> H. J. Lu wrote: > > >On Fri, Jun 15, 2007 at 06:21:53PM -0700, Ian Lance Taylor wrote: > > > > > >>This is hardly a new thought, but I believe that for the i386 gcc is > >>handicapped by reload. No matter how smart we are before reload, it > >>just take one poor decision by reload in an inner loop and we've lost > >>all the gains. Reload has enormous complexities which are mostly > >>irrelevant for the i386. And I think that the idea of doing register > >>allocation separately from spill code generation does not make sense > >>on the i386. > >> > >> > >> > > > >Why don't we turn on vectorizer at -O3 or even -O2, depending on > >ISA? I added -ftree-vectorize to BOOT_CFLAGS on x86-64. According to > >-ftree-vectorizer-verbose=1, there are 82 loops vectorized in > >gcc source. There are no regressions. There are not much changes > >in bootstrap time as well as "make check" time. > > > > > > > A mount ago I did some measurements of the effect of the vectorizer for > Core2 in 64-bit mode. Unfortunately, I saw small improvement (as I > remember less than 1% for SPECInt2000). >
As far as I know there aren't many opportunities for vectorization in SPECInt2000 that have much impact on performance. The one exception is gzip that has 2 loops that when vectorized give ~8% improvement on the entire benchmark. We vectorized it and got this improvement with an old version of autovect-branch that had some hacks to remove redundant casts. If I remember correctly there's also a small hot initialization loop in eon that gets vectorized but doesn't improve performance because of the cost of creating the constant in a vector (unless inter-procedural constant propagation and loop specialization takes place). I think these two cases are pretty much it. > Vectorizer is a big a project and may be we will see more improvements > in future. They promissed implement SLP two years ago and now I see it > happens. It would be nice to see it not only in loops. > Do you have specific examples where SLP helps performance out of loops? thanks, dorit > Vlad >
