on 2021/5/27 下午8:55, Richard Sandiford wrote:
> Sorry for the slow reponse.
>
> "Kewen.Lin" <li...@linux.ibm.com> writes:
>> diff --git a/gcc/vec-perm-indices.c b/gcc/vec-perm-indices.c
>> index ede590dc5c9..57dd11d723c 100644
>> --- a/gcc/vec-perm-indices.c
>> +++ b/gcc/vec-perm-indices.c
>> @@ -101,6 +101,70 @@ vec_perm_indices::new_expanded_vector (const
>> vec_perm_indices &orig,
>> m_encoding.finalize ();
>> }
>>
>> +/* Check whether we can switch to a new permutation vector that
>> + selects the same input elements as ORIG, but with each element
>> + built up from FACTOR pieces. Return true if yes, otherwise
>> + return false. Every FACTOR permutation indexes should be
>> + continuous separately and the first one of each batch should
>> + be able to exactly modulo FACTOR. For example, if ORIG is
>> + { 2, 3, 4, 5, 0, 1, 6, 7 } and FACTOR is 2, the new permutation
>> + is { 1, 2, 0, 3 }. */
>> +
>> +bool
>> +vec_perm_indices::new_shrunk_vector (const vec_perm_indices &orig,
>> + unsigned int factor)
>> +{
>> + gcc_assert (factor > 0);
>> +
>> + if (maybe_lt (orig.m_nelts_per_input, factor))
>> + return false;
>> +
>> + poly_uint64 nelts;
>> + /* Invalid if vector units number isn't multiple of factor. */
>> + if (!multiple_p (orig.m_nelts_per_input, factor, &nelts))
>> + return false;
>> +
>> + /* Only handle the case that npatterns is multiple of factor.
>> + FIXME: Try to see whether we can reshape it by factor npatterns. */
>> + if (orig.m_encoding.npatterns () % factor != 0)
>> + return false;
>> +
>> + unsigned int encoded_nelts = orig.m_encoding.encoded_nelts ();
>> + auto_vec<element_type> encodings (encoded_nelts);
>
> auto_vec<element_type, 32> would avoid memory allocations in the
> same cases that m_encoding can. “encoding” might be better than
> “encodings” since there's only really one encoding here.
>
>> + /* Separate all encoded elements into batches by size factor,
>> + then ensure the first element of each batch is multiple of
>> + factor and all elements in each batch is consecutive from
>> + the first one. */
>> + for (unsigned int i = 0; i < encoded_nelts; i += factor)
>> + {
>> + element_type first = orig.m_encoding[i];
>> + element_type new_index;
>> + if (!multiple_p (first, factor, &new_index))
>> + return false;
>> + for (unsigned int j = 1; j < factor; ++j)
>> + {
>> + if (maybe_ne (first + j, orig.m_encoding[i + j]))
>> + return false;
>> + }
>
> Formatting nit: unnecessary braces around if.
>
>> + encodings.quick_push (new_index);
>> + }
>> +
>> + m_ninputs = orig.m_ninputs;
>> + m_nelts_per_input = nelts;
>> + poly_uint64 full_nelts = exact_div (orig.m_encoding.full_nelts (),
>> factor);
>> + unsigned int npatterns = orig.m_encoding.npatterns () / factor;
>> +
>> + m_encoding.new_vector (full_nelts, npatterns,
>> + orig.m_encoding.nelts_per_pattern ());
>> +
>> + for (unsigned int i = 0; i < encodings.length (); i++)
>> + m_encoding.quick_push (encodings[i]);
>
> I think this can be:
>
> m_encoding.splice (encodings);
>
> OK with those changes, thanks. Thanks also for doing it in a
> variable-length-friendly way.
>
Thanks for the comments, Richard! The patch was updated as them,
re-tested and committed in r12-1103.
BR,
Kewen
