[patch] for PRs 27639 and 26719

[Eric Botcazou]

> We need to ensure that scev_probably_wraps_p returns a correct answer (it is
> not clear from your explanation to me whether this is currently the case,
> but given that your patch makes things less conservative, I think so),

Sorry for being unclear, but scev_probably_wraps_p returns the correct answer.
The problem is that it is not called by convert_affine_scev because of the 
return value of nowrap_type_p.  So, in some sense, my patch makes things more 
conservative, not less.

> Introducing this type of hack into a correct analysis is a bad thing.

Well, not that correct for unsigned subtypes.

> I will have a look at the problem today.

Thanks in advance.

-- 
Eric Botcazou