[Inliner] Restrict soft-float inlining penalty.

[efriedma]

Summary

The penalty is currently getting applied in a bunch of places where it doesn't make sense, like bitcasts (which are free) and calls (which were getting the call penalty applied twice). Instead, just apply the penalty to binary operators and floating-point casts.

While I'm here, also fix getFPOpCost() to do the right thing in more cases, so we don't have to dig into function attributes.

(Not sure if I should also apply this to fcmp instructions.)

[echristo]

lib/Analysis/InlineCost.cpp

1102

I realize you're just moving this code, but we might want to make this a TTI backend call for the function since there's also the subtarget feature on some targets as well.

Would it make since for this to be a part of getFPOpCost instead?

[efriedma]

Fixed getFPOpCost().

[efriedma]

lib/Analysis/InlineCost.cpp

1102

I think this code was originally written before we had per-function TTI... but we should take advantage of it now. Fixed the default implementation of getFPOpCost to do the right thing, and removed the explicit check for "use-soft-float".

[echristo]

LGTM, thanks!

include/llvm/CodeGen/BasicTTIImpl.h

return TLI->isTypeLegal(VT) &&
            TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
   }
 
   bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
     return true;
   }
 
   unsigned getFPOpCost(Type *Ty) {
-    // By default, FP instructions are no more expensive since they are
-    // implemented in HW.  Target specific TTI can override this.
+    // Check whether FADD is available, as a proxy for floating-point in
+    // general.
+    const TargetLoweringBase *TLI = getTLI();
+    EVT VT = TLI->getValueType(DL, Ty);
+    if (TLI->isOperationLegalOrCustomOrPromote(ISD::FADD, VT))
       return TargetTransformInfo::TCC_Basic;
+    return TargetTransformInfo::TCC_Expensive;
   }
 
   unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
     const TargetLoweringBase *TLI = getTLI();
     switch (Opcode) {
     default: break;
     case Instruction::Trunc:
       if (TLI->isTruncateFree(OpTy, Ty))

lib/Analysis/InlineCost.cpp

bool CallAnalyzer::visitCastInst(CastInst &I) {
   if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
         return ConstantExpr::getCast(I.getOpcode(), COps[0], I.getType());
       }))
     return true;
 
   // Disable SROA in the face of arbitrary casts we don't whitelist elsewhere.
   disableSROA(I.getOperand(0));
 
+  // If this is a floating-point cast, and the target says this operation
+  // is expensive, this may eventually become a library call. Treat the cost
+  // as such.
+  switch (I.getOpcode()) {
+  case Instruction::FPTrunc:
+  case Instruction::FPExt:
+  case Instruction::UIToFP:
+  case Instruction::SIToFP:
+  case Instruction::FPToUI:
+  case Instruction::FPToSI:
+    if (TTI.getFPOpCost(I.getType()) == TargetTransformInfo::TCC_Expensive)
+      Cost += InlineConstants::CallPenalty;
+  default:
+    break;
+  }
+
   return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
 }
 
 bool CallAnalyzer::visitUnaryInstruction(UnaryInstruction &I) {
   Value *Operand = I.getOperand(0);
   if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
         return ConstantFoldInstOperands(&I, COps[0], DL);
       }))
bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) {
 
   if (simplifyInstruction(I, Evaluate))
     return true;
 
   // Disable any SROA on arguments to arbitrary, unsimplified binary operators.
   disableSROA(LHS);
   disableSROA(RHS);
 
+  // If the instruction is floating point, and the target says this operation
+  // is expensive, this may eventually become a library call. Treat the cost
+  // as such.
+  if (I.getType()->isFloatingPointTy() &&
+      TTI.getFPOpCost(I.getType()) == TargetTransformInfo::TCC_Expensive)
+    Cost += InlineConstants::CallPenalty;
+
   return false;
 }
 
 bool CallAnalyzer::visitLoad(LoadInst &I) {
   Value *SROAArg;
   DenseMap<Value *, int>::iterator CostIt;
   if (lookupSROAArgAndCost(I.getPointerOperand(), SROAArg, CostIt)) {
     if (I.isSimple()) {
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
     // Skip ephemeral values.
     if (EphValues.count(&*I))
       continue;
 
