[Bug target/43635] [4.5/4.6 Regression] ICE in s390_emit_call, at config/s390/s390.c:9484

["rguenther at suse dot de" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #4 from rguenther at suse dot de  2010-04-08 08:11 -------
Subject: Re:  [4.5/4.6 Regression] ICE in s390_emit_call,
 at config/s390/s390.c:9484

On Wed, 7 Apr 2010, krebbel at gcc dot gnu dot org wrote:

> ------- Comment #3 from krebbel at gcc dot gnu dot org  2010-04-07 15:56 -------
> Created an attachment (id=20328)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20328&action=view)
>  --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20328&action=view)
> Reduced testcase
> 
> The check for valid sibcalls in the S/390 back-end relies on the decl to be
> available in the DECL parameter:
> 
>   /* The 31 bit PLT code uses register 12 (GOT pointer - caller saved)
>      which would have to be restored before the sibcall.  */
>   if (!TARGET_64BIT && flag_pic && decl && !targetm.binds_local_p (decl))
>     return false;
> 
> The type cast of the function pointer in the testcase introduces a temporary
> variable on 4.5 which is not linked with the function decl anymore:
> 
>   x_1 = (void (*<T244>) (float) (*<T248>) (void *)) d;
>   D.1966_3 = x_1 (p_2(D)); [tail call]
>   return D.1966_3;
> 
> With 4.4:
> 
>   D.1224 = (void (*<T242>) (float) (*<T246>) (void *)) d (p); [tail call]
>   return D.1224;
> 
> So we probably have to always reject sibcalls if the fndecl is not reachable.

The extra cast happens if the function signatures are not compatible
(by means of useless_type_conversion_p).  I can't deciper the
function type from (void (*<T244>) (float) (*<T248>) (void *))
right now, but it's early in the morning ;)

Richard.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43635