add support for building outlined aarch64 intrinsics #407
Conversation
llvm/llvm-project@a4ac434 saw the addition of out-of-line aarch64 atomic intrinsics. LLVM will sometimes emit these, so we need to ensure they're included in Rust's compiler-rt.
|
The c feature is disabled by default. A rust port of these would be nice? That would probably require |
| // Ideally, this would be a Vec of object files, but cc doesn't make it *entirely* | ||
| // trivial to build an individual object. | ||
| let mut atomics_libraries = Vec::new(); | ||
| for instruction_type in &["cas", "cwp", "ldadd", "ldclr", "ldeor", "ldset"] { |
There was a problem hiding this comment.
Should this be swp instead of cwp?
Actually it is enabled by default in |
|
This turns out to be an issue even without LLVM emitting references to these intrinsics. If rustc links in a C static library that was built to expect these outlined intrinsics, and it doesn't link in libgcc (which it doesn't on aarch64 musl), that'll result in a build error due to these missing symbols. It'd be helpful to have these symbols available even before we generate code using them ourselves. |
|
Closed in favor of #416 . |
llvm/llvm-project@a4ac434 saw the
addition of out-of-line aarch64 atomic intrinsics. LLVM will sometimes
emit these, so we need to ensure they're included in Rust's compiler-rt.
===
This probably isn't an issue that we'll see in ToT Rust until we pull in a new LLVM. Seems reasonable to me to hold off on merging this PR until then; I just had to do it anyway, so thought I'd provide the result in case it's helpful. :)
The approach here is admittedly ugly. If we want to stick more closely with the "all builds are done by
cfg.compile("libcompiler-rt.a");," a potentially cleaner alternative would be to generate a temp .S file for each combination of these things. These would boil down to something like:LLVM's cmake approach (AIUI) is "create a lot of symlinks to lse.S, and compile those with differing flags." Adopting that approach would still give us badness here, since it doesn't seem
cclets us apply file-specific flags.I'm still doing a bit of extra testing to ensure this does the right thing, but aarch64 builds successfully for me with a local backport of the recent lse.S changes.