[llvm] [AMDGPU] Fix canonicalization of truncated values. (PR #83054)

[Jay Foad via llvm-commits]

jayfoad wrote:

> > Thanks! If I happen to end up figuring out how to simplify the TableGen I can create a followup PR, otherwise I look forward to seeing how someone else manages to do :)
> 
> I managed to get it to work with this:
> 
> ```
> def is_canonicalized : PatLeaf<(fAny srcvalue:$src),
> [{
>     const SITargetLowering &Lowering =
>               *static_cast<const SITargetLowering *>(getTargetLowering());
>     return Lowering.isCanonicalized(*CurDAG, SDValue(N, 0));
> }]
> >;
> 
> def : GCNPat<
>   (fcanonicalize (f32 is_canonicalized:$src)),
>    (COPY f32:$src)
> >;
> ```

Awesome! I wonder if we could do something similar for is_divergent and is_uniform, so we don't need separate DivergentUnaryFrag/DivergentBinFrag/etc.

https://github.com/llvm/llvm-project/pull/83054