consider using a different hash than SipHash for integer keys in rustc

[thestinger]

You'll always be able to craft inputs for rustc making it hit a pathological compile-time and I don't think we need to protect against this. SipHash is pretty fast for strings, but on short fixed-size key types it's very slow. The compiler depends heavily on the speed of looking up integer keys in maps, so this is a huge performance cost.

(I'll post benchmark numbers soon)

[thestinger]

A fast block cipher would also work here, but most are probably even slower than SipHash and are likely all going to be slower than just a multiplication...

[huonw]

cc #11783

[erickt]

nominating for closure since @alexcrichton's FNV hasher landed in #12635.

[alexcrichton]

This has indeed be done now, hurray!