On Thu, 22 Mar 2018, Jakub Jelinek wrote:
> Hi!
>
> Something I wasn't really aware, apparently we allow extern vars with
> incomplete types in "m" and "=m" constrained asm. The problem is that
> incomplete vars have VOIDmode mode and so their MEMs do as well, apparently
> everything works with it except a 2013-ish assert in alias.c.
>
> The following patch just makes sure x_mode is not VOIDmode if x doesn't have
> VOIDmode, if x has VOIDmode, then it is fine for x_mode to be VOIDmode too.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK.
The question now is whether we handle dependences correctly for
VOIDmode MEMs? Do they have !MEM_SIZE_KNOWN_P?
Thanks,
Richard.
> 2018-03-22 Jakub Jelinek <ja...@redhat.com>
>
> PR inline-asm/85022
> * alias.c (write_dependence_p): Don't require for x_canonicalized
> non-VOIDmode if x has VOIDmode.
>
> * c-c++-common/torture/pr85022.c: New test.
>
> --- gcc/alias.c.jj 2018-03-01 11:29:06.725104754 +0100
> +++ gcc/alias.c 2018-03-22 10:35:15.649546027 +0100
> @@ -2999,7 +2999,8 @@ write_dependence_p (const_rtx mem,
> int ret;
>
> gcc_checking_assert (x_canonicalized
> - ? (x_addr != NULL_RTX && x_mode != VOIDmode)
> + ? (x_addr != NULL_RTX
> + && (x_mode != VOIDmode || GET_MODE (x) == VOIDmode))
> : (x_addr == NULL_RTX && x_mode == VOIDmode));
>
> if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem))
> --- gcc/testsuite/c-c++-common/torture/pr85022.c.jj 2018-03-22
> 10:44:40.520679173 +0100
> +++ gcc/testsuite/c-c++-common/torture/pr85022.c 2018-03-22
> 10:44:29.550675460 +0100
> @@ -0,0 +1,9 @@
> +/* PR inline-asm/85022 */
> +
> +extern struct B b;
> +
> +void
> +foo ()
> +{
> + __asm ("" : "+m" (b));
> +}
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB
21284 (AG Nuernberg)
