> I've tested and pushed a patch to remove this old cruft.  The comments appear 
> to be lifted directly from arm.cc, which isn't a surprise because Julian 
> probably wrote both instances, a few years apart.
>
> I don't really understand what the "is_packed" means, but I get an ICE 
> caused by "subreg:SI (reg:V4DI ...)" if I allow that.  The hardware will 
> allow loading any data from any alignment, but you can't take arbitrary 
> subregs of vectors.

Is that with the patch or without?  is_packed tells you if the originally 
accessed scalar object was not naturally aligned.  Allowing such an access 
should be OK according to your description. Maybe it just exposes a latent
issue? For riscv we prevent certain subregs in can_change_mode_class. 

If it's with the patch then maybe try the respun version (v3).

> Sorry about the merge conflict, but hopefully now you don't need to do 
> anything here.

Yep, much easier that way, thanks.  I split off the alignment patch since and 
it has been approved but I'll probably respin it once more before pushing.

-- 
Regards
 Robin

Reply via email to