[PATCH] [LoopAccesses] Cache the result of canVectorizeMemory

[hfinkel at anl.gov]

In http://reviews.llvm.org/D7682#124926, @anemet wrote:

> In http://reviews.llvm.org/D7682#124591, @anemet wrote:
>
> > In http://reviews.llvm.org/D7682#124581, @hfinkel wrote:
> >
> > > Why not do the analysis in the constructor, and remove the implicit ordering dependency between calling analyzeLoop and calling the other accessor functions?
> >
> >
> > I think that should work, let me give it a try.
>
>
> OK, I know now why we can't do it in *this* patch.  LAI is constructed when LVLegality is constructed so we would effectively move running canVectorizeMemory much earlier than before.
>
> On the other hand I can do this after we have the LAA-LAI separation later in http://reviews.llvm.org/D7684 (where the pass is actually created).  There, LAI is only created from LVLegality::canVectorizeMem so there is no functional change.  Are you OK with that?


Yes, that's fine with me. Go ahead with this, it will last for a few minutes ;) LGTM.


http://reviews.llvm.org/D7682

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/