[LLVMdev] Folding cast of a global address to boolean

[Chris Lattner]

On Thu, 6 Jan 2005, Morten Ofstad wrote:

> I noticed that folding a cast of a global address to boolean doesn't work 
> like I expected - there is a comment in the code that casts of external 
> symbols should not happen, but I don't understand the rationale for this.
>
> // FIXME: When we support 'external weak' references, we have to prevent
> // this transformation from happening.  In the meantime we avoid folding
> // any cast of an external symbol.
>    if (!GV->isExternal())
>      return ConstantBool::True;

The basic idea is that we currently do not support globals with "external 
weak" linkage, but we will someday.  The semantics of an external weak 
symbol are somewhat interesting: if a definition of the symbol is ever 
linked into the program, the symbol gets that address (obviously).  If a 
definition is never linked into the program, all uses of the global get a 
null pointer instead.

This is often used in situations where you want to have code conditional 
on whether the library is linked into a pthreads-enabled program.  It 
allows you to do stuff like this:

"external weak" pthread_create(...);

...
   if (pthread_create) {
     // Linked to -lpthread.
   }

If a body for pthread_create is linked into the program, the conditional 
will evaluate to true.

> This particular transformation is quite important for me because I have some 
> symbols which are either ConstantPointerNull or a GlovalVariable depending on 
> which context the program is compiled in. This is used to implement a kind of 
> conditional compilation, and I'm relying on the constant folding to remove 
> branches in my code. I tried to change the linkage type of the GlobalValue, 
> but that's not what the isExternal() function checks so it did not have any 
> effect.

Yeah, external functions (those without a body) must have external linkage 
right now.

> If I understood better why this is disabled, maybe I can fix it. Otherwise I 
> will have to implement the conditional compilation as part of the code 
> generation and that will be quite a bit more complicated.

There is no good reason for it to be disabled, so I enabled it.  When we 
add support for "external weak", we should add a check in that code path 
to make sure the global is not external weak.  Until then, we shouldn't 
pessimize code.

http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20050103/022961.html

-Chris

-- 
http://nondot.org/sabre/
http://llvm.cs.uiuc.edu/