[llvm-commits] [PATCH] Added topological graph sort routine to llvm/ADT

[Delesley Hutchins]

AFAICT, the RPO iterator does not do what we need.  For example,
assume we have the following graph (pardon the bad ASCII graphics):

   /---> (2) ----------\
(1)                     --> (5)
   \---> (3) ---> (4)--/

I believe the RPO iterator will traverse this graph as 5-2-4-3-1,
which is a standard depth-first post-order traversal.

What we want is a topological traversal, which is 1-2-3-4-5; every
predecessor of a node must be traversed before the node itself.
Notice that a topological traversal is neither depth-first (that would
be 1-2-5-3-4), nor breadth-first (1-2-3-5-4); we want to ensure that 2
and 4 are traversed before 5.

This patch does three things to achieve the topological traversal:
(1) Depth-first traversal to identify back-edges and assign depths.
(2) Sort nodes according to depth.
(3) Traversal of the sorted list of nodes.

  -DeLesley

On Wed, Aug 17, 2011 at 10:00 PM, Jakob Stoklund Olesen <stoklund at 2pi.dk> wrote:
>
> On Aug 17, 2011, at 9:17 PM, Chandler Carruth wrote:
>
> On Wed, Aug 17, 2011 at 8:56 PM, Jakob Stoklund Olesen <stoklund at 2pi.dk>
> wrote:
>>
>> On Aug 17, 2011, at 3:52 PM, Delesley Hutchins wrote:
>>
>> > This patch adds a topological sort routine for indexed graphs to
>> > llvm/ADT.  This sort routine is currently used to traverse CFGs in
>> > Clang when doing thread safety analysis (patch to be released
>> > shortly).
>>
>> Could you explain why we need this in addition to
>> include/llvm/ADT/PostOrderIterator.h?
>
> FYI, I'm not terribly familiar w/ PostOrderIterator (so maybe it would
> actually serve here), but JSYK, the motivating use case is in a Clang
> patch: http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20110815/045262.html
>
> I see.
> AFAICT, the existing RPO iterator will give you the exact same ordering of
> blocks, at least it will give you a topo-sort of the graph after removing
> back-edges.
> You can detect back-edges by keeping track of visited blocks in a BitVector
> as you iterate through the RPO.
> /jakob
>



-- 
DeLesley Hutchins | Software Engineer | delesley at google.com | 505-206-0315