Hi Marco, Thank you for initiating the discussion. You mentioned concerns with checking in your patch (http://homepages.inf.ed.ac.uk/s0787712/res/mcversi_gem5.patch). Why do you think it needs more work? You already have two ship its for your Ruby changes (review 3398). I believe most of aspects of your larger patch should not be controversial. For instance I would be disappointed if folks gave you a hard time for adding new pseudo instructions. Perhaps some will dislike the additions to the O3 model, but personally they seem pretty minor. It would probably help if all your changes to the O3 code included comments on how they are used and why they are needed.
I would suggest posting the entire patch to reviewboard. As I mentioned to you in Barcelona, eventually it would be great to add at least long regression tests that used it. There is precedent to have long regression tests rely on code in the ext/ directory. Thanks, Brad -----Original Message----- From: gem5-dev [mailto:[email protected]] On Behalf Of Andreas Hansson Sent: Tuesday, April 12, 2016 3:08 PM To: gem5 Developer List <[email protected]> Subject: Re: [gem5-dev] Memory consistency verification (and some bugs) Hi Marco, This sounds good. Regarding the external component, I certainly do not mind. I would suggest we prune the old and crufty McPat. That should free up some 45k LoC, and I there is not really any integration in place in any case. Do you also have any updates to the MemChecker? It is currently very limited in that the memory usage grows very quickly. Could you perhaps shed some light on how this all fits together? Thanks, Andreas On 12/04/2016, 22:57, "gem5-dev on behalf of Marco Elver" <[email protected] on behalf of [email protected]> wrote: >Hi all, > >I've been working on integrating McVerSi [1] (a framework for >simulation-based memory consistency verification) into the latest Gem5 >version; in doing so I have also discovered new bugs (see end of email). > >At [1] I provide a patch and instructions on how to use McVerSi with >the current Gem5 version. There is a chance that (due to differing >checkpoints, timing, etc.) other bugs may be discovered than the >mentioned ones here. > >Ultimately, it would be good to have this in upstream, but currently >doesn't appear clear to me what the best approach for this is. It was >raised at HPCA (during discussion with Brad and David), that it might >be nice to have for regression testing. Comments would be appreciated. > >The patch at [1] needs some work to be suitable to be accepted into >upstream; does it still make sense to upload to RB to discuss? I am >also unsure if Gem5 currently provides a clean way to implement the >hooks required, and would rather not need to change other subsystems. > >(A small portion of the patch at [1] is already on the RB: >http://reviews.gem5.org/r/3398/) > >Furthermore, it requires a substantial external component >(https://github.com/melver/mc2lib) to live in ext/; should this be put >onto RB at the same time? > >Many thanks, > >-- >Marco Elver >Research Associate | University of Edinburgh | ac.marcoelver.com > >[1] http://ac.marcoelver.com/research/mcversi > > >The following are new bugs found; neither look particularly easy to fix: > >1. In the classic memory system, same address read->read reordering (SC >PER LOCATION violation) can occur with O3CPU (found while using ARM >ISA, but probably affects all other ISAs requiring at least same >address >read->read ordering): > >Consider the case of a replacement (Cache::allocateBlock), which >chooses a block corresponding to a prior request that has already been >satisfied, i.e. no longer in MSHR. However, the request is still not >committed in the LQ on CPU side, and if we don't send an invalidation >to the CPU side on a replacement we may miss any potential external >invalidations after. This bug is found after running for about 30 min. > >2. In O3CPU LSQ, write->read reordering (SC PER LOCATION violation) can >occur in the same thread when using virtual address synonyms (found >with >X86+Ruby, but will affect all other configurations that use the O3CPU): > >When forwarding in-flight writes from the SQ, only virtual addresses >are checked; if two virtual addresses map to the same physical address, >no forwarding takes place. I realize this might be a rare use-case for >the kinds of workloads being run on Gem5 right now? Found with first >generated test, i.e. instantly. > >-- >The University of Edinburgh is a charitable body, registered in >Scotland, with registration number SC005336. > >_______________________________________________ >gem5-dev mailing list >[email protected] >http://m5sim.org/mailman/listinfo/gem5-dev IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you. _______________________________________________ gem5-dev mailing list [email protected] http://m5sim.org/mailman/listinfo/gem5-dev _______________________________________________ gem5-dev mailing list [email protected] http://m5sim.org/mailman/listinfo/gem5-dev
