[WIP] Performance

[noughtmare]

Please test the performance of this.

Currently I've tried to optimize the receive function. I might look at send later.

The correctness should still be checked before this can be merged.

This pull request includes:

• Adding inline pragmas to small functions
• Using unsafe indexing (unsafeRxMap and unsafeRxGetMapping).
• Remove the rxqDescriptor function from the RxQueue record and write it as a separate inlinable function.

Now the go loop inside the receive function uses unboxed operations almost exclusively, which means that no garbage gets generated. The only non-unboxed operations are due to IORefs. That might be a further opportunity for optimization. There is also allocation for the bufPtr that is consed to the bufs list, so using a mutable vector for that would probably also improve performance.

[eleksaggr]

@emmericp

[TravisWhitaker]

Nice, this looks like it should do a bit better.

Would be good to change the GC parameters as well (at least a bigger -A and -H), as the GHC defaults are extremely conservative.

[noughtmare]

I have made some more changes to make the code even more performant: https://github.com/noughtmare/ixy.hs/commit/2c24c5cc7a58df1bbd689b6b9b19e6f71b1abefe. But that changes a lot about the code. It replaces all IOUArrays with unboxed vectors and uses unboxed IORefs instead of the standard boxed ones. And it uses a storable vector instead of the standard list for the send and receive functions. That makes the code uglier (there are ways to make it prettier again, but I didn't do that yet), so I haven't pushed it to this branch.