     ++NumInstructions;
     if (isa<ExtractElementInst>(I) || I->getType()->isVectorTy())
       ++NumVectorInstructions;
 
-    // If the instruction is floating point, and the target says this operation
-    // is expensive or the function has the "use-soft-float" attribute, this may
-    // eventually become a library call. Treat the cost as such.
-    if (I->getType()->isFloatingPointTy()) {
-      // If the function has the "use-soft-float" attribute, mark it as
-      // expensive.
-      if (TTI.getFPOpCost(I->getType()) == TargetTransformInfo::TCC_Expensive ||
-          (F.getFnAttribute("use-soft-float").getValueAsString() == "true"))
-        Cost += InlineConstants::CallPenalty;
-    }
-
     // If the instruction simplified to a constant, there is no cost to this
     // instruction. Visit the instructions using our InstVisitor to account for
     // all of the per-instruction logic. The visit tree returns true if we
     // consumed the instruction in any way, and false if the instruction's base
     // cost should count against inlining.
     if (Base::visit(&*I))
       ++NumInstructionsSimplified;
     else

lib/Target/ARM/ARMTargetTransformInfo.h

public:
   int getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
                          const Instruction *I = nullptr);
 
   int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index);
 
   int getAddressComputationCost(Type *Val, ScalarEvolution *SE, 
                                 const SCEV *Ptr);
 
-  int getFPOpCost(Type *Ty);
-
   int getArithmeticInstrCost(
       unsigned Opcode, Type *Ty,
       TTI::OperandValueKind Op1Info = TTI::OK_AnyValue,
       TTI::OperandValueKind Op2Info = TTI::OK_AnyValue,
       TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
       TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
       ArrayRef<const Value *> Args = ArrayRef<const Value *>());
 

lib/Target/ARM/ARMTargetTransformInfo.cpp

if (Ty->isVectorTy() && SE &&
       !BaseT::isConstantStridedAccessLessThan(SE, Ptr, MaxMergeDistance + 1))
     return NumVectorInstToHideOverhead;
 
   // In many cases the address computation is not merged into the instruction
   // addressing mode.
   return 1;
 }
 
-int ARMTTIImpl::getFPOpCost(Type *Ty) {
-  // Use similar logic that's in ARMISelLowering:
-  // Any ARM CPU with VFP2 has floating point, but Thumb1 didn't have access
-  // to VFP.
-
-  if (ST->hasVFP2() && !ST->isThumb1Only()) {
-    if (Ty->isFloatTy()) {
-      return TargetTransformInfo::TCC_Basic;
-    }
-
-    if (Ty->isDoubleTy()) {
-      return ST->isFPOnlySP() ? TargetTransformInfo::TCC_Expensive :
-        TargetTransformInfo::TCC_Basic;
-    }
-  }
-
-  return TargetTransformInfo::TCC_Expensive;
-}
-
 int ARMTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
                                Type *SubTp) {
   // We only handle costs of reverse and alternate shuffles for now.
   if (Kind != TTI::SK_Reverse && Kind != TTI::SK_Alternate)
     return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
 
   if (Kind == TTI::SK_Reverse) {
     static const CostTblEntry NEONShuffleTbl[] = {

