Richard Biener <richard.guent...@gmail.com> writes:
> On Fri, Dec 15, 2017 at 1:29 AM, Richard Sandiford
> <richard.sandif...@linaro.org> wrote:
>> This patch just adds VEC_DUPLICATE_EXPR, since the VEC_DUPLICATE_CST
>> isn't needed with the new VECTOR_CST layout. It's really just the
>> original patch with bits removed, but just in case:
>>
>> Tested on aarch64-linux-gnu, x86_64-linux-gnu and powerpc64-linux-gnu.
>> OK to install?
>
> To keep things simple at this point OK. Note that I'd eventually
> like to see this as VEC_PERM_EXPR <scalar_type_1, scalar_type_1, { 0, ... }>.
> For reductions when we need { x, 0, ... } we now have to use a
> VEC_DUPLICATE_EXPR to make x a vector and then a VEC_PERM_EXPR
> to merge it with {0, ... }, right? Rather than VEC_PERM_EXPR <x_1, 0,
> { 0, 1, 1, 1.... }>
That's where the shift-left-and-insert-scalar thing (IFN_SHL_INSERT)
comes in. But yeah, allowing scalars as operands to VEC_PERM_EXPRs
would mean it could represent both VEC_DUPLICATE_EXPR and IFN_SHL_INSERT.
I guess the question is whether that's better than extending CONSTRUCTOR
(or a replacement) to use the VECTOR_CST encoding. I realise you don't
like CONSTRUCTOR in gimple though...
I promise to look at either of those for GCC 9 if you think they're
better, but they'll be more invasive for other targets.
Thanks,
Richard
