faster rand(::MersenneTwister, ::Union{[U]Int{8,16,32},Bool})

[rfourquet]

Now that generating scalar Float64 values (i.e. 52 random bits) is even faster,
it becomes faster to just generate such values and make small integers out of
them, rather than saving entropy bits with using the integer cache.

I tested on two different machines, the observed speed-up ranges from 15% to 40%.

[vtjnash]

IIUC, this changes the random number stream? We should make sure to run nanosoldier on this to have an idea of which tests that’ll affect.

[rfourquet]

this changes the random number stream?

Yes.

We should make sure to run nanosoldier on this to have an idea of which tests that’ll affect.

My recent invocation of Nanosoldier were not always showing the expected improvements, but let's see.

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

[nanosoldier]

Your benchmark job has completed - no performance regressions were detected. A full report can be found here. cc @ararslan

[rfourquet]

It's funny: Nanosoldier shows two improvements, which are consequences of the change here (for ComplexF32 and Complex{UInt32}), but not the most direct improvements for the small integer types.

[vtjnash]

I mean we should run it on all tests. Some of them are sensitive to the numbers themselves, so we will want to know which changes in the next daily are “noise”.

[rfourquet]

Ah right, got it!
@nanosoldier runbenchmarks(ALL, vs=":master")

[nanosoldier]

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

[rfourquet]

Freebsd CI failure is a timeout.