullptr);
 }
 

llvm/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp

 
 int SystemZTTIImpl::getArithmeticInstrCost(
     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
     TTI::OperandValueKind Op1Info,
     TTI::OperandValueKind Op2Info, TTI::OperandValueProperties Opd1PropInfo,
     TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
     const Instruction *CxtI) {
 
+  // TODO: Handle more cost kinds.
+  if (CostKind != TTI::TCK_RecipThroughput)
+    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
+                                         Op2Info, Opd1PropInfo,
+                                         Opd2PropInfo, Args, CxtI);
+
   // TODO: return a good value for BB-VECTORIZER that includes the
   // immediate loads, which we do not want to count for the loop
   // vectorizer, since they are hopefully hoisted out of the loop. This
   // would require a new parameter 'InLoop', but not sure if constant
   // args are common enough to motivate this.
 
   unsigned ScalarBits = Ty->getScalarSizeInBits();
 

llvm/lib/Target/X86/X86TargetTransformInfo.cpp

 int X86TTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
                                        TTI::TargetCostKind CostKind,
                                        TTI::OperandValueKind Op1Info,
                                        TTI::OperandValueKind Op2Info,
                                        TTI::OperandValueProperties Opd1PropInfo,
                                        TTI::OperandValueProperties Opd2PropInfo,
                                        ArrayRef<const Value *> Args,
                                        const Instruction *CxtI) {
+  // TODO: Handle more cost kinds.
+  if (CostKind != TTI::TCK_RecipThroughput)
+    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
+                                         Op2Info, Opd1PropInfo,
+                                         Opd2PropInfo, Args, CxtI);
   // Legalize the type.
   std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
 
   int ISD = TLI->InstructionOpcodeToISD(Opcode);
   assert(ISD && "Invalid opcode");
 
   static const CostTblEntry GLMCostTable[] = {
     { ISD::FDIV,  MVT::f32,   18 }, // divss