[Bug tree-optimization/100239] [10/11/12 Regression] ICE: in expand_expr_real_2, at expr.c:9865 with __builtin_shuffle()

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100239

--- Comment #4 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-11 branch has been updated by Jakub Jelinek
<jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:80dc24f813a9d8951d75eea7324f6d713b872bca

commit r11-8300-g80dc24f813a9d8951d75eea7324f6d713b872bca
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Tue Apr 27 15:42:47 2021 +0200

    veclower: Fix up vec_shl matching of VEC_PERM_EXPR [PR100239]

    The following testcase ICEs at -O0, because lower_vec_perm sees the
      _1 = { 0, 0, 0, 0, 0, 0, 0, 0 };
      _2 = VEC_COND_EXPR <_1, { -1, -1, -1, -1, -1, -1, -1, -1 }, { 0, 0, 0, 0,
0, 0, 0, 0 }>;
      _3 = { 6, 0, 0, 0, 0, 0, 0, 0 };
      _4 = VEC_PERM_EXPR <{ 0, 0, 0, 0, 0, 0, 0, 0 }, _2, _3>;
    and as the ISA is SSE2, there is no support for the particular permutation
    nor for variable mask permutation.  But, the code to match vec_shl matches
    it, because the permutation has the first operand a zero vector and the
    mask picks all elements randomly from that vector.
    So, in the end that isn't a vec_shl, but the permutation could be in theory
    optimized into the first argument.  As we keep it as is, it will fail
    during expansion though, because that for vec_shl correctly requires that
    it actually is a shift:
          unsigned firstidx = 0;
          for (unsigned int i = 0; i < nelt; i++)
            {
              if (known_eq (sel[i], nelt))
                {
                  if (i == 0 || firstidx)
                    return NULL_RTX;
                  firstidx = i;
                }
              else if (firstidx
                       ? maybe_ne (sel[i], nelt + i - firstidx)
                       : maybe_ge (sel[i], nelt))
                return NULL_RTX;
            }

          if (firstidx == 0)
            return NULL_RTX;
          first = firstidx;
    The if (firstidx == 0) return NULL; is what is missing a counterpart
    on the lower_vec_perm side.
    As with optimize != 0 we fold it in other spots, I think it is not needed
    to optimize this cornercase in lower_vec_perm (which would mean we'd need
    to recurse on the newly created _4 = { 0, 0, 0, 0, 0, 0, 0, 0 };
    whether it is supported or not).

    2021-04-27  Jakub Jelinek  <jakub@redhat.com>

            PR tree-optimization/100239
            * tree-vect-generic.c (lower_vec_perm): Don't accept constant
            permutations with all indices from the first zero element as
vec_shl.

            * gcc.dg/pr100239.c: New test.

    (cherry picked from commit 83d26d0e1b3625ab6c2d83610a13976b52f63e0a)