[Bug tree-optimization/115841] 521.wrf_r ICEs when building with -march=znver4 -Ofast -flto --param vect-partial-vector-usage=1

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115841

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2024-07-16
             Status|UNCONFIRMED                 |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot 
gnu.org
     Ever confirmed|0                           |1

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
Reduced testcase, fails with -Ofast -mavx512vl -mtune=znver4 --param
vect-partial-vector-usage=1 -fcommon (-fcommon is the key from fortran so we
can't re-align xl).
For an arch with "proper" costs (aligned loads cheaper) one would swap
the static/not static but even with -fno-vect-cost-model which should make
three loads aligned here it doesn't reproduce.

unsigned char xl[192];
static unsigned char A170[192*3];

void jerate (unsigned char *, unsigned char *);
float foo (unsigned n)
{
  jerate (xl, A170);

  unsigned i = 32;
  int kr = 1;
  float sfn11s = 0.f;
  float sfn12s = 0.f;
  do
    {
      int krm1 = kr - 1;
      long j = krm1;
      float a = (*(float(*)[n])A170)[j];
      float b = (*(float(*)[n])xl)[j];
      float c = a * b;
      float d = c * 6.93149983882904052734375e-1f;
      float e = (*(float(*)[n])A170)[j+48];
      float f = (*(float(*)[n])A170)[j+96];
      float g = d * e;
      sfn11s = sfn11s + g;
      float h = f * d;
      sfn12s = sfn12s + h;
      kr++;
    }
  while (--i != 0);
  float tem = sfn11s + sfn12s;
  return tem;
}