On Fri, Dec 8, 2023 at 8:19 AM Miguel Ojeda <miguel.ojeda.sandonis@xxxxxxxxx> wrote: > > On Fri, Dec 8, 2023 at 6:28 AM comex <comexk@xxxxxxxxx> wrote: > > > > Regarding the issue of wrappers not being inlined, it's possible to get LLVM to optimize C and Rust code together into an object file, with the help of a compatible Clang and LLD: > > > > @ rustc -O --emit llvm-bc a.rs > > @ clang --target=x86_64-unknown-linux-gnu -O2 -c -emit-llvm -o b.bc b.c > > @ ld.lld -r -o c.o a.bc b.bc > > > > Basically LTO but within the scope of a single object file. This would be redundant in cases where kernel-wide LTO is enabled. > > > > Using this approach might slow down compilation a bit due to needing to pass the LLVM bitcode between multiple commands, but probably not very much. > > > > Just chiming in as someone not involved in Rust for Linux but familiar with these tools. Perhaps this has been considered before and rejected for some reason; I wouldn’t know. > > Thanks comex for chiming in, much appreciated. > > Yeah, this is what we have been calling the "local-LTO hack" and it > was one of the possibilities we were considering for non-LTO kernel > builds for performance reasons originally. I don't recall who > originally suggested it in one of our meetings (Gary or Björn > perhaps). > > If LLVM folks think LLVM-wise nothing will break, then we are happy to On paper, nothing comes to mind. No promises though. >From a build system perspective, I'd rather just point users towards LTO if they have this concern. We support full and thin lto. This proposal would add a third variant for just rust drivers. Each variation on LTO has a maintenance cost and each have had their own distinct fun bugs in the past. Not sure an additional variant is worth the maintenance cost, even if it's technically feasible. > go ahead with that (since it also solves the performance side), but it > would be nice to know if it will always be OK to build like that, i.e. > I think Andreas actually tried it and it seemed to work and boot, but > the worry is whether there is something subtle that could have bad > codegen in the future. > > (We will also need to worry about GCC.) > > Cheers, > Miguel -- Thanks, ~Nick Desaulniers
