https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124138

--- Comment #6 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <[email protected]>:

https://gcc.gnu.org/g:b95f529df46e33c7ee7058b7cc51473e7a0d2f4c

commit r16-7565-gb95f529df46e33c7ee7058b7cc51473e7a0d2f4c
Author: Jakub Jelinek <[email protected]>
Date:   Wed Feb 18 10:28:34 2026 +0100

    i386: Fix up avx_vpermilp_parallel [PR124138]

    r16-3677 extended avx_vpermilp_parallel so that it handles also V8DImode
    (among others) and broke this testcase.
    For V8DFmode the bug existed there already since r0-127405 which extended
    avx_vpermilp_parallel to handle V8DFmode.
    For V16SImode/V16SFmode I think the code is correct, in that case it can
    for V4SImode/V4SFmode choose any permutation it wants and the
    V8S{I,F}mode and V16S{I,F}mode cases before FALLTHRU verify the upper lanes
    match those lower ones.
    For V[48]D[IF]mode the function uses different checking, where it checks
    each pair of ipar separately:
        case E_V8DFmode:
        case E_V8DImode:
          /* In the 512-bit DFmode case, we can only move elements within
             a 128-bit lane.  First fill the second part of the mask,
             then fallthru.  */
          for (i = 4; i < 6; ++i)
            {
              if (ipar[i] < 4 || ipar[i] >= 6)
                return 0;
              mask |= (ipar[i] - 4) << i;
            }
          for (i = 6; i < 8; ++i)
            {
              if (ipar[i] < 6)
                return 0;
              mask |= (ipar[i] - 6) << i;
            }
          /* FALLTHRU */

        case E_V4DFmode:
        case E_V4DImode:
          /* In the 256-bit DFmode case, we can only move elements within
             a 128-bit lane.  */
          for (i = 0; i < 2; ++i)
            {
              if (ipar[i] >= 2)
                return 0;
              mask |= ipar[i] << i;
            }
          for (i = 2; i < 4; ++i)
            {
              if (ipar[i] < 2)
                return 0;
              mask |= (ipar[i] - 2) << i;
            }
          break;
    so that it directly computes corresponding mask bit.
    Earlier in the function it verified ipar[i] wasn't above
    nelts (nor negative), so just checking if (ipar[i] >= 2) is ok, it can't
    be negative, and just checking if (ipar[i] < 6) is also fine, because
    it can never be 8 or above (because nelts is 8 in that case).
    Though, the if (ipar[i] < 2) case check used to be correct only when
    nelts could be only 4, when it can be 8 too, it will as in the following
    testcase happily accept values in the [4, 7] range which shouldn't be
    accepted (it can only handle [2, 3]).

    Fixed thusly.

    2026-02-17  Jakub Jelinek  <[email protected]>

            PR target/124138
            * config/i386/i386.cc (avx_vpermilp_parallel): Verify
            ipar[2] and ipar[3] aren't larger than 3.

            * gcc.dg/pr124138.c: New test.
            * gcc.target/i386/avx512f-pr124138-1.c: New test.
            * gcc.target/i386/avx512f-pr124138-2.c: New test.

Reply via email to