On Fri, Jan 27, 2012 at 5:11 PM, Steve Reinhardt <[email protected]> wrote:
> > > On Fri, Jan 27, 2012 at 9:31 AM, nathan binkert <[email protected]> wrote: > >> > This all looks good, with the one exception being that an SE compile now >> > takes 31% more user time. It's true that that's negligible wall clock >> time >> > on a large-scale MP, but not everyone always has access to that. I >> would >> > still really like us to figure out a way to make compilation more >> > configurable so we can leave out the components that no one needs in SE >> > mode (though ideally that's just one application of a more general >> > mechanism rather than a lingering hard-coded "SE vs FS" difference). >> >> I certainly hope that the 31% longer SE compile won't make us not >> choose to do this. I agree about the configurable thing, but we >> should go in this direction first. Most developers should be >> compiling *everything* regularly when they're running regressions, and >> this change should reduce that time significantly. >> > > I think there's a distinction between "developers" and "users"... I agree > that if you're working on core parts of the code then you want to test > everything so you'll be building all these pieces anyway. But if you're > more of a "user" just hacking on one little corner of the system to do your > own specific experiments, that's not the case. Ideally it wouldn't matter > much because you'd just compile the extra bits once and scons would never > make you compile it again, but we all know that's not how life works. > > In general I think we as core developers over-project our usage modes and > patterns onto what most people do with the code. > Just to be a little more specific: I'm not saying that we need to develop a general compile-time configuration mechanism before we merge these changes back in. I'm fine with something as simple as leaving the existing SE/FS distinction in the SConscripts, but having it such that the only distinction between the SE and FS builds is that the latter includes devices etc. (and whatever else constitutes the 31% extra work) but the former doesn't. We can rename FS to something else to indicate that it's a superset. Or we can rename them both. My discussion of the general compile-time mechanism is just to emphasize that the maintenance of this one aspect of the current SE/FS difference is a stopgap replacement for this alternate ideal, not a perpetuation of the now-meaningless distinction that Gabe has worked so hard to eradicate, so that he doesn't accuse me of bitterly clinging to the past... Steve _______________________________________________ gem5-dev mailing list [email protected] http://m5sim.org/mailman/listinfo/gem5-dev
