Fix simd_backend + --no-default-features

[tarcieri]

...and one other alloc cleanup.

simd_backend requires alloc and so this combination previously caused a compile error.

Also one additional test that works with an alloc-only profile wasn't enabled, causing a warning in this case.

Finally, changes CI to test this configuration.

[tarcieri]

Well that's fun...

https://github.com/dalek-cryptography/curve25519-dalek/actions/runs/3456308070/jobs/5768944920

error: failed to run custom build command for `typenum v1.15.0`

Caused by:
  process didn't exit successfully: `/home/runner/work/curve25519-dalek/curve25519-dalek/target/debug/build/typenum-e88f9767296c4856/build-script-main` (signal: 4, SIGILL: illegal instruction)
  --- stdout
  cargo:rustc-env=TYPENUM_BUILD_CONSTS=/home/runner/work/curve25519-dalek/curve25519-dalek/target/debug/build/typenum-11848c0af93[14](https://github.com/dalek-cryptography/curve25519-dalek/actions/runs/3456308070/jobs/5768944920#step:5:15)bcc/out/consts.rs
warning: build failed, waiting for other jobs to finish...
Error: Process completed with exit code 101.

It seems the RUSTFLAGS to enable the SIMD target features are being applied to a build script, which is then executed in CI, and dies due to SIGILL.

I'm not sure if there's a way to selectively enable RUSTFLAGS only for the target executable and not for the build scripts.

[rozbb]

Ugggh I've fought this before. I don't think there's a way to do crate-specific RUSTFLAGS, but I might be wrong.

Is the only alternative the feature flag route?

[tarcieri]

@rozbb my best guess is it's a change in a recent nightly which is auto-vectorizing something which wasn't being vectorized before. See #434 as a trivial code change which is also breaking in CI now.

I tried pinning the nightly version in #435 and that seems to provide a band-aid until we can figure out a better solution.

[tarcieri]

@rozbb this is green now

[rozbb]

Actually, it seems that only tests require alloc. cargo +nightly build --no-default-features --features "simd_backend" works for me

[rozbb]

Finally, changes CI to test this configuration.

I don't see the changes anymore

[tarcieri]

Actually, it seems that only tests require alloc. cargo +nightly build --no-default-features --features "simd_backend" works for me

Aah hmmm, ok. Maybe the tests should be gated on alloc.

I don't see the changes anymore

Was lost in a rebase unfortunately

[rozbb]

If it's a pain to feature-gate the tests, I don't think it matters that much. I don't think there's an AVX2-capable machine that also has no RAM

[tarcieri]

Gated the alloc-dependent tests in a94629c.

If it's a pain to feature-gate the tests, I don't think it matters that much. I don't think there's an AVX2-capable machine that also has no RAM

I'm a little curious why this only pops up with simd_backend enabled. But also, I think it's generally good to ensure all combinations of features work unless you explicitly bail out with compile_error! for invalid combinations.