Expose target-cpu=native as a profile setting

[bluss]

For a lot of projects you want that the compiler uses every cpu feature available. Not just numerics, but loops that copy or search in general can profit from using wider registers or vector instructions.

target-cpu=native makes plain old Rust code much more powerful. We don't need to write simd intrinsics, simd types, we can write one version of the code that works everywhere but the compiler can compile it well. It's not perfect, not always a replacement for hand implemented vectorization, but it's pretty good.

As of this writing, matrixmultiply is in plain rust. It should compile with any settings, but using -C target-cpu=native is a big benefit (more than 50% increase in throughput if you have avx). It would therefore be great if projects using it had an easy way to just flip the switch that they want the compiler to use all available cpu features.

[bluss]

Re #2112 (comment) this falls between the two scenarios in the comment. Neither testing nor required, just a very good thing to use when it's available.

[alexcrichton]

Yeah I don't think we've quite figured out a great way to deal with these sorts of options today. There's RUSTFLAGS which can suffice in the meantime, but it's not a great interface unfortunately.

To me this is essentially the same question of exposing target-feature in Cargo to crates somehow. Ideally, though, the entire DAG is compiled with the same options (including the standard library) which is the hard part.

[retep998]

I'd really like if I could set this to be the default for all my builds locally since I typically don't care about distributing binaries.

[bluss]

@retep998 That seems to be possible using build.rustflags in the user-wide cargo config.