Hi folks, specifically ARM folks. We've been seeing a problem with O3 where when switching vector register renaming modes (full vectors vs vector elements), the CPU checks its bookkeeping and finds that a vector register is missing, ie with no instructions in flight, the free list has one fewer register in it than the difference between the total number of physical vector registers, and the number that should be taken up with architectural state.
This problem has been somewhat difficult to reproduce, although we can get it to happen, and it does happen often enough that it's been a real pain for us. Given that it's not very easy to get it to happen which makes it hard to observe, I've been digging around in the code trying to understand what all the pieces do and why the bookkeeping might be wrong. The most promising thing I've found so far is that when squashing, the rename stage looks at its history and rolls back renames for squashed instructions. Some registers are fixed and not renamed, so rolling back those would be pointless. Also those registers should not go on the free list. The way O3 detects those special registers is that they have the same index before and after renaming. If that is the case, O3 ignores those entries, and does not roll them back or mark their target as free. This check is slightly out of date though, since with the recently added pinned register writes, a register will be renamed to the same thing several times in a row. When these entries are checked, they will not be rolled back (I think this part is still fine), but they will also not be marked as free. This isn't exactly a smoking gun though, since the more I think about it, the more I think this may actually be ok. If one of the later writes is squashed, the register isn't "free" since it still holds the (partially written) architectural state. If everything gets squashed all the way back to the first entry which did change what register to use, then the slightly outdated check won't trigger and things should be freed up correctly (I think). This code is mostly new to me though, so I'm not super confident making any grand declarations about what's going on. All the pieces seem to be there though, which makes me very suspicious. Maybe something goes wrong if the right number of writes never happens because later writers get squashed? Gabe
_______________________________________________ gem5-dev mailing list -- email@example.com To unsubscribe send an email to gem5-dev-le...@gem5.org %(web_page_url)slistinfo%(cgiext)s/%(_internal_name)s