Re: [RFC PATCH, vectorizer]: Vectorize int -> double conversions

["Daniel Berlin" <dberlin at dberlin dot org>]

On 4/23/07, Uros Bizjak <ubizjak@gmail.com> wrote:
> Hello!
>
> Attached (RFC!) patch implements vectorization of int -> double
> conversion. The testcase:
>
>   /* int -> double */
>   for (i = 0; i < N; i++)
>     {
>       da[i] = (double) ib[i];
>     }
>
> compiles on i686 -msse2 target into:
>
> .L3:
>         movdqa  (%eax,%ecx), %xmm0
>         cvtdq2pd        %xmm0, %xmm1
>         pshufd  $238, %xmm0, %xmm0
>         movapd  %xmm1, (%edx,%eax,2)
>         cvtdq2pd        %xmm0, %xmm0
>         movapd  %xmm0, 16(%edx,%eax,2)
>         addl    $16, %eax
>         cmpl    $128, %eax
>         jne     .L3
>
> (pshufd is there to shuffle SImode vector from x0x1x2x3 into x2x3x2x3.
>
> Regarding the patch: in vectorizable_conversion() we detect
> (nunits_out == nunits_in / 2) as EXPAND case and handle conversion in
> the same way as vectorizable_type_promotion(). Unfortunatelly, we need
> a couple of new tree codes and 4 new optabs to handle signed and
> unsigned conversion.
Dorit, are there any common architectures which support odd size promotions?
If so, maybe VEC_UNPACK_FLOAT{HI,LOW} should simply be
VEC_UNPACK_FLOAT, specifying the start and length into the vector in
it's operands.

Then you can use HI/LOW optabs if the size == vec_size /2 and the
start pos == 0 || vec_size/2 (and there is no better vec_unpack optab)

>