i686 target spec disagrees with downstream definitions

[sanxiyn]

See https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=973414.

[jyn514]

IIUC, the issue is that rustc generates SSE2 instructions on a CPU that doesn't support SSE2? What would the fix be, to add a new target triple for Geode?

[sanxiyn]

I think what's going on is that Debian and Rust disagree on what i686 is. Debian is clearly right and Rust is clearly wrong (80686 lacks SSE2), but it's what it is and neither here nor there. The issue is analogous to ARM v6/v7 situation.

Rust's i686-unknown-linux-gnu target requires SSE2, so can't run on Geode. Rust's i586-unknown-linux-gnu will run on Geode, but that's unfortunate, because Geode is 686 and not 586 and will miss 686 instructions (including CMOV).

[cuviper]

Debian could further patch rustc_target to disable target-features on i686 -- it seems they already lowered the cpu target from "pentium4" to "pentiumpro".

[steveklabnik]

I've changed the title to something more descriptive.

[Shnatsel]

Geode does not support some instructions that Pentium Pro does, so this behavior seems WAI:

Debian's rustc has a patch to reduce the i386 baseline from upstream's pentium4 to pentiumpro
https://sources.debian.org/src/rustc/1.48.0+dfsg1-1/debian/patches/d-i686-baseline.patch/
but apparently that's not sufficient for a Geode LX. i686 is in a weird situation where the Pentium Pro was the first i686 CPU, but is not a baseline i686 CPU.

from https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=973414#61

I believe the resolution would be dropping the target from pentiumpro to Geode in Debian.

[sanxiyn]

I would like to note that FreeBSD also changes i686 baseline from pentium4 to pentiumpro: https://github.com/freebsd/freebsd-ports/blob/master/lang/rust/files/patch-compiler_rustc__target_src_spec_i686__unknown__freebsd.rs.

Ignoring Geode question for a moment, does it make sense for Rust to set i686 baseline to pentium4? Who is i686 for anyway? Wouldn't people use x86_64 for new hardwares?

[negge]

Just ran into this with gentoo on a Pentium III today (yes in 2021):

https://bugs.gentoo.org/741708

Cross compiling your own dev-lang/rust-bin is not an option for everyone, and as more system packages like librsvg are written in Rust this cuts off more and more systems from reaping the benefit of a safe systems language.

I am one who would appreciate if i686 definition matched what distro's use. Certainly the rust compiler can emit SSE2 instructions on those systems with these instructions present, even if it does not itself contain SSE2 instructions.

[cuviper]

I am one who would appreciate if i686 definition matched what distro's use.

There's no consensus between distros either. Fedora now builds C and C++ with -march=i686 -mtune=generic -msse2 -mfpmath=sse -mstackrealign, and that change called out the fact that i686 is primarily only used for multi-lib compatibility on x86_64 hosts. My impression is that Ubuntu is similar, but I'm not sure about that.

[sanxiyn]

@cuviper linked to a change proposal accepted in Fedora 29, but that's misleading, since Fedora 31 dropped i686 install anyway: https://fedoraproject.org/wiki/Changes/Stop_Building_i686_Kernels. Of course Fedora's i686 target is solely used for multi-lib since Fedora can't be installed on i686.

[thomcc]

Ignoring Geode question for a moment, does it make sense for Rust to set i686 baseline to pentium4? Who is i686 for anyway? Wouldn't people use x86_64 for new hardwares?

Anecdotally, I've seen most use it as "32-bit x86 on more modernish hardware", and the older i586 is "32-bit x86 on older hardware". I think this distinction is somewhat nice, even if it means the targets are misnamed.

Users of machines where this isn't viable might be better off with i586... But I guess that doesn't solve things for distros...