[cfe-commits] r154794 - in /cfe/trunk: lib/AST/ExprConstant.cpp test/Sema/const-eval-64.c test/Sema/const-eval.c

[Richard Smith]

On Mon, April 16, 2012 06:16, Eli Friedman wrote:
> On Sun, Apr 15, 2012 at 10:00 PM, Richard Smith <richard at metafoo.co.uk>
> wrote:
>
>> Hi Eli,
>>
>>
>> This "breaks" pointer arithmetic for some cases with undefined behavior
>> (but
>> which are probably common anyway):
>>
>>   int a;
>>   bool b = &a - 1 < &a;
>>
>>
>> I suspect (without evidence) that this happens frequently in real
>> code. Perhaps we should refuse to fold a pointer comparison for a non-null
>> base if either (unsigned) offset is greater than the size of the base
>> object?
>
> I hope it doesn't happen frequently... but yes, we should refuse to
> fold it.  That's basically the same as my FIXME: in the general case, we can't
> tell whether &a - N < &a is true at compile-time.
>
> Can we use Context.getTypeSizeInChars(LHSValue.Base->getType()) to
> compute the size, or is there some other way you would suggest?

That seems fine to me (though use getType(LHSValue.Base) to handle an Expr* or
ValueDecl* base), along with a check for an incomplete base type.