[Bug tree-optimization/63665] [4.8/4.9 Regression] wrong code with signed overflow even with -fwrapv

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63665

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
Latent issue.

pr45752.c:40:3: note: Load permutation 2 2 2 2 2 3 3 3 3 3 0 0 0 0 0 1 1 1 1 1
4 4 4 4 4

We have a SLP group size of 5 and computed an unroll factor of 2 (ok).  It
looks
to me that

Index: gcc/tree-vect-slp.c
===================================================================
--- gcc/tree-vect-slp.c (revision 216771)
+++ gcc/tree-vect-slp.c (working copy)

@@ -3041,6 +3041,7 @@ vect_transform_slp_perm_load (slp_tree n
                                          &number_of_mask_fixes, &mask_fixed,
                                          &needs_first_vector))
                return false;
+             gcc_assert (current_mask_element < 2 * nunits);
              mask[index++] = current_mask_element;

               if (index == nunits)

makes x86_64 ICE as well.  The issue is that we suddenly need to skip _two_
vectors in vect_get_mask_element which isn't implemented so we get
out-of-bound indexes.  Seems we need to iterate instead in which case
we will reject vectorization correctly with

pr45752.c:40:3: note: Load permutation 2 2 2 2 2 3 3 3 3 3 0 0 0 0 0 1 1 1 1 1
4 4 4 4 4
pr45752.c:40:3: note: permutation requires at least three vectors c_71 =
MEM[(unsigned int *)pInput2_135 + 8B];

pr45752.c:40:3: note: Build SLP failed: unsupported load permutation
*pOutput2_136 = _83;