Reporting a bug when Tullio being included with LoopVectorization

[zxm403089989]

In this simple example, if just including Tullio, everythings works well.

using Tullio,BenchmarkTools
t=collect(range(start=-3,step=0.01,stop=3));
nb(e)::Float64=inv(expm1(e/1e-4))
@btime @tullio test1[n]:=nb(t[n])
1.800 μs (1 allocation: 4.88 KiB)
601-element Vector{Float64}:

However if I include LoopVectorization and TensorCast next:

using LoopVectorization,TensorCast
@btime @tullio test2[n]:=nb(t[n])
7.850 μs (49 allocations: 7.62 KiB)
┌ Warning: #= C:\Users\.julia\packages\Tullio\NGyNM\src\macro.jl:1093 =#:
│ `LoopVectorization.check_args` on your inputs failed; running fallback `@inbounds @fastmath` loop instead.
│ Use `warn_check_args=false`, e.g. `@turbo warn_check_args=false ...`, to disable this warning.
└ @ Main C:\Users\.julia\packages\LoopVectorization\FMfT8\src\condense_loopset.jl:1049
601-element Vector{Float64}:

After I several times tries, I found this error is related with function type declear. When I define function, no matter declear argument type of return type, with the coexistence with LoopVectorization and TensorCast, LoopVectorization would throw this error.

[mcabbott]

I think the issue is the ::Float64 annotation with LoopVectorization:

julia> using Tullio, LoopVectorization

julia> t=collect(range(start=-3,step=0.01,stop=3));

julia> nb(e)::Float64 = inv(expm1(e/1e-4));

julia> @tullio test1[n] := nb(t[n]);
┌ Warning: #= /Users/me/.julia/packages/Tullio/NGyNM/src/macro.jl:1093 =#:
│ `LoopVectorization.check_args` on your inputs failed; running fallback `@inbounds @fastmath` loop instead.
│ Use `warn_check_args=false`, e.g. `@turbo warn_check_args=false ...`, to disable this warning.
└ @ Main ~/.julia/packages/LoopVectorization/7xFHi/src/condense_loopset.jl:1053

julia> nb_any(e) = inv(expm1(e/1e-4));  # the same, without ::Float64

julia> @tullio test1[n] := nb_any(t[n]);

LoopVectorization works by sending special Vec types contianing several floats through your code, and I presume the above warning comes from noticing that this will fail:

julia> x = LoopVectorization.VectorizationBase.Vec{4,Float64}(1,2,3,4) * 1e-6
VectorizationBase.Vec{4, Float64}<1.0e-6, 2.0e-6, 3.0e-6, 4.0e-6>

julia> nb_any(x)
VectorizationBase.Vec{4, Float64}<99.50083333194446, 49.50166665555567, 32.835833295834135, 24.503333244447838>

julia> nb(x)
ERROR: TypeError: in typeassert, expected Float64, got a value of type VectorizationBase.Vec{4, Float64}

BenchmarkTools and TensorCast are uninvolved.