[LLVMdev] Indirect function call

[Andrew Lenharth]

On Tue, 2006-05-23 at 13:32 +0800, Nai Xia wrote:
> On Monday 22 May 2006 22:22, Andrew Lenharth wrote:
> > On Mon, 2006-05-22 at 15:33 +0800, 澶忎竴姘� wrote:
> > > But my code does not always works: if the arguments are not pointer,
> > > CompleteBUDataStructures not records it. So, if you want to find all indirect
> > > calls, you maybe have to repair CompleteBUDataStructures. :)
> > 
> > Not surprising, CBU is trying to do something entirely different that
> > what you are.
> > 
> > > If you do not use BUDataStructures, you can do it yourself: find all load/store
> > > instructions with its destination is function type.
> > 
> > You may want to look at how the call graph builder works.  It finds all
> > indirect call sites, and also finds all functions whose address escapes
> > (this is, may be called indirectly).
> 
>   The BasicCallGraph class only lines out the indirect calls(makes the caller point to external node), 
> but do not resolves them using alias analysis such as DSA. 
> I think DSA solve this problem for interested call sites by finding the corresponding globals(i.e. the functions) 
> for the callsite DSnode. Maybe 夏一民 just wanted to point out that DSA does not take all callsite into count.
> But just as suggested in callgraph.h, "As an extension in the future, there may be multiple nodes with a null
> function.  These will be used when we can prove (through pointer analysis) that an indirect call site can 
> call only a specific set of functions." 

First, if you want call site information, you need TD not BU.  Second,
TD still isn't perfect.  I have a series of patches that improve DSA's
indirect call handling, but they are ugly and not yet ready for
incorporation into mainline.  If you really want to try them I can send
you patches, or you can check out my monotone tree.

I also have some clients of the TD call graph that make use of it (well
make use of the Global list in the function pointer's DSNode), such as a
devirtualizer.  Again, I can send you a copy or you can checkout my
tree.

Andrew