[LLVMdev] Canonicalizing vector masking.

[Rafael Espíndola]

On 26 September 2014 19:22, Sean Silva <chisophugis at gmail.com> wrote:
> Hi, I received an internal test case from a game team (it wasn't about this
> in particular), and I was wondering if there was maybe an opportunity to
> canonicalize a particular code pattern:
>
>   %inputi = bitcast <4 x float> %input to <4 x i32>
>
>   %row0i = and <4 x i32> %inputi, <i32 -1, i32 0, i32 0, i32 0>
>   %row0 = bitcast <4 x i32> %row0i to <4 x float>
>
>   %row1i = and <4 x i32> %inputi, <i32 0, i32 -1, i32 0, i32 0>
>   %row1 = bitcast <4 x i32> %row1i to <4 x float>
>
>   %row2i = and <4 x i32> %inputi, <i32 0, i32 0, i32 -1, i32 0>
>   %row2 = bitcast <4 x i32> %row2i to <4 x float>
>
>   %row3i = and <4 x i32> %inputi, <i32 0, i32 0, i32 0, i32 -1>
>   %row3 = bitcast <4 x i32> %row3i to <4 x float>
>
> This arises from code which expands a vector of scale factors into the
> diagonal of a 4x4 diagonal matrix. This code pattern is coming from
> intrinsics which are explicitly doing the masking like this.
>
> My question is: should we canonicalize this to:
>
>   %row0 = shufflevector <4 x float> %input, <4 x float> zeroinitializer, <4
> x i32> <i32 0, i32 4, i32 4, i32 4>
>   %row1 = shufflevector <4 x float> %input, <4 x float> zeroinitializer, <4
> x i32> <i32 4, i32 1, i32 4, i32 4>
>   %row2 = shufflevector <4 x float> %input, <4 x float> zeroinitializer, <4
> x i32> <i32 4, i32 4, i32 2, i32 4>
>   %row3 = shufflevector <4 x float> %input, <4 x float> zeroinitializer, <4
> x i32> <i32 4, i32 4, i32 4, i32 3>
>
> which seems to better express the intent, or a sequence of insertelement and
> extract element (which is what we get for the attached code), or leave it as
> is? (or any better ideas?)
>
> Forgive my naivete if there's something obvious I'm missing since I haven't
> done much w.r.t. vectors in LLVM.

shufflevector does look more canonical. In the past I think we avoided
creating shufflevector for fear of producing bad code in CodeGen, but
I think Chandler just fixed that :-)

Cheers,
Rafael