Switch to FxHasher hash function in int64hash and bitmix

[Zentrik]

This pr switches the current hash function to FxHasher as discussed in #52440.
On the allinference benchmarks in BaseBenchmarks I didn't see any regressions and generally a 5% performance improvement.

[Zentrik]

To get doctests passing I can either make this code use the old hash function

julia/src/symbol.c

Lines 22 to 27 in eba10dd

	static uintptr_t hash_symbol(const char *str, size_t len) JL_NOTSAFEPOINT
	{
	uintptr_t oid = memhash(str, len) ^ ~(uintptr_t)0/3*2;
	// compute the same hash value as v1.6 and earlier, which used `hash_uint(3h - objectid(sym))`
	return inthash(-oid);
	}

or update the docs. Not sure if we care about the symbol hash changing.

[simeonschaub]

We generally don't guarantee stable hashes across minor versions so just changing the doctests should be fine

[KristofferC]

@nanosoldier runbenchmarks("inference", vs=":master")

[Zentrik]

It seems performance has regressed significantly since the first commit, at least locally.

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[vchuravy]

A good benchmark would be the time it takes to build core.ji and the sysimg. At least in my recent profile int64hash is 4% e.g. 20s on my machine.

[Zentrik]

I've had trouble getting reproducible benchmarks results but I'm fairly confident this pr should now be an improvement on the allinference benchmarks against master.

As for the time taken to build corecompiler.ji and the sysimage I didn't notice a difference when running make but this pr does make bitmix much faster so it should eliminate the int64hash hotspot from your profile. I'll have to see if I can get perf or some other tool working properly to confirm.

[Zentrik]

Coming back to this, I only see small (~5%) regressions. Not sure how I ever saw improvements.
I suspect this is due to poor hash quality causing more rehashing and objectid still being slow due to the cold path and preventing in lining (like you probably see when hashing in rust). I'll try again with an alternate hash function.