[LLVMdev] unwinds to in the CFG

[Chris Lattner]

On Thu, 3 Apr 2008, Duncan Sands wrote:
>> bb1:
>>    %x = invoke i32 @f() to label %normal unwind label %unwind
>> normal:
>>    phi i32 [%x, %bb1]
>>    ret i32 %x
>> unwind:
>>    phi i32 [%x, %bb1]  ; illegal
>>    ret i32 %x
>>
>> The PHI in %unwind must mention %bb1, but may not refer to %x.
>
> In some sense this PHI is not dominated by its operands.  There
> is an edge to %unwind coming out of the invoke, but it is coming
> out of the RHS, i.e. before the assignment to %x.  If I wrote
> the invoke as:

On IRC, I suggested that Nicholas consider making the unwind block not be 
dominated by *any* instruction in the unwinding block.  This gets us out 
of weird partial dominance cases and makes the CFG checks simpler.  For 
example, this would be illegal:

foo:  unwinds to bar:
   %a = add ...
   call ...
   ...


bar:
   use %a


Because foo doesn't dominate bar at all.  In practice, unwind blocks (at 
least in C++) almost never refer to computation in the unwinding block, so 
I think that this is just fine in pratice.  For cases where a reference 
does exist, it is sufficient to split the block:


foo:
   %a = add ..
   br foo2

foo2: unwinds to bar
   call ...

bar:
   use %a

This means that simplifycfg would need to know that it can't merge 
foo/foo2, etc.

I think this would significantly simplify the dominance issues with this 
approach.  Maybe even the quiche eaters will be able to handle it! ;-)

What do you think?

-Chris

-- 
http://nondot.org/sabre/
http://llvm.org/