static_parameter expr

[KDr2]

Fix #174

[yebai]

@penelopeysm @willtebbutt, can you help review this PR? The failing Turing integration tests are due to Mooncake, so they can be ignored.

[willtebbutt]

This LGTM. You'll need to bump the version number before merging though so that we can tag a release.

[penelopeysm]

Yes, it does fix the examples in #174! To be precise, the first one still gives an error, but it now gives another error that isn't due to Libtape (it's a meaningless model after all).

[penelopeysm]

@KDr2 I think this might be useful in writing a test 😄 it fails on master and runs correct on this PR.

using Libtask
function f()
    r = (a = 1)
    return nothing
end
tf = Libtask.TapedFunction(f)
r = tf()

I'm still trying to work out the RandomMeasures example. Will update if I get it.

[KDr2]

I think this might be useful in writing a test

Thanks, do you have any clue about how to reproduce the second one with a simple function (not a Turing model)?

[penelopeysm]

Haha, sorry, I just edited the comment (probably after you saw it) to say that I hadn't worked that one out yet. I'm looking at it now!

Thank you for looking into this so quickly!!

[penelopeysm]

@KDr2 I think it's the ::Type{T} that makes this fail on master (but is fixed on this PR):

using Libtask
function g(::Type{T}) where {T}
    a = zeros(T, 10)
end
tf = Libtask.TapedFunction(g, Float64)

#= on master
┌ Error: Unknown Expression:
│   typeof(var) = Core.SSAValue
│   var = :(%1)
│   typeof(line) = Expr
│   line = :($(Expr(:static_parameter, 1)))
└ @ Libtask ~/.julia/packages/Libtask/0vMZ5/src/tapedfunction.jl:443
=#

[penelopeysm]

using Libtask
function g(::Type{T}) where {T}
    a = zeros(T, 10)
end
tf = Libtask.TapedFunction(g, Float64)

The above seems fine, but then subsequently calling r = tf(Float64) fails, at least on <=1.10. This isn't a regression introduced by this PR, it's also on master.

[KDr2]

This isn't a regression introduced by this PR, it's also on master.

Fixed, but what happened to the Integeration test? It passes now...

[penelopeysm]

It's probably because of Hong's commit which bumped the minor version, and Turing.jl's compat entry for Libtask doesn't allow it to be tested with this.

I'm curious as to whether we can keep 1.7 compat here, if you don't mind, I'll test it out :)

[KDr2]

we can keep 1.7 compat

We can keep both 1.7 and 1.10 in the CI unit tests.

[penelopeysm]

Thank you again to all of you who contributed here ❤️