use ScopedValues for MPFR precision and rounding

[simonbyrne]

Should fix thread safety issues.

Actually fixes #27139

[longemen3000]

Could this fix #51058?

[vchuravy]

You could set it to default precision? Or is it important that the default precision could be changed globally?

[simonbyrne]

We support it at the moment, so dropping it would be breaking, but I'm not sure how necessary it is?

[MasonProtter]

If you need to be able to change it, don't you need this to be

const CURRENT_PRECISION = ScopedValue{RefValue{Clong}}(Ref{Clong}())

since ScopedValue is immutable?

[simonbyrne]

No, but the default field is marked const:

julia/base/scopedvalues.jl

Lines 40 to 46 in b189bed

	mutable struct ScopedValue{T}
	const has_default::Bool
	const default::T
	ScopedValue{T}() where T = new(false)
	ScopedValue{T}(val) where T = new{T}(true, val)
	ScopedValue(val::T) where T = new{T}(true, val)
	end

[MasonProtter]

I didn't mean that it's an immutable struct, I just mean that you can't actually change the value other than by introducing a new dynamic scope.

[simonbyrne]

Yes, that's correct.

[vchuravy]

Nice! I would be curious to see what's the performance impact.

[simonbyrne]

@vchuravy should this overload Base.get?

[vchuravy]

It has different API so I decided not to.

get(dict, key, value) vs get(ScopedValue). But that was me being conservative, also it returns Some/Nothing.

I could see an argument for get(SValue, value)

[simonbyrne]

I see a minor impact:

function f(N)
    x = big(0.0)
    for i = 1:N
        x += BigFloat("0.1")
    end
    return x
end

function g(N)
    setrounding(BigFloat, RoundDown) do
        f(N)
    end
end

On master:

julia> @btime f(1_000_000)
  197.769 ms (5000002 allocations: 202.18 MiB)
100000.0000000000000000000000000000000000000000000000000000000000000000001391156

julia> @btime g(1_000_000)
  202.435 ms (5000002 allocations: 202.18 MiB)
99999.99999999999999999999999999999999999999999999999999999999999999999955990192

On this branch:

julia> @btime f(1_000_000)
  208.775 ms (5000002 allocations: 202.18 MiB)
100000.0000000000000000000000000000000000000000000000000000000000000000001391156

julia> @btime g(1_000_000)
  251.740 ms (5000007 allocations: 202.18 MiB)
99999.99999999999999999999999999999999999999999999999999999999999999999955990192

So a bit of overhead, but not too bad.

[simonbyrne]

To clarify, this would access the scoped values around 4 million times (in each iteration, access both rounding mode and precision for thee decimal to binary conversion, then the addition), so 40 milliseconds of overhead would be about 10ns per call.

[simonbyrne]

Could this fix #51058?

No.

[vchuravy]

Changing the precision has a bit more overhead, but we would assume that's a rare operation to begin with?

[simonbyrne]

Changing the precision has a bit more overhead, but we would assume that's a rare operation to begin with?

Actually, that's probably more common. However I see roughly the same overhead when using the following:

function h(N)
    setprecision(BigFloat, 256) do
        f(N)
    end
end

[KristofferC]

What's the status here @simonbyrne? Rebase and merge?

[simonbyrne]

I think this is ready to go now.

[nsajko]

PR #55911 fixes a regression introduced by this PR, could anyone help review it?