On Tue, Apr 09, 2024 at 12:50:15PM +0200, Peter Zijlstra wrote: > On Mon, Mar 25, 2024 at 01:59:55PM -0700, Boqun Feng wrote: > > On Mon, Mar 25, 2024 at 10:44:45AM +0000, Mark Rutland wrote: > > [...] > > > > > > > > * I choose to re-implement atomics in Rust `asm` because we are still > > > > figuring out how we can make it easy and maintainable for Rust to call > > > > a C function _inlinely_ (Gary makes some progress [2]). Otherwise, > > > > atomic primitives would be function calls, and that can be performance > > > > bottleneck in a few cases. > > > > > > I don't think we want to maintain two copies of each architecture's atomics. > > > This gets painful very quickly (e.g. as arm64's atomics get patched between > > > LL/SC and LSE forms). > > > > > > > No argument here ;-) > > Didn't we talk about bindgen being able to convert inline C functions > into equivalent inline Rust functions? ISTR that getting stuck on Rust Yes, we did. > not having a useful inline asm. > Mostly two features were missing: 1) asm goto and 2) memory operands, #1 gets implemented[1] by Gary, and should be available in Rust 1.78 (plan to release at May 2, 2024); For #2, my understanding is that arch-specific effort is needed (since different architectures may have different contraints on memory operands), I haven't yet found anyone is working on this. (background explanation for broader audience: in GCC's inline asm, you can specify an memory location, other than a register location, as an input or output of an asm block's operand[2], but current Rust inline asm doesn't provide this functionality, by default, without option "pure", "nomem", etc, every asm block in Rust can be thought as a C asm block with "memory" clobber) That being said, if you look at the link I shared or this gist from Gary: https://gist.github.com/nbdd0121/d4bf7dd7f9b6d6b50fa18b1092f45a3c there is another way (yeah, we probably also have discussed this previously), basically what it does is compiling the functions in a C file as LLVM IR, so that Rust can call these functions at LLVM IR level. This in theory is doing some local LTO, and I've tested that it works for asm blocks. We still need to tweak our build system to make this work, but should it work, it would mean that Rust can call a C function in a pretty efficient way. > But fixing all that in a hurry seems like the much saner path forward. So a sane plan to me is wiring our atomics into Rust functions via what Mark has (i.e. starting off as FFI calls), then we can switch to the "local LTO" approach when it's ready. In case that "local LTO" needs more work and we do have performance need, we can always either 1) manually implement some primitives in Rust asm, or 2) look into how bindgen or other tools can translate simple C functions (asm blocks) into Rust. Regards, Boqun [1]: https://github.com/rust-lang/rust/pull/119365 [2]: https://gcc.gnu.org/onlinedocs/gcc/Simple-Constraints.html#index-m-in-constraint
