forked from OSchip/llvm-project
Teach InstCombine to hoist FABS and FNEG through FPTRUNC instructions. The application of these operations commutes with the truncation, so we should prefer to do them in the smallest size we can, to save register space, use smaller constant pool entries, etc.
llvm-svn: 172117
This commit is contained in:
parent
929cf5615f
commit
dbf0ca523d
|
@ -1204,8 +1204,34 @@ Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
|
|||
}
|
||||
break;
|
||||
}
|
||||
|
||||
// (fptrunc (fneg x)) -> (fneg (fptrunc x))
|
||||
if (BinaryOperator::isFNeg(OpI)) {
|
||||
Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
|
||||
CI.getType());
|
||||
return BinaryOperator::CreateFNeg(InnerTrunc);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
|
||||
if (II) {
|
||||
switch (II->getIntrinsicID()) {
|
||||
default: break;
|
||||
case Intrinsic::fabs: {
|
||||
// (fptrunc (fabs x)) -> (fabs (fptrunc x))
|
||||
Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
|
||||
CI.getType());
|
||||
Type *IntrinsicType[] = { CI.getType() };
|
||||
Function *Overload =
|
||||
Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(),
|
||||
II->getIntrinsicID(), IntrinsicType);
|
||||
|
||||
Value *Args[] = { InnerTrunc };
|
||||
return CallInst::Create(Overload, Args, II->getName());
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
|
||||
CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
|
||||
if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) &&
|
||||
|
|
|
@ -13,3 +13,22 @@ define i8 @test2() {
|
|||
; CHECK: ret i8 -1
|
||||
}
|
||||
|
||||
; CHECK: test3
|
||||
define half @test3(float %a) {
|
||||
; CHECK: fptrunc
|
||||
; CHECK: llvm.fabs.f16
|
||||
%b = call float @llvm.fabs.f32(float %a)
|
||||
%c = fptrunc float %b to half
|
||||
ret half %c
|
||||
}
|
||||
|
||||
; CHECK: test4
|
||||
define half @test4(float %a) {
|
||||
; CHECK: fptrunc
|
||||
; CHECK: fsub
|
||||
%b = fsub float -0.0, %a
|
||||
%c = fptrunc float %b to half
|
||||
ret half %c
|
||||
}
|
||||
|
||||
declare float @llvm.fabs.f32(float) nounwind readonly
|
||||
|
|
Loading…
Reference in New Issue