[llvm] r277749 - GlobalISel: also add G_TRUNC to IRTranslator.

[Quentin Colombet via llvm-commits]

> On Aug 18, 2016, at 1:33 PM, Tim Northover <tnorthover at apple.com> wrote:
> 
> On 18 Aug 2016, at 12:10, Quentin Colombet <qcolombet at apple.com> wrote:
>> We were discussing this offline and came to the conclusion that G_TRUNC is a special case of G_EXTRACT.
>> In other words, we wouldn’t need that instruction.
> 
> I think it's different for vectors (I considered reusing it at the time).
> 
>    %trunc = G_EXTRACT { <4 x i16>, <4 x i32> } %src
> 
> would extract the low 64-bits of the %src reg.
> 
> It could probably be represented as a G_EXTRACT followed by a G_SEQUENCE though:
> 
>   %e0, %e1, %e2, %e3 = G_EXTRACT {s16, s16, s16, s16, <4 x s32> } %src, 0, 32, 48, 64
>   %vec = G_SEQUENCE { <4 x s16>, s16, s16, s16, s16 } %e0, %e1, %e2, %e3
> 
> I'm not entirely convinced it would be an improvement though. Depends on how orthogonal we want these things to be.

Good point we haven’t though about the vector case.
What that means is that we should canonicalize on either G_TRUNC or G_EXTRACT when their semantics overlap. G_TRUNC is probably more readable.

> 
>> This would be good for us. We currently have to deal with a lot of DAG combines switching between trunc and vector extracts
> 
> I don't quite follow this. They seem pretty incompatible even in the SDag world.
> 
> Tim.
> 
>