[Polly] Allow multiple reductions in one statement

Tobias Grosser tobias at grosser.es
Fri Jun 27 12:30:54 PDT 2014


On 27/06/2014 21:04, Johannes Doerfert wrote:
> I will change the style/typos before I commit it.

Good.

>
> What about A[i] = A[i] + A[i]?
> We will not mark it as reduction like at the moment and that is a good thing (in my opinion).
> Case 1: The reads are different LLVM-IR Values:
> 	The "collectPossibleLoads" for the A[i] store will return both A[i] loads, but if we test for non overlapping accesses will always find the other load as overlapping => Invalid
> Case 2: The reads are the same LLVM-IR Value:
> 	The "collectPossibleLoads" won't return this load ad it has at least two uses (in the binary operator).

OK. I did not see this. Maybe its worth a test case.

> I will remove the XFAIL and make it not detect the reduction/scop.

Good.

> I will introduce unary reductions in 2 commits (hopefully). I will simply remove the FIXME's for now if they are confusing.

Good.

> What do you think?

LGTM.

Tobias



More information about the llvm-commits mailing list