Support vec-perm for NEON

VEC_PERM_EXPR is used for general permutation of vector elements. The vectorizer generates VEC_PERM_EXPR in two cases: for load permutation in SLP and for backward accesses. We cover the first case with vldN. The second case only requires to revert the order of vector elements, which is a much easier task.

Implementing general permutation and using it for backward accesses may not be the most efficient solution. It may be better to introduce a new optab for vector revert and use it in the vectorizer for such cases. Another option is to generate a vtbl.8 in the most generic permute case, and split the reverse case into a special case using the mask to differentiate between the cases. To implement reverse we can use vrev followed by vswp or just a vrev depending on the number of bits in the vectors under consideration.

We already have a BSWAP operation on RTL .

This will help vectorize libav dcadsp.c: ff_dca_lfe_fir().

Another case libav fdct (will require some support in SLP):

    tmp0 = dataptr[0] + dataptr[7];
    tmp1 = dataptr[1] + dataptr[6];
    tmp2 = dataptr[2] + dataptr[5];
    tmp3 = dataptr[3] + dataptr[4];

    tmp7 = dataptr[0] - dataptr[7];
    tmp6 = dataptr[1] - dataptr[6];
    tmp5 = dataptr[2] - dataptr[5];
    tmp4 = dataptr[3] - dataptr[4];