[Bug tree-optimization/107451] [11/12/13 Regression] Segmentation fault with vectorized code since r11-6434

rguenth at gcc dot gnu.org gcc-bugzilla@gcc.gnu.org
Thu Nov 17 14:58:20 GMT 2022


https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107451

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
Interestingly the following variant of the testcase falls back to
VMAT_ELEMENTWISE but does have the same problem there fixed up by later
folding, but it will segfault when using -O2 -mavx2 -fno-vect-cost-model
-fdisable-tree-vrp2 -fdisable-free-forwprop4 which then keeps the bogus

  _62 = *ivtmp_64;
  _61 = MEM[(const double *)ivtmp_64 + 8B];
  ivtmp_60 = ivtmp_64 + _65;
  _59 = *ivtmp_60;
  _58 = MEM[(const double *)ivtmp_60 + 8B];
  ivtmp_57 = ivtmp_60 + _65;
  vect_cst__56 = {_62, _61, _59, _58};
  vect__4.7_55 = VEC_PERM_EXPR <vect_cst__56, vect_cst__56, { 0, 1, 0, 1 }>;

that problem should be present even before the r11-6434 change.  In fact
this segfaults on the GCC 10 branch with just -O2 -ftree-loop-vectorize -mavx2
generating the same load/permute as trunk for the reduction (so there's some
half-way "fix" on the later branches).  Also broken with GCC 9.5.

static void __attribute__((noipa))
setdot(int n, const double *x, int inc_x, const double *y, double * __restrict
dot)
{
  int i, ix = 0;

  for(i = 0; i < n; i++) {
      dot[i*4+0] = x[ix]          * y[ix]   ;
      dot[i*4+1] = x[ix+1] * y[ix+1] ;
      dot[i*4+2] = x[ix]          * y[ix+1] ;
      dot[i*4+3] = x[ix+1] * y[ix]   ;
      ix += inc_x ;
  }
}

int main(void)
{
  double x[2] = {0, 0}, y[2] = {0, 0};
  double dot[4];
  setdot(1, x, 4096*4096, y, dot);
  return 0;
}


More information about the Gcc-bugs mailing list