Don't overwrite Base.vect for numbers with a different implementation

[ararslan]

Fixes jump-dev#241

[iamed2]

Double-checked that this compiles out the any call?

[iamed2]

Should we do this with vcat as well?

[ararslan]

Double-checked that this compiles out the any call?

It does not, unfortunately, and there is a performance impact on the Base method with this change.

After loading Convex:

julia> @benchmark Base.vect($x...)
BenchmarkTools.Trial: 
  memory estimate:  102.72 KiB
  allocs estimate:  375
  --------------
  minimum time:     36.807 μs (0.00% GC)
  median time:      38.912 μs (0.00% GC)
  mean time:        45.703 μs (12.67% GC)
  maximum time:     33.698 ms (99.77% GC)
  --------------
  samples:          10000
  evals/sample:     1

(here x is a 74-element Vector{Any} consisting of Ints, Float32s, and Complex{Int}s.)

Without loading Convex, same x:

julia> @benchmark Base.vect($x...)
BenchmarkTools.Trial: 
  memory estimate:  55.84 KiB
  allocs estimate:  148
  --------------
  minimum time:     22.036 μs (0.00% GC)
  median time:      23.726 μs (0.00% GC)
  mean time:        28.634 μs (13.89% GC)
  maximum time:     25.832 ms (99.89% GC)
  --------------
  samples:          10000
  evals/sample:     1

[iamed2]

Can we do it the old way then? I know it's technically type piracy...

[ararslan]

Okay, I've reinstated the old code but corrected the implementation.

[iamed2]

Hmm what if we changed it to use invoke unconditionally for vect and vcat?

[ararslan]

I checked that earlier and there is still a performance impact, unfortunately. Smaller than with the any branch at least, but still slower than without loading Convex. With the current state of the PR, the performance is unchanged after loading Convex (since the code does exactly the same thing).