test/Transforms/Inline/ARM/inline-fp.ll

+; RUN: opt -S -inline -mtriple=arm-eabi -pass-remarks=.* -pass-remarks-missed=.* < %s 2>&1 | FileCheck %s -check-prefix=NOFP
+; RUN: opt -S -inline -mtriple=arm-eabi -mattr=+vfp2 -pass-remarks=.* -pass-remarks-missed=.* < %s 2>&1 | FileCheck %s -check-prefix=FULLFP
+; RUN: opt -S -inline -mtriple=arm-eabi -mattr=+vfp2,+fp-only-sp -pass-remarks=.* -pass-remarks-missed=.* < %s 2>&1 | FileCheck %s -check-prefix=SINGLEFP
+; Make sure that soft float implementations are calculated as being more expensive
+; to the inliner.
+
+; NOFP-DAG: single not inlined into test_single because too costly to inline (cost=125, threshold=75)
+; NOFP-DAG: single not inlined into test_single because too costly to inline (cost=125, threshold=75)
+; NOFP-DAG: single_cheap inlined into test_single_cheap with cost=-15 (threshold=75)
+; NOFP-DAG: single_cheap inlined into test_single_cheap with cost=-15015 (threshold=75)
+; NOFP-DAG: double not inlined into test_double because too costly to inline (cost=125, threshold=75)
+; NOFP-DAG: double not inlined into test_double because too costly to inline (cost=125, threshold=75)
+; NOFP-DAG: single_force_soft not inlined into test_single_force_soft because too costly to inline (cost=125, threshold=75)
+; NOFP-DAG: single_force_soft not inlined into test_single_force_soft because too costly to inline (cost=125, threshold=75)
+
+; FULLFP-DAG: single inlined into test_single with cost=0 (threshold=75)
+; FULLFP-DAG: single inlined into test_single with cost=-15000 (threshold=75)
+; FULLFP-DAG: single_cheap inlined into test_single_cheap with cost=-15 (threshold=75)
+; FULLFP-DAG: single_cheap inlined into test_single_cheap with cost=-15015 (threshold=75)
+; FULLFP-DAG: double inlined into test_double with cost=0 (threshold=75)
+; FULLFP-DAG: double inlined into test_double with cost=-15000 (threshold=75)
+; FULLFP-DAG: single_force_soft not inlined into test_single_force_soft because too costly to inline (cost=125, threshold=75)
+; FULLFP-DAG: single_force_soft not inlined into test_single_force_soft because too costly to inline (cost=125, threshold=75)
+
+; SINGLEFP-DAG: single inlined into test_single with cost=0 (threshold=75)
+; SINGLEFP-DAG: single inlined into test_single with cost=-15000 (threshold=75)
+; SINGLEFP-DAG: single_cheap inlined into test_single_cheap with cost=-15 (threshold=75)
+; SINGLEFP-DAG: single_cheap inlined into test_single_cheap with cost=-15015 (threshold=75)
+; SINGLEFP-DAG: double not inlined into test_double because too costly to inline (cost=125, threshold=75)
+; SINGLEFP-DAG: double not inlined into test_double because too costly to inline (cost=125, threshold=75)
+; SINGLEFP-DAG: single_force_soft not inlined into test_single_force_soft because too costly to inline (cost=125, threshold=75)
