[PATCH 2/2] ARM cost model: Add vector reverse shuffle costs

[Renato Golin]

I think all three patches look good, feel free to chose the right patch
first, to reduce future changes.


On 9 February 2013 19:54, Arnold Schwaighofer <aschwaighofer at apple.com>wrote:

> Yes I will, depending on which patch goes in first.
>
> Thanks
>
> On Feb 9, 2013, at 1:35 PM, Renato Golin <renato.golin at linaro.org> wrote:
>
> +    // Currently, not well supported by backend and scalarized to memory.
> This
> +    // number is somewhat arbitrary we just want to make it really
> expensive.
> +    { ISD::VECTOR_SHUFFLE, MVT::v8i16, 8*4 },
> +    { ISD::VECTOR_SHUFFLE, MVT::v16i8, 16*4 },
>
> You should change it, now that you implemented it correctly.
>
> Code looks good.
>
> --renato
>
>
> On 8 February 2013 17:42, Arnold Schwaighofer <aschwaighofer at apple.com>wrote:
>
>> A reverse shuffle is lowered to a vrev and possibly a vext instruction
>> (quad
>> word). The backend has a deficiency in lowering native v8i16 and v16i8
>> types.
>> They are scalarized to memory. Reflect this in their cost.
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130209/34a22319/attachment.html>