Cleanup Custom Tests

[josephlr]

This moves the tests for the custom RNG registration into custom.rs. It also:

• Simplifies the registered custom RNG
• Makes sure the custom RNG is not used on supported platforms
• Makes sure the custom RNG is used on unsupported platforms

Split out from #471 CC @briansmith

[newpavlov]

This looks fine to me, but I am not sure we need to specifically test that custom implementations do not overwrite the default backends.

Also, I have an alternative proposal. I think we can do the following:

// in lib.rscfg_if!{if #[cfg(all(test, test_rdrand))]{mod lazy;
#[path = "rdrand.rs"]mod imp;} else if #[cfg(all(test, test_custom, feature = "custom"))]{use custom as imp;} else if #[cfg( ... ))]{// ...}}// in tests module:register_custom_getrandom!(mock_rng);#[cfg(not(all(feature = "custom", test_custom)))]#[test]fnmock_rng_is_not_used(){// ..}

This will eliminate the weird juggling around RDRAND and will allow us to test a custom implementation on standard Linux targets. Using cfg(test) will ensure that the testing branches will not escape outside of this crate.

[josephlr]

@newpavlov@briansmith PTAL. I moved the custom.rs tests back to being an integration test, so both supported and unsupported platforms are now tested without changing the cfg_if! in lib.rs.

This looks fine to me, but I am not sure we need to specifically test that custom implementations do not overwrite the default backends.

Also, I have an alternative proposal. I think we can do the following:

I think its important to test that:

• The custom backend works
• The custom backend won't override existing implementations

My concern with changing the lib.rs logic is that by adding too many test cfgs, we might end up not actually testing what we want.

This will eliminate the weird juggling around RDRAND and will allow us to test a custom implementation on standard Linux targets. Using cfg(test) will ensure that the testing branches will not escape outside of this crate.

I agree that RDRAND has some weird juggling, but that's probably best left for discussion in #476

[briansmith]

This makes a lot more sense to me. Just a couple comments.