[llvm-commits] [llvm] r83324 - /llvm/trunk/lib/Analysis/MallocHelper.cpp

[Chris Lattner]

On Oct 6, 2009, at 12:12 AM, Duncan Sands wrote:

> Hi Chris,
>
>>>> +  // Check malloc prototype.
>>>> +  // FIXME: this will be obsolete when nobuiltin attribute will  
>>>> exist.
>>>
>>> why would it be obsolete then?  If LLVM sees "malloc" with a funky
>>> prototype it will always need to bail out, no?
>> Why would it in this case?  Though it doesn't make a lot of sense,  
>> it wouldn't cause the program to crash, and the code would have  
>> undefined behavior.
>
> if the standard malloc for some operating system has an unusual  
> prototype, for example it returns more info than usual malloc, then
> it is still a builtin, just a builtin that LLVM doesn't understand
> yet.  I don't see why it shouldn't be marked with a "builtin"  
> attribute.
> A "builtin" attribute presumably wouldn't be for marking functions
> that LLVM knows about, but for marking standard functions.  LLVM may
> or may not know about any particular function or particular standard.
> That said, I appreciate that there are limits to this line of  
> reasoning.

I think it only makes sense to consider something "builtin" if it  
follows the C standard.  You can't define a 'malloc' with a different  
prototype than what the spec indicates (at least without using -ffree- 
standing).

-Chris