[cfe-dev] Fixing selector types on the GNU runtime

[David Chisnall]

On 21 Jan 2009, at 03:14, Mike Stump wrote:

> So, what are you interested in, the theoretics of hard core
> pedantictry or the engineering realities of having FreeBSD be warning
> free?

I'm interested in fixing the warning only in as far as the warning  
indicates potentially unsafe code.  The code, as written and according  
to the language specification, says pass a zero integer value of some  
indeterminate size to terminate the variadic function. The callee  
expects a zero pointer value.  This is categorically wrong.  If some  
compilers on some architectures do the right thing, then that's a nice  
bonus.  On x86/FreeBSD, the compiler does the right thing (since NULL  
is a pointer-sized int) but it warns that this is a coincidence.

The #ifdef block you suggest makes the warning go away on FreeBSD/x86,  
but this is not a platform where the real problem actually occurs,  
just one which notices that the code is wrong and issues a warning.   
Warning that the code does not say what you think it does is not a  
bug, it is a (useful) feature.

The diff I submitted does the right thing, according to the standard.   
If this is not what you want, then please update the coding  
conventions to specify this exemption.  As it stands, my diff will  
avoid real issues on platforms like Win64, where 0 and 0L, the two  
suggested definitions of NULL from the standard, are both smaller than  
pointers.  I have variadic function calls in this file with more  
pointer arguments than there are integer registers on x86-64 for  
argument passing, so the 0 will be passed on the stack on this  
platform and not extended to 64-bits.

Honestly, the amount of effort people have put in to arguing against  
fixing this one file is less than the amount of effort it would have  
taken to make the change everywhere in LLVM.  I put these changes in  
with the fixing the type conversion diff because I assumed there would  
be no debate about fixing incorrect code.  It seems I was wrong, so  
have attached a diff that only fixes the type conversion error, and  
leaves the unsafe variadic function calls in.

David
-------------- next part --------------
A non-text attachment was scrubbed...
Name: clang.diff
Type: application/octet-stream
Size: 1386 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20090121/9efccadc/attachment.obj>
-------------- next part --------------