[Bug tree-optimization/110541] New: Invalid VEC_PERM_EXPR mask element size

[Bug tree-optimization/110541] New: Invalid VEC_PERM_EXPR mask element size

[kristerw at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110541

            Bug ID: 110541
           Summary: Invalid VEC_PERM_EXPR mask element size
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: kristerw at gcc dot gnu.org
  Target Milestone: ---

tree.def says:
  The number of MASK elements must be the same with the
  number of elements in V0 and V1.  The size of the inner type
  of the MASK and of the V0 and V1 must be the same.

But tree-vectorizer creates permutations where the MASK element size is
different than for V0 and V1, such as

   vector(8) unsigned short _79;
   ...
  _79 = VEC_PERM_EXPR <_78, { 0, 0, 0, 0, 0, 0, 0, 0 }, { 4, 5, 6, 7, 8, 9, 10,
11 }>;

where the MASK elements are of a 64-bit type.

This can be seen when compiling the following function (from
gcc.c-torture/compile/20000717-1.c) as "gcc -S -O3" for x86_64:

short
inner_product (short *a, short *b)
{
  int i;
  short sum = 0;

  for (i = 9; i >= 0; i--)
    sum += (*a++) * (*b++);

  return sum;
}

[Bug tree-optimization/110541] Invalid VEC_PERM_EXPR mask element size

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110541

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-07-05
             Status|UNCONFIRMED                 |ASSIGNED

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
The restriction as documented in tree.def applies to non-constant MASKs only,
the case for constant MASKs was relaxed.  Like

/* Vector permutation expression.  A = VEC_PERM_EXPR<v0, v1, mask> means

   N = length(mask)
   foreach i in N:
     M = mask[i] % (length(v0) + length(v1))
     A[i] = M < length(v0) ? v0[M] : v1[M - length(v0)]

   V0 and V1 are vectors of the same type.

   When MASK is not constant:
     MASK is an integer-typed vector.  The number of MASK elements must 
     be the same with the number of elements in V0 and V1.  The size of 
     the inner type of the MASK and of the V0 and V1 must be the same.

   When MASK is constant:
     MASK is an integer-typed vector.   MASK elements outside of 
     [0, length(V0) + length(V1) - 1] invoke undefined behavior (the 
     modulo operation above doesn't apply).
*/
DEFTREECODE (VEC_PERM_EXPR, "vec_perm_expr", tcc_expression, 3)

I will submit this change.

[Bug tree-optimization/110541] Invalid VEC_PERM_EXPR mask element size

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110541

--- Comment #2 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:9fed1ec67f83f3c56c030fe2a501b01ceb6e243f

commit r14-2311-g9fed1ec67f83f3c56c030fe2a501b01ceb6e243f
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Jul 5 08:53:01 2023 +0200

    middle-end/110541 - VEC_PERM_EXPR documentation is off

    The following adjusts the tree.def documentation about VEC_PERM_EXPR
    which wasn't adjusted when the restrictions of permutes with constant
    mask were relaxed.

            PR middle-end/110541
            * tree.def (VEC_PERM_EXPR): Adjust documentation to reflect
            reality.

[Bug tree-optimization/110541] Invalid VEC_PERM_EXPR mask element size

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110541

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|ASSIGNED                    |RESOLVED
           Keywords|                            |documentation

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed as documentation issue.  Note

"MASK elements outside of [0, length(V0) + length(V1) - 1] invoke undefined
behavior (the modulo operation above doesn't apply)."

isn't actually the case, it would be a missed canonicalization only.