Richard Biener <richard.guent...@gmail.com> writes:
> On Tue, Dec 12, 2017 at 4:32 PM, Richard Sandiford
> <richard.sandif...@linaro.org> wrote:
>> Richard Biener <richard.guent...@gmail.com> writes:
>>> On Sun, Dec 10, 2017 at 12:06 AM, Richard Sandiford
>>> <richard.sandif...@linaro.org> wrote:
>>>> This series is a replacement for:
>>>> https://gcc.gnu.org/ml/gcc-patches/2017-11/msg00747.html
>>>> based on the feedback that using VEC_PERM_EXPR would be better.
>>>>
>>>> The changes are:
>>>>
>>>> (1) Remove the restriction that the selector elements have to have the
>>>> same width as the data elements, but only for constant selectors.
>>>> This lets through the cases we need without also allowing
>>>> potentially-expensive ops. Adding support for the variable
>>>> case can be done later if it seems useful, but it's not trivial.
>>>>
>>>> (2) Encode the integer form of constant selectors (vec_perm_indices)
>>>> in the same way as the new VECTOR_CST encoding, so that it can
>>>> cope with variable-length vectors.
>>>>
>>>> (3) Remove the vec_perm_const optab and reuse the target hook to emit
>>>> code. This avoids the need to create a CONST_VECTOR for the wide
>>>> selectors, and hence the need to have a corresponding wide vector
>>>> mode (which the target wouldn't otherwise need or support).
>>>
>>> Hmm. Makes sense I suppose.
>>>
>>>> (4) When handling the variable vec_perm optab, check that modes can store
>>>> all element indices before using them.
>>>>
>>>> (5) Unconditionally use ssizetype selector elements in permutes created
>>>> by the vectoriser.
>>>
>>> Why specifically _signed_ sizetype? That sounds like an odd choice.
>>> But I'll
>>> eventually see when looking at the patch.
>>
>> Sorry, should have said. The choice doesn't matter for vector lengths
>> that are a power of 2,
>
> which are the only ones we support anyway?
Yeah, for fixed-length at the tree level. The variable-length support
allows (2^N)*X vectors for non-power-of-2 X though, and we support
non-power-of-2 fixed-length vectors in RTL (e.g. V12QI).
>> but for others, using a signed selector means that
>> -1 always selects the last input element, whereas for unsigned selectors,
>> the element selected by -1 would depend on the selector precision. (And the
>> use of sizetype precision is pretty arbitrary.)
>
> hmm, so you are saying that vec_perm <v1, v2, { -1, -2, ... }> is equal
> to vec_perm <v1, v2, {2*n-1, 2*n-2, ....}?
Yeah.
> tree.def defines VEC_PERM_EXPR via
>
> N = length(mask)
> foreach i in N:
> M = mask[i] % (2*N)
> A = M < N ? v0[M] : v1[M-N]
>
> which doesn't reflect this behavior. Does this behavior persist for variable
> vector permutations?
__builtin_shuffle is defined to wrap though:
The elements of the input vectors are numbered in memory ordering of
@var{vec0} beginning at 0 and @var{vec1} beginning at @var{N}. The
elements of @var{mask} are considered modulo @var{N} in the single-operand
case and modulo @math{2*@var{N}} in the two-operand case.
I think we need to preserve that for VEC_PERM_EXPR, otherwise we'd need
to introduce the masking operation when lowering __builtin_shuffle to
VEC_PERM_EXPR.
>>> Does that mean we have a VNDImode vector unconditionally for the
>>> permute even though a vector matching the width of the data members
>>> would work?
>>
>> A VECTOR_CST of N DIs, yeah. It only becomes a VNDI at the rtl level
>> if we're selecting 64-bit data elements.
>
> And on GIMPLE? Do we have a vector type with ssizetype elements
> unconditionally?
For autovectorised permutes, yes. Other permutes we keep the existing types.
Thanks,
Richard
