On 27 June 2012 21:27, Richard Henderson <r...@redhat.com> wrote:
> On 06/20/2012 05:44 AM, Ramana Radhakrishnan wrote:
>> + case NEON_DUP:
>> + if (TREE_CODE (argp[0]) == INTEGER_CST
>> + || TREE_CODE (argp[0]) == REAL_CST)
>> + return build_vector_from_val (result_type, argp[0]);
>> + return NULL_TREE;
>
> You can expand this in all cases. Constants go to VECTOR_CST,
> as you're doing, but variables can be expanded via a CONSTRUCTOR.
>
> Check out what we generate for
>
> (v4si){ x, x, x, x }I finally wrote something like this up. It has to special case the fact the DUP intrinsics take a non-vector 64 bit form and hence you just can't use expand_vector_from_val. This however has fallout in the neon intrinsics test results especially all the vdup* and the vmov*_n* intrinsics. Sigh and that needs some cleanup. regards, Ramana > > > r~
vdup-improvement.patch
Description: Binary data