+; SINGLEFP-DAG: single_force_soft not inlined into test_single_force_soft because too costly to inline (cost=125, threshold=75)
+
+define i32 @test_single(i32 %a, i8 %b, i32 %c, i8 %d) #0 {
+  %call = call float @single(i32 %a, i8 zeroext %b)
+  %call2 = call float @single(i32 %c, i8 zeroext %d)
+  ret i32 0
+}
+
+define i32 @test_single_cheap(i32 %a, i8 %b, i32 %c, i8 %d) #0 {
+  %call = call float @single_cheap(i32 %a, i8 zeroext %b)
+  %call2 = call float @single_cheap(i32 %c, i8 zeroext %d)
+  ret i32 0
+}
+
+define i32 @test_double(i32 %a, i8 %b, i32 %c, i8 %d) #0 {
+  %call = call double @double(i32 %a, i8 zeroext %b)
+  %call2 = call double @double(i32 %c, i8 zeroext %d)
+  ret i32 0
+}
+
+define i32 @test_single_force_soft(i32 %a, i8 %b, i32 %c, i8 %d) #1 {
+  %call = call float @single_force_soft(i32 %a, i8 zeroext %b) #1
+  %call2 = call float @single_force_soft(i32 %c, i8 zeroext %d) #1
+  ret i32 0
+}
+
+define internal float @single(i32 %response, i8 zeroext %value1) #0 {
+entry:
+  %conv = zext i8 %value1 to i32
+  %sub = add nsw i32 %conv, -1
+  %conv1 = sitofp i32 %sub to float
+  %0 = tail call float @llvm.pow.f32(float 0x3FF028F5C0000000, float %conv1)
+  %mul = fmul float %0, 2.620000e+03
+  %conv2 = sitofp i32 %response to float
+  %sub3 = fsub float %conv2, %mul
+  %div = fdiv float %sub3, %mul
+  ret float %div
+}
+
+define internal float @single_cheap(i32 %response, i8 zeroext %value1) #0 {
+entry:
+  %conv = zext i8 %value1 to i32
+  %sub = add nsw i32 %conv, -1
+  %conv1 = bitcast i32 %sub to float
+  %conv2 = bitcast i32 %response to float
+  %0 = tail call float @llvm.pow.f32(float %conv2, float %conv1)
+  %1 = tail call float @llvm.pow.f32(float %0, float %0)
+  %2 = tail call float @llvm.pow.f32(float %1, float %1)
+  ret float %2
+}
+
+define internal double @double(i32 %response, i8 zeroext %value1) #0 {
+entry:
+  %conv = zext i8 %value1 to i32
+  %sub = add nsw i32 %conv, -1
+  %conv1 = sitofp i32 %sub to double
+  %0 = tail call double @llvm.pow.f64(double 0x3FF028F5C0000000, double %conv1)
+  %mul = fmul double %0, 2.620000e+03
+  %conv2 = sitofp i32 %response to double
+  %sub3 = fsub double %conv2, %mul
+  %div = fdiv double %sub3, %mul
+  ret double %div
+}
+
+define internal float @single_force_soft(i32 %response, i8 zeroext %value1) #1 {
+entry:
+  %conv = zext i8 %value1 to i32
+  %sub = add nsw i32 %conv, -1
+  %conv1 = sitofp i32 %sub to float
+  %0 = tail call float @llvm.pow.f32(float 0x3FF028F5C0000000, float %conv1)
+  %mul = fmul float %0, 2.620000e+03
+  %conv2 = sitofp i32 %response to float
+  %sub3 = fsub float %conv2, %mul
+  %div = fdiv float %sub3, %mul
+  ret float %div
+}
+
+declare float @llvm.pow.f32(float, float) optsize minsize
+declare double @llvm.pow.f64(double, double) optsize minsize
+
+attributes #0 = { optsize }
+attributes #1 = { optsize "use-soft-float"="true" "target-features"="+soft-float" }

