[llvm-commits] SIV tests in LoopDependence Analysis, Sanjoy's patch

[Hal Finkel]

On Fri, 16 Mar 2012 23:43:19 +0530
Sanjoy Das <sanjoy at playingwithpointers.com> wrote:

> Hi Preston, Hal,
> 
> I do have commit access, but I don't think the work is ready for
> merging.  It does feel good to get positive feedback, of course. :)
> 
> The issue is, the dependence analysis does not calculate direction
> vectors.  And, without direction vectors, it isn't possible to combine
> analysis of separate subscripts (in a multi-dimensional array) into an
> analysis of the array access as a whole.
> 
> For instance, the access in the loop
> 
> for (i = 0; i < N; i++)
>   for (j = 0; j < N; j++)
>     array[i + 1][j - 1] = array[i][j] + C
> 
> is (according to the code I wrote) dependent on the first subscript
> but independent on the second, which is plain wrong since that way the
> entire access works out to be independent.
> 
> I don't think individual subscript analysis should even return a
> Independent / Dependent result.  Instead, it should return a subset of
> { <, =, >, Provably independent, Unknown }.  That way, the dependence
> direction in the above loop comes out to be [<, >] (and hence
> dependent).  Breaking conditions also need to be worked in and all
> this needs to be public since loop transformations will need this
> information.
> 
> I'm already working towards this, and will probably have to change the
> interface LDA exposes.  Hopefully I will have some code ready by this
> weekend.

Sounds great, I look forward to seeing it.

 -Hal

> 
> This is the first time I'm working on a analysis pass proper, and I
> really don't want to commit half-baked code. :) Meanwhile, I'll push
> my work here: https://github.com/sanjoy/llvm/tree/lda
> 
> Thanks!



-- 
Hal Finkel
Postdoctoral Appointee
Leadership Computing Facility
Argonne National Laboratory