llvm dev - Feb 2021 - InstCombine makes assumptions on math functions

Hello,

I see InstCombine makes assumptions on functions named with C math function
names.

For example it will optimize the following code:

target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define dso_local double @fabs(double %x) {
  %res = fmul double %x, %x
  ret double %res
}

define dso_local float @_Z3foof(float %x) {
entry:
  %conv = fpext float %x to double
  %0 = call double @fabs(double %conv)
  %conv1 = fptrunc double %0 to float
  ret float %conv1
}

Into the following:

target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define dso_local float @_Z3foof(float %x) {
entry:
  %conv1 = call float @llvm.fabs.f32(float %x)
  ret float %conv1
}

; Function Attrs: nounwind readnone speculatable
declare double @llvm.fabs.f64(double) #0

; Function Attrs: nounwind readnone speculatable
declare float @llvm.fabs.f32(float) #0

attributes #0 = { nounwind readnone speculatable }

It seems to have assumed my @fabs function was the @llvm.fabs.f64 intrinsic.

Is that correct behavior?

-- 
*Alexandre Isoard*
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210225/effbc66a/attachment.html>

Correct, that can be debated. Expected, yes.
To avoid this, use `nobuiltin`: https://godbolt.org/z/TGsnnM

~ Johannes


On 2/25/21 4:06 PM, Alexandre Isoard via llvm-dev wrote:
> Hello,
>
> I see InstCombine makes assumptions on functions named with C math function
> names.
>
> For example it will optimize the following code:
>
> target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
> target triple = "x86_64-unknown-linux-gnu"
>
> define dso_local double @fabs(double %x) {
>    %res = fmul double %x, %x
>    ret double %res
> }
>
> define dso_local float @_Z3foof(float %x) {
> entry:
>    %conv = fpext float %x to double
>    %0 = call double @fabs(double %conv)
>    %conv1 = fptrunc double %0 to float
>    ret float %conv1
> }
>
> Into the following:
>
> target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
> target triple = "x86_64-unknown-linux-gnu"
>
> define dso_local float @_Z3foof(float %x) {
> entry:
>    %conv1 = call float @llvm.fabs.f32(float %x)
>    ret float %conv1
> }
>
> ; Function Attrs: nounwind readnone speculatable
> declare double @llvm.fabs.f64(double) #0
>
> ; Function Attrs: nounwind readnone speculatable
> declare float @llvm.fabs.f32(float) #0
>
> attributes #0 = { nounwind readnone speculatable }
>
> It seems to have assumed my @fabs function was the @llvm.fabs.f64
intrinsic.
>
> Is that correct behavior?
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev