[Bug fortran/80645] FAIL: gfortran.dg/elemental_subroutine_3.f90 -O1 (test for excess errors)

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80645

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |WAITING

--- Comment #12 from Richard Biener <rguenth at gcc dot gnu.org> ---
Looking again for the issue in the original report (the comment#1 issue has
been fixed):

    arg:0 <var_decl 0x7ffff7fee5a0 x type <array_type 0x7ffff69a4930>
        addressable used static decl_0 BLK
/space/rguenther/src/svn/trunk2/gcc/testsuite/gfortran.dg/elemental_subroutine_3.f90:29:0
size <integer_cst 0x7ffff697d930 192> unit-size <integer_cst 0x7ffff699c588 24>

and we warn about

__builtin_memcpy (_153, &MEM[(void *)&x + 16B], 12)

with x having size 24 there's not 12 bytes at &x + 16.  It is DSE3 first
having the size of 12 here:

   <bb 18> [local count: 20817692]:
   _95 = __builtin_malloc (16);
-  __builtin_memcpy (_95, &x[3], 16);
+  _153 = _95 + 4;
+  __builtin_memcpy (_153, &MEM[(void *)&x + 16B], 12);

this is from the following code in .original:

    parm.11.dtype = 297;
    parm.11.dim[0].lbound = 0;
    parm.11.dim[0].ubound = 3;
    parm.11.dim[0].stride = -1;
    parm.11.data = (void *) &x[3];
    parm.11.offset = 0;
        typedef struct mytype [4];
...
    data.13 = (struct mytype *) _gfortran_internal_pack (&parm.11);
    if ((struct mytype[0:] *) parm.11.data == data.13)
      {
        {
          void * restrict D.2228;

          D.2228 = (void * restrict) __builtin_malloc (16);
          data.13 = (struct mytype *) D.2228;
          (void) __builtin_memcpy ((void *) data.13, parm.11.data, 16);
        }
      }
    atmp.12.data = (void * restrict) data.13;
    atmp.12.offset = 0;
    {
      integer(kind=8) S.14;

      S.14 = 0;
      while (1)
        {
          if (S.14 > 3) goto L.5;
          {
            integer(kind=4) D.2230;

            D.2230 = (*(integer(kind=4)[0:] *)
parm.10.data)[parm.10.dim[0].stride * NON_LVALUE_EXPR <S.14>];
            myassign (&(*(struct mytype[4] * restrict) atmp.12.data)[S.14],
&x[(integer(kind=8)) D.2230 + -1]);
          }
          S.14 = S.14 + 1;
        }
      L.5:;

so we somehow allocate a temporary for the view on the LHS even though
it is packed in-place (and later unpacked again) but somehow miscalculate
the size of the required temporary.  Possibly because of the negative
stride.

   x(4:1:-1) = x((/1,3,2,4/))

-> gfortran wrong-code bug.  The reduced testcase from comment#10 does _not_
do this pack/unpack and temporary dance.  Probably requires the myassign thing?

Fortran folks, please investigate again!