[Bug tree-optimization/18241] [4.0 Regression] linux kernel loop gets miscompiled

["law at redhat dot com" <gcc-bugzilla at gcc dot gnu dot org>]

------- Additional Comments From law at redhat dot com  2004-12-13 21:01 -------
(In reply to comment #15)
> Confirmed,  this looks like a jump threading bug.
> Before we have:
>  goto <bb 9> (<L10>);
> 
> <L8>:;
>   D.7815_42 = pathp_2->node;
>   tag_43 = tag_27;
>   idx_44 = idx_6;
>   D.7819_45 = D.7815_42->tags[tag_43][idx_44];
>   if (D.7819_45 != 0) goto out; else goto <L9>;
> 
> <L9>:;
>   idx_46 = idx_6 + 1;
> 
>   # idx_6 = PHI <idx_39(6), idx_46(8)>;
> <L10>:;
>   if (idx_6 == 0) goto <L8>; else goto <L11>;
> 
> 
> But after:
>  goto <bb 8> (<L10>);
> 
> <L9>:;
>   idx_46 = 1;
>   goto <bb 10> (<L15>);
> 
>   # idx_57 = PHI <0(6)>;
> <L10>:;
>   D.7815_42 = pathp_2->node;
>   tag_43 = tag_27; 
>   idx_44 = 0;   <---- this is just plainly wrong.
>   D.7819_45 = D.7815_42->tags[tag_27][0];
>   if (D.7819_45 != 0) goto out; else goto <L9>;

Huh?  idx_44 must be zero, I don't see how any other value can be correct here.

THe only way to get to L8 (block #7) is when idx_6 == 0 and idx_44 is set 
unconditionally from idx_6.

So what about that assignment do you think is wrong?

Note that we also know that idx_46 must have the value 1 because it's only
reachable from block #7 where we know that idx_6 == 0 and thus idx_46 must
have the value 1.




-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18241