Is it reasonable to use runtime.fasthash on platforms other than linux_amd64?

[jxskiss]

Since drwmutex is mainly to improve the performance of read-most workload, use a fast hash function also can distribute the lock evenly, which may help to reduce lock contention.
Benchmark shows cost of runtime.fasthashn is very cheap.

func BenchmarkCpuid(b *testing.B) {
	for i := 0; i < b.N; i++ {
		_ = cpu()
	}
}

func BenchmarkFastrand(b *testing.B) {
	for i := 0; i < b.N; i++ {
		_ = fastrandn(CPU_COUNT)
	}
}

func cpu() uint64

//go:noescape
//go:linkname fastrandn runtime.fastrandn
func fastrandn(x uint32) uint32

goos: darwin
goarch: amd64

BenchmarkCpuid-12               16798030                66.3 ns/op
BenchmarkFastrand-12            504372906                2.39 ns/op

goos: linux
goarch: amd64

BenchmarkCpuid-4         1471983               813 ns/op
BenchmarkFastrand-4     324570566                3.58 ns/op

So is it reasonable to use this rand method to choose read lock on platforms other than linux_amd64?

[jxskiss]

@jonhoo

[jonhoo]

Hmm.. While this would help somewhat, it won't have anywhere near the same scalability as using a core-local lock (which CPUID) gives you. It could be that it's a reasonable fall-back for other platforms for the time being, but at the same time I think I'd prefer to say that this is linux_amd64-only so that people don't depend on the scaling behavior of the CPUID implementations and then get disappointed when they don't get the same behavior on other platforms.