Re: PR 6212

[Richard Henderson <rth at redhat dot com>]

On Mon, May 06, 2002 at 02:56:09PM -0700, Mark Mitchell wrote:
> I can verify the patch on GNU/Linux, too.  Is there someone who would
> care to try this on a MIPS or Alpha box to confirm that it works OK
> there?  (Alpha was also affected by this problem.)

I'll test Alpha, but I'm pretty sure this patch is incorrect.

Consider

	int foo(int *p)
	{
	  return *((char *)p);
	}

	char c[2] __attribute__((aligned));

	int bar()
	{
	  return foo((int *)&c[1]);
	}

Since we recurse all the way down to the PARM_DECL, we decide that P
is 4-byte aligned.  We already discussed cases with pointers having
strange values (so long as they weren't dereferenced) and declared
them acceptable in gcc.

No, I think a proper solution is to let highest_pow2_factor increase
the alignment, but not decrease it.  Assuming the alignment of the
type *that we dereference* would match the behaviour of previous
gcc versions.


r~