Look up correct framecode for invoke

[timholy]

Prior to this the tests failed due to the following:

julia> using JuliaInterpreter

julia> sin(1)
0.8414709848078965

julia> @interpret sin(1)
0.8370984093755471

This turns out to be due to the new evalpoly, it was looking up the wrong framecode. This might have been a bug on all previous versions, but only detected now because of the change from macro to generated function.

[timholy]

OSX and Windows test failures seem unrelated (install_archive).

[jmert]

For what it's worth, this helps with a case that interprets an invoked call with a where clause — I was trying to debug a code path, but kept getting that the type variables were unknown. Merging this locally fixed the issue.

julia> using Debugger, LinearAlgebra

julia> f(d::Diagonal{T}) where {T} = invoke(f, Tuple{AbstractMatrix}, d)
f (generic function with 1 method)

julia> f(m::AbstractMatrix{T}) where {T} = T
f (generic function with 2 methods)

julia> D = Diagonal([1.0, 2.0])
2×2 Diagonal{Float64,Array{Float64,1}}:
 1.0   ⋅ 
  ⋅   2.0

julia> f(D)
Float64

julia> @run f(D)
Breaking for error:
ERROR: UndefVarError: T not defined
Stacktrace:
 [1] f(::Diagonal{Float64,Array{Float64,1}}) at REPL[3]:1
 [2] f(::Diagonal{Float64,Array{Float64,1}}) at REPL[2]:1

In f(m) at REPL[3]:1
>1  f(m::AbstractMatrix{T}) where {T} = T

About to run: return $(QuoteNode(:($(Expr(:static_parameter, 1)))))
1|debug> q

julia> 

[timholy]

Thanks @jmert, I shamelessly stole this test and added it here. I can verify that it fails on current master.

[timholy]

Why is this still using

Julia Version 1.4.0-DEV.535
Commit 9babbf5dc6 (2019-11-30 20:29 UTC)

There's a newer version available:

Julia Version 1.4.0-DEV.546
Commit 84e839227f (2019-12-03 02:31 UTC)

Interestingly, over at timholy/SnoopCompile.jl#43 one sees the "push" build is using a different nightly from the "pr" build on linux. (OSX is using a consistent nightly.)

CC @StefanKarpinski

[StefanKarpinski]

I don't know. I'm not really an expert on these CI systems. Perhaps @staticfloat knows?

[staticfloat]

I don't know why. I re-ran the job and it got a newer one. There are levels of caching that happen between the CI services and the buildbot output, and the caching is implemented differently for different CI providers (such as MacOS on Travis or Linux on Travis, which are within different datacenters)

[timholy]

While nightly continues to be problematic, on my machine master passes, so I'll merge this.