> Richard, > Will gcc 4.3.0's release be held up until all of the major > architectures have fully optimized cost models for vectorization? > I ask because as far as I can tell the powerpc cost model changes > haven't been submitted yet.
At this point it doesn't look like there will be any cost model changes submitted for powerpc - experimentation so far on powerpc970 with Spec2006 and Polyhedron benchmark suites didn't bring up any issues that needed to be fixed (in both suites vectorization either had no impact (on most benchmarks), or had a significant positive impact (on 1-2 benchmarks); the cost model had no impact). We are continuing to experiment on power, also with other benchmarks, in search for cases that do require cost model tuning, but there are no patches ready at the moment. There are however a couple of small cost-model changes that were going to be submitted this week for the Cell SPU - it's unfortunate if these cannot get into 4.3. dorit (P.S. thanks to Razya for powerpc Polyhedron testing) > It certainly would be nice if all > of the major targets could have -fvect-cost-model enabled by > default for -O3. > Jack > > On Wed, Jan 23, 2008 at 12:06:22PM +0100, Richard Guenther wrote: > > > > As we now reached the goal of less than 100 open serious regressions > > against GCC 4.3, we are as of now in regression and documentation fixes > > only mode. This means that for patches going on the trunk the same > > rules as for release branches apply. > > > > The next milestone before the release of GCC 4.3.0 is to get down > > the priority one (P1) regressions against GCC 4.3 down to zero. At > > the point we reach that goal, either by fixing all P1 bugs or by > > downgrading less important ones to P2, the mainline will be freezed > > to prepare for a release candidate. Around the same time we will > > branch and the opening of stage1 for GCC 4.4 development will be > > announced. > > > > There are 5 P1 bugs open against GCC 4.3, one of it has patches, > > three of them are C++ regressions assigned to Mark and one bug > > does not seem likely to be fixed (PR31529), which makes it a likely > > candidate for downgrading. > > > > I will update the GCC frontpage with the trunk state asap. > > > > Thanks, > > Richard.
