Compiling against libcore with a soft-float target appears to require sin/cos/sqrt since recently

[alex]

Since #51695 (I believe), the project I have for compiling linux kernel modules has started failing due to sin/cos/sqrt symbols not being available: fishinabarrel/linux-kernel-module-rust#61 (comment)

This uses a custom target file which is x86_64 with +soft-float.

I believe the issue is that stdsimd uses the llvm.sin.* intrinsics, and LLVM will emit a call to libm if the platform is soft-float (this is speculation, but well founded I think :-)).

Being able to depend on libcore without needing any additional runtime libraries is a hard requirement for compiling for targets like kernel modules. I'm not sure what the right direction for resolving this is.

[gnzlbg]

I think we just have to fix this in stdsimd, by requiring those symbols only when used through std and not when used through core. @alexcrichton thoughts?

To test this upstream, maybe we could try to add some soft float targets that reproduce the issue, and then commit a fix. We will probably need a stdbuild feature in stdsimd (like the libc one) to detect when it is being built as part of std or core.

---

Or in other words, the real issue here is that floats are part of the language, but some operations on float require library support on some targets, and therefore those operations cannot be part of core.

[alexcrichton]

I believe this isn't the first of its kind for libm intrinsics in libcore, but the general story around them is "if you use --gc-sections and don't use the code that uses the intrinsics it doesn't matter" . @alex do you know if you're compiling with --gc-sections and/or enabling that sort of flag in the linker?

[alex]

We're not using gc-sections, we used to, but then we ran into some issues so we stopped :-)

…

On Thu, Jul 12, 2018, 11:59 AM Alex Crichton ***@***.***> wrote: I believe this isn't the first of its kind for libm intrinsics in libcore, but the general story around them is "if you use --gc-sections and don't use the code that uses the intrinsics it doesn't matter" . @alex <https://github.com/alex> do you know if you're compiling with --gc-sections and/or enabling that sort of flag in the linker? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#52301 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAADBOYG9x15GpkF0tqkASiF9lj5ZAjLks5uF3JsgaJpZM4VMsWH> .

[alexcrichton]

Ah ok, in general Rust code is designed to be compiled with --gc-sections, otherwise there's binary bloat issues as well. I believe that compiling with --gc-sections will remove the dependence on these symbols as you're not likely using the code anyway

[alex]

@alexcrichton I haven't dug in any depth, but adding --gc-sections doesn't appear to have resolved it (fishinabarrel/linux-kernel-module-rust#61 latest commit + travis run).

[alexcrichton]

@alex hm is it possible to get the full linker command line printed out by the makefiles? If these symbols are still being referenced when --gc-sections is passed that's perhaps a bug! Or perhaps a facet of another flag being passed to the linker.

[geofft]

Or in other words, the real issue here is that floats are part of the language, but some operations on float require library support on some targets, and therefore those operations cannot be part of core.

Yes, this seems right - while digging into this I noticed that f32::sqrt() and f64::sqrt() are in libstd and not libcore specifically for this reason (#27823).

Rethinking that and saying "use --gc-sections" seems reasonable from first principles but I think it's not what libcore/libstd have done historically (or in other words, if you want to do this it seems like it would make sense to reintroduce these functions to libcore :) ).

Also --gc-sections has caused weird bugs for us (see fishinabarrel/linux-kernel-module-rust@7cd1003) so I'm not a huge fan but it's potentially the right thing long-term....

[geofft]

@alexcrichton For what it's worth, --gc-sections does in fact fix this on Ubuntu 18.04 (kernel 4.15.0-1007-aws, binutils 2.30) but not on Travis's Ubuntu 14.04 image (kernel 4.4.0-101-generic, binutils 2.24 I think). On 18.04 the relevant linker command, from running the kernel buildsystem with make V=1, is

  ld -m elf_x86_64 -z max-page-size=0x200000   --entry=init_module --gc-sections -r -o /home/ubuntu/t2/example-sysctl/examplesysctl.o /home/ubuntu/t2/example-sysctl/target/x86_64-linux-kernel-module/debug/libexample_sysctl.a

On Travis, the corresponding command is the same except without -z max-page-size (i.e., it does also include --gc-sections). See https://travis-ci.org/alex/linux-kernel-module-rust/builds/403377295#L731

I wonder if libcore isn't being compiled with -ffunction-sections (or the Rust equivalent of it?) on 14.04 and therefore it can't be GC'd. Note that we are compiling libcore ourselves via cargo-xbuild.

[alexcrichton]

@geofft FWIW it's always been the case that --gc-sections is recommended with Rust, otherwise rust binaries have a lot of bloat they otherwise don't need. The core/std libraires are compiled to just a handful of object files (used to be just two!) and you don't need 90% of the code most of the time, and we rely on -ffunction-sections and -fdata-sections (turned on by default for Rust) plus --gc-sections to remove all the unneeded code.

While items like sqrt may have been moved from core to std items like a % b which are lowered to fmod by LLVM were not moved. We're somewhat inconsistent about this!

That linker command looks like it's doing a relocatable link (producing another object file with -r) and I think that ignores --gc-sections? Did the build process change between those kernel versions?

[reynoldsbd]

I think I'm also running up against this issue, although my circumstances are slightly different:

• Target does not have +soft-float, but the cos/cosf/sin/sinf symbols are still failing to resolve
• Target uses the lld-link flavor, so --gc-sections is not available (the LINK analog /OPT:REF,NOICF is specified)

[alexcrichton]

@reynoldsbd I wonder if that's maybe a bug in LLD? Or maybe a version of LLD which didn't have it implemented?

[reynoldsbd]

@alexcrichton I've tried a few different versions of LLD, including version 7, but the issue does not go away. Could definitely be a bug, but I'm also wondering if PE/COFF linking just has different semantics when it comes to GC'ing. According to the docs linked in my previous comment, /OPT:REF only applies to COMDAT sections. Perhaps the sections referencing these symbols are not COMDATs?

That being said, @gnzlbg's solution of just removing the references when building via core seems more sustainable to me.

[gnzlbg]

Can you try with the latest nightly again? The modules causing this issue have been removed, so this should have been fixed.

[reynoldsbd]

@gnzlbg that did it! My issue has disappeared on the latest nightly. Thank you!