exploit independences during dependence analysis
commitc63f95d222df046327ecac450a58273652639a21
authorSven Verdoolaege <skimo@kotnet.org>
Thu, 2 Apr 2015 14:33:58 +0000 (2 16:33 +0200)
committerSven Verdoolaege <skimo@kotnet.org>
Fri, 24 Apr 2015 08:37:03 +0000 (24 10:37 +0200)
treeac9ae3a1d34b889b22c751740a9dcd0e2a91c4e8
parentcc7fe95981686bdfdc6b714682e01a7241d2c1aa
exploit independences during dependence analysis

In e57e679 (gpu.c: compute_schedule: exploit independences,
Mon Feb 3 17:57:58 2014 +0100), we started exploiting independences,
but only very conservatively.  In particular, we would only use
them to allow more schedule dimensions to be considered parallel and
not to create more freedom for the scheduler.

Also exploit the independences during dependence analysis,
which does create more freedom for the scheduler.
We still only exploit independences when live-range reordering
is enabled since the live-range reordering takes care
of ensuring that the live-ranges of local variables will not overlap.

Signed-off-by: Sven Verdoolaege <skimo@kotnet.org>
ppcg.c