test/Transforms/Inline/inline-fp.ll

-; RUN: opt -S -inline < %s | FileCheck %s
-; RUN: opt -S -passes='cgscc(inline)' < %s | FileCheck %s
-; Make sure that soft float implementations are calculated as being more expensive
-; to the inliner.
-
-define i32 @test_nofp() #0 {
-; f_nofp() has the "use-soft-float" attribute, so it should never get inlined.
-; CHECK-LABEL: test_nofp
-; CHECK: call float @f_nofp 
-entry:
-  %responseX = alloca i32, align 4
-  %responseY = alloca i32, align 4
-  %responseZ = alloca i32, align 4
-  %valueX = alloca i8, align 1
-  %valueY = alloca i8, align 1
-  %valueZ = alloca i8, align 1
-
-  call void @getX(i32* %responseX, i8* %valueX)
-  call void @getY(i32* %responseY, i8* %valueY)
-  call void @getZ(i32* %responseZ, i8* %valueZ)
-
-  %0 = load i32, i32* %responseX
-  %1 = load i8, i8* %valueX
-  %call = call float @f_nofp(i32 %0, i8 zeroext %1)
-  %2 = load i32, i32* %responseZ
-  %3 = load i8, i8* %valueZ
-  %call2 = call float @f_nofp(i32 %2, i8 zeroext %3)
-  %call3 = call float @fabsf(float %call)
-  %cmp = fcmp ogt float %call3, 0x3FC1EB8520000000
-  br i1 %cmp, label %if.end12, label %if.else
-
-if.else:                                          ; preds = %entry
-  %4 = load i32, i32* %responseY
-  %5 = load i8, i8* %valueY
-  %call1 = call float @f_nofp(i32 %4, i8 zeroext %5)
-  %call4 = call float @fabsf(float %call1)
-  %cmp5 = fcmp ogt float %call4, 0x3FC1EB8520000000
-  br i1 %cmp5, label %if.end12, label %if.else7
-
-if.else7:                                         ; preds = %if.else
-  %call8 = call float @fabsf(float %call2)
-  %cmp9 = fcmp ogt float %call8, 0x3FC1EB8520000000
-  br i1 %cmp9, label %if.then10, label %if.end12
-
-if.then10:                                        ; preds = %if.else7
-  br label %if.end12
-
-if.end12:                                         ; preds = %if.else, %entry, %if.then10, %if.else7
-  %success.0 = phi i32 [ 0, %if.then10 ], [ 1, %if.else7 ], [ 0, %entry ], [ 0, %if.else ]
-  ret i32 %success.0
-}
-
-define i32 @test_hasfp() #0 {
-; f_hasfp()  does not have the "use-soft-float" attribute, so it should get inlined.
-; CHECK-LABEL: test_hasfp
-; CHECK-NOT: call float @f_hasfp 
-entry:
-  %responseX = alloca i32, align 4
-  %responseY = alloca i32, align 4
-  %responseZ = alloca i32, align 4
-  %valueX = alloca i8, align 1
-  %valueY = alloca i8, align 1
-  %valueZ = alloca i8, align 1
-
-  call void @getX(i32* %responseX, i8* %valueX)
-  call void @getY(i32* %responseY, i8* %valueY)
-  call void @getZ(i32* %responseZ, i8* %valueZ)
-
-  %0 = load i32, i32* %responseX
-  %1 = load i8, i8* %valueX
-  %call = call float @f_hasfp(i32 %0, i8 zeroext %1)
-  %2 = load i32, i32* %responseZ
-  %3 = load i8, i8* %valueZ
-  %call2 = call float @f_hasfp(i32 %2, i8 zeroext %3)
-  %call3 = call float @fabsf(float %call)
-  %cmp = fcmp ogt float %call3, 0x3FC1EB8520000000
-  br i1 %cmp, label %if.end12, label %if.else
-
-if.else:                                          ; preds = %entry
-  %4 = load i32, i32* %responseY
-  %5 = load i8, i8* %valueY
-  %call1 = call float @f_hasfp(i32 %4, i8 zeroext %5)
-  %call4 = call float @fabsf(float %call1)
-  %cmp5 = fcmp ogt float %call4, 0x3FC1EB8520000000
-  br i1 %cmp5, label %if.end12, label %if.else7
-
-if.else7:                                         ; preds = %if.else
-  %call8 = call float @fabsf(float %call2)
-  %cmp9 = fcmp ogt float %call8, 0x3FC1EB8520000000
-  br i1 %cmp9, label %if.then10, label %if.end12
-
-if.then10:                                        ; preds = %if.else7
-  br label %if.end12
-
-if.end12:                                         ; preds = %if.else, %entry, %if.then10, %if.else7
-  %success.0 = phi i32 [ 0, %if.then10 ], [ 1, %if.else7 ], [ 0, %entry ], [ 0, %if.else ]
-  ret i32 %success.0
-}
-
-declare void @getX(i32*, i8*) #0
-
-declare void @getY(i32*, i8*) #0
-
-declare void @getZ(i32*, i8*) #0
-
-define internal float @f_hasfp(i32 %response, i8 zeroext %value1) #0 {
-entry:
-  %conv = zext i8 %value1 to i32
-  %sub = add nsw i32 %conv, -1
-  %conv1 = sitofp i32 %sub to float
-  %0 = tail call float @llvm.pow.f32(float 0x3FF028F5C0000000, float %conv1)
-  %mul = fmul float %0, 2.620000e+03
-  %conv2 = sitofp i32 %response to float
-  %sub3 = fsub float %conv2, %mul
-  %div = fdiv float %sub3, %mul
-  ret float %div
-}
-
-define internal float @f_nofp(i32 %response, i8 zeroext %value1) #1 {
-entry:
-  %conv = zext i8 %value1 to i32
-  %sub = add nsw i32 %conv, -1
-  %conv1 = sitofp i32 %sub to float
-  %0 = tail call float @llvm.pow.f32(float 0x3FF028F5C0000000, float %conv1)
-  %mul = fmul float %0, 2.620000e+03
-  %conv2 = sitofp i32 %response to float
-  %sub3 = fsub float %conv2, %mul
-  %div = fdiv float %sub3, %mul
-  ret float %div
-}
-
-declare float @fabsf(float) optsize minsize
-
-declare float @llvm.pow.f32(float, float) optsize minsize
-
-attributes #0 = { optsize }
-attributes #1 = { optsize "use-soft-float"="true" }