[Bug rtl-optimization/67676] New: Implicit alignment of struct not applied to members

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67676

            Bug ID: 67676
           Summary: Implicit alignment of struct not applied to members
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: olegendo at gcc dot gnu.org
  Target Milestone: ---
            Target: sh*-*-*

I've noticed this while doing PR 67675...


On strict-alignment targets (like SH) a pointer to a struct is always assumed
to be aligned:

struct S
{
  int a, b, c;
  char s[64];
};

'struct S*' will always be 4 byte aligned.

Accessing a member works as expected:

int foo4 (struct S* x)
{
  return x->b;
}

        mov.l   @(4,r4),r0
        rts/n

However, when the address of a member is passed to __builtin_strcmp, the
alignment is lost somehow:

int foo5 (struct S* x)
{
  return __builtin_strcmp (x->s, "1234");
}

Here the builtin strcmp expander will see x->s with an alignment of 1, although
it's 4.

Adding an alignment hint gives the expected result:

int foo6 (struct S* x)
{
  struct S* xx = (struct S*)__builtin_assume_aligned (x, 4);
  return __builtin_strcmp (xx->s, "1234");
}

I'm not sure whether this is a tree or RTL problem.  Happens for C and C++.
Possibly related: PR 23467, PR 50417