[PATCH] Checked arithmetic builtins

[Eli Friedman]

On Thu, Jun 20, 2013 at 2:39 PM, Michael Gottesman <mgottesman at apple.com>wrote:

>
> On Jun 20, 2013, at 2:16 PM, Eli Friedman <eli.friedman at gmail.com> wrote:
>
> On Tue, Jun 18, 2013 at 3:11 PM, Michael Gottesman <mgottesman at apple.com>
> wrote:
>
>> Hello cfe commits!
>>
>> The attached patch adds checked-arithmetic builtins to ameliorate such
>> code in security critical applications (for instance webkit). It simply
>> exposes {u,s}{add,sub,mul}.
>>
>> *NOTE* The u{add,sub} overlaps with the multi precision built-ins. I
>> decided to add in the additional builtin since users are going to see
>> s{add,sub} and look for u{add,sub}. We could add in a Builtins.h header
>> where I could implement the checked arithmetic with the multi precision
>> arithmetic builtins but I felt that was a bigger change than this.
>>
>>
> It looks like you copy-pasted some stuff you didn't mean to in the changes
> to LanguageExtensions.rst.
>
>
> = /. Fixed.
>
>
> +    return RValue::get(Builder.CreateZExt(Carry, X->getType()));
>
> These builtins all return bool, right?  Why are you zero-extending here?
>  (You might be able to get away with this in C because of integer
> promotions, but I'm pretty sure it'll explode in C++.)
>
>
> = /. Fixed. The reason I was zero extending was that I was basing this off
> of the multiprecision builtin code and forgot to remove it.
>
>
> IIRC, we don't actually support CodeGen of llvm.smul.with.overflow with
> 64-bit operands on x86-32; it would be nice to print a proper error message
> instead of crashing.  (At least, we didn't at one point; I don't recall if
> it ever got fixed.)
>
>
> I just codegened my test file on OS X with -arch i386 and it codegened
> fine. So it looks like it was fixed.
>
> Hows this look:
>


LGTM.

-Eli
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20130620/2bb10992/attachment.html>