On 05/22/2013 01:13 PM, Iyer, Balaji V wrote:
Hi Jeff, Yes, converting the array notation expansion to #pragma simd
(or #pragma omp simd) trees will be beneficial performance wise. But,
it will require semi-significant re-write and this can destabilize a
currently stable implementation. IMHO, for now I would like to keep
array notation as is and let us see how it interacts with other
flavors of gcc backends. At the moment, we do not see any reason for
instability in this, but we have not done any significant testing on
non x86 backends. After OMP 4 hits the main-line and everything is
stable, we could proceed as you indicated.
I was thinking more from maintenance standpoint. It seems kindof silly
to have multiple blobs of code generating loops. The fact that it gets
us vectorized code more often is just icing on the cake.
This is certainly something that I want to see happen, though I'll defer
to Richard, Joseph, Aldy, etc on the timing of that change.
jeff
