CI: fix build-simd job

[tarcieri]

The RUSTFLAGS were getting applied to build scripts, which caused them to crash with SIGILL.

According to this issue, RUSTFLAGS won't be applied to build scripts when cross-compiling by passing the --target attribute:

rust-lang/cargo#4423

This attempts to work around the problem by explicitly passing:

--target x86_64-unknown-linux-gnu

[rozbb]

This fix works for me, but it kinda sucks that the build is broken this bad. Do you know if anyone upstream knows about this? I guess we can just document in the README that this is how you have to compile with SIMD for now (since it already requires a few steps)

[tarcieri]

@rozbb I think a less brittle and more permanent fix would be to use runtime CPU feature detection rather than RUSTFLAGS. It would even be possible to run the tests with simd_backend enabled in that case.

But this should be enough to get the build working again.

[rozbb]

Don't you need the RUSTFLAGS part for codegen? If we want runtime detection, then every build needs to include the AVX2/AVX512 code, no?

[tarcieri]

For ideal codegen you need RUSTFLAGS, but you can still achieve baseline functionality without it.

Instead you can annotate functions with #[target_feature(enable = "avx...")] and LLVM will still inline across those with the assumption the given features are available.

Such functions must also be annotated unsafe and have the potential to SIGILL if you don't do runtime feature detection.

If we want runtime detection, then every build needs to include the AVX2/AVX512 code, no?

Only if the simd_backend feature is enabled.

[rozbb]

Ah I see. I assumed the RUSTFLAGS were necessary but I suppose they're not. Just generate the code for everything if simd_backend is enabled. It's only a redundancy of 1 backend, at most, so not bad. Is that right?

[tarcieri]

Yeah, you include code for all backends for a particular target architecture, and detect which are available at runtime.

That eliminates the need for most users to enable RUSTFLAGS to take advantage of the backend, and also makes the binary relocatable.