* Numbers for SWIM with -ftree-loop-linear
@ 2004-09-14 2:15 Daniel Berlin
0 siblings, 0 replies; only message in thread
From: Daniel Berlin @ 2004-09-14 2:15 UTC (permalink / raw)
To: gcc
<pinskia> DannyB: at -O1 -ftree-loop-linear: from ??? (I don't have
numbers for -O1) to 1111 (I will test with higher optimizations soon)
<pinskia> for -O3 -funroll-loops -mcpu=G5 -ffast-math -ftree-loop-linear:
288 to 1260
That's in terms of spec score on a g5 running darwin 10.3.4.
So uh, it helps a bit.
To be fair, swim's inner loop accesses arrays with a stride of 1336, and
with -ftree-loop-linear, it's now stride 1.
:)
I've also determined why running ivopts first on g95 doesn't help (g95
doesn't use the original specified induction variables as the normal ivs,
but ivopts should eliminate one in favor of the other).
ivopts canonicalizes the iv properly, but it also does tree strength
reduction, and replaces an array reference with a pointer deref,
which we can't analyze data dependence for, and give up.
It would be nice if ivopts was split into strength reduction, and
induction variable elimination, seperately.
So that way, we could run IVE before linear loop transforms, and strength
reduction after.
--Dan
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2004-09-14 1:55 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2004-09-14 2:15 Numbers for SWIM with -ftree-loop-linear Daniel Berlin
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).