[llvm-commits] [llvm] r89434 - /llvm/trunk/lib/Analysis/CaptureTracking.cpp

[Chris Lattner]

On Nov 20, 2009, at 10:04 AM, Duncan Sands wrote:

> Hi Dan,
> 
>> Hi Duncan, thanks for pointing these out. I removed the same-object rule,
>> which wasn't essential to my original testcase, and refined the null
>> comparison rule to only apply to noalias return values, which is still
>> sufficient for the original testcase.
> 
> sorry for annoying you with all these crazy examples!  Personally I would
> be happy with a notion of pointer capture based on alias analysis (i.e.
> a pointer escapes if there is some other pointer which might alias it),
> with rules saying that dissecting pointers and reconstructing them from
> the pieces "doesn't count", i.e. you are allowed to say that the new one
> doesn't alias the original, and thus that the original was not captured.
> But Chris rejected this - perhaps we should gang up on him? :)

The only case I'm really concerned about are things like PointerIntPair that bitmangle things into the low bits of pointers.

-Chris