llvm dev - Sep 2011 - [LLVMdev] question on difference of bitcode between C and C++

On 22 September 2011 03:30, Eli Friedman <eli.friedman at gmail.com>
wrote:
>> I am trying to find such difference of bitcode between C and C++.
>
> There isn't any difference in that sense... in IR, a constructor is
> just a function call, a reference is just a pointer, etc.
Hi Fei,

While Clang (like others) lowers C++ into C semantics and lower that
into IR, there are some changes that exist in C++ and not in C. The IR
has the same features, but some assumptions on the semantics are
different.

I can give you two examples:

1. Classes in C++ are like C structures in IR, but the C++ ABI makes
it difficult to express Base/Derived classes in pure structures.
(http://www.systemcall.org/blog/2011/01/cpp-class-sizes/ and
http://www.systemcall.org/blog/2011/03/cpp-class-sizes-2/).

So, if your pass depends on identifying the same types, you could end
up thinking that the types are different, but they're not. They're
just different struct representations (base vs. derived) of the same
type.

2. Virtual table tables encode offsets in two different ways:
addresses and offsets, and the two representations are normally on the
same global static array in IR. So, while the type of the array is int
(or pointer), it contains addresses and offsets from addresses
bitcasting to the type of the global.

These are not the only C-lowering that C++ front-ends do, but it gives
you a taste of what to expect. As Eli said, most of C++ can be lowered
into C-like structures and the IR is very similar, but some semantic
interpretations are done differently, and the internal IR (that deals
with those objects) is slightly different.


-- 
cheers,
--renato

http://systemcall.org/

Another difference is the presence of exceptions in C++, which would require you
to handle more IR instructions. This might not matter depending on type of
analysis you do.
See: http://llvm.org/docs/LangRef.html#i_invoke
(Note that there is a substantial rewrite of exception handling going into 3.0)

Anna.
On Sep 22, 2011, at 3:29 AM, Renato Golin wrote:
> On 22 September 2011 03:30, Eli Friedman <eli.friedman at gmail.com>
wrote:
>>> I am trying to find such difference of bitcode between C and C++.
>> 
>> There isn't any difference in that sense... in IR, a constructor is
>> just a function call, a reference is just a pointer, etc.
> 
> Hi Fei,
> 
> While Clang (like others) lowers C++ into C semantics and lower that
> into IR, there are some changes that exist in C++ and not in C. The IR
> has the same features, but some assumptions on the semantics are
> different.
> 
> I can give you two examples:
> 
> 1. Classes in C++ are like C structures in IR, but the C++ ABI makes
> it difficult to express Base/Derived classes in pure structures.
> (http://www.systemcall.org/blog/2011/01/cpp-class-sizes/ and
> http://www.systemcall.org/blog/2011/03/cpp-class-sizes-2/).
> 
> So, if your pass depends on identifying the same types, you could end
> up thinking that the types are different, but they're not. They're
> just different struct representations (base vs. derived) of the same
> type.
> 
> 2. Virtual table tables encode offsets in two different ways:
> addresses and offsets, and the two representations are normally on the
> same global static array in IR. So, while the type of the array is int
> (or pointer), it contains addresses and offsets from addresses
> bitcasting to the type of the global.
> 
> These are not the only C-lowering that C++ front-ends do, but it gives
> you a taste of what to expect. As Eli said, most of C++ can be lowered
> into C-like structures and the IR is very similar, but some semantic
> interpretations are done differently, and the internal IR (that deals
> with those objects) is slightly different.
> 
> 
> -- 
> cheers,
> --renato
> 
> http://systemcall.org/
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20110922/9610829d/attachment.html>

Thanks! Your suggestions are really helpful to me. I am verifying the
differences and will provide my findings soon.

--
Best Regards,
Fei Jia




On Thu, Sep 22, 2011 at 2:33 PM, Anna Zaks <ganna at apple.com> wrote:
> Another difference is the presence of exceptions in C++, which would
> require you to handle more IR instructions. This might not matter depending
> on type of analysis you do.
> See: http://llvm.org/docs/LangRef.html#i_invoke
> (Note that there is a substantial rewrite of exception handling going into
> 3.0)
>
> Anna.
> On Sep 22, 2011, at 3:29 AM, Renato Golin wrote:
>
> On 22 September 2011 03:30, Eli Friedman <eli.friedman at gmail.com>
wrote:
>
> I am trying to find such difference of bitcode between C and C++.
>
>
> There isn't any difference in that sense... in IR, a constructor is
>
> just a function call, a reference is just a pointer, etc.
>
>
> Hi Fei,
>
> While Clang (like others) lowers C++ into C semantics and lower that
> into IR, there are some changes that exist in C++ and not in C. The IR
> has the same features, but some assumptions on the semantics are
> different.
>
> I can give you two examples:
>
> 1. Classes in C++ are like C structures in IR, but the C++ ABI makes
> it difficult to express Base/Derived classes in pure structures.
> (http://www.systemcall.org/blog/2011/01/cpp-class-sizes/ and
> http://www.systemcall.org/blog/2011/03/cpp-class-sizes-2/).
>
> So, if your pass depends on identifying the same types, you could end
> up thinking that the types are different, but they're not. They're
> just different struct representations (base vs. derived) of the same
> type.
>
> 2. Virtual table tables encode offsets in two different ways:
> addresses and offsets, and the two representations are normally on the
> same global static array in IR. So, while the type of the array is int
> (or pointer), it contains addresses and offsets from addresses
> bitcasting to the type of the global.
>
> These are not the only C-lowering that C++ front-ends do, but it gives
> you a taste of what to expect. As Eli said, most of C++ can be lowered
> into C-like structures and the IR is very similar, but some semantic
> interpretations are done differently, and the internal IR (that deals
> with those objects) is slightly different.
>
>
> --
> cheers,
> --renato
>
> http://systemcall.org/
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20110922/82408238/attachment.html>