On Wed, 23 Sep 2020 at 08:17, Yonghong Song <yhs@xxxxxx> wrote: > > Hi, > > I have spent some time to add additional logic in llvm BPF backend > in order to generate verifier friendly code. > > The first patch is: > https://reviews.llvm.org/D87153 > which moves CORE relocation builtin handling from in late IR > optimization (after inlining and major optimizations) > to in early IR optimization (before inlining and any optimizations). > The reason is to prevent harmful CSEs. > > But this change may change how compiler do optimizations. > The patch can pass bpf selftests in latest bpf-next. > Andrii helped it can also pass bcc/libbpf-tools. > > If your code uses COREs, esp. having a lot of subroutines > and/or loops, it would be good to give a try with new patch > to see whether there are any issues or not. In my case, > for one of our internal applications with lots of subroutines > and loops, inlining all subroutines and unrolling all loops > will cause register spills which cannot be handled by > the verifier, while existing llvm won't have issue. Hi Yonghong, We currently don't use CORE (outside of bcc, etc.), so there isn't much I can test I guess? Please let me know if there is something I can do for your follow up patches. Best Lorenz -- Lorenz Bauer | Systems Engineer 6th Floor, County Hall/The Riverside Building, SE1 7PB, UK www.cloudflare.com
