map recommendations in deprecation warnings and NEWS harmful?

[garborg]

For the deprecated lowercase int, bool, etc.

Timings (where vector is sized so map takes ~1 second):
Int -> Bool: map 20x slower than Vector{Bool} which is 25% slower than round
Bool -> Int: map 20x slower than Vector{Int} or round
Int8 -> Int: map 15x slower than Vector{Int} or round
Int -> Float: map 10x slower than Vector{Float64} or float
Float64 -> Float32: map 50x slower than Vector{Float32}

(Aside: I'm also unsure why Bool(200) works when Int8(200) doesn't, if anyone wants to point me to the right discussion.)

[simonster]

Looks like there is a regression that may explain part (but not all) of the speed gap. On 0.3.6:

julia> immutable A; x::Int; end

julia> x = rand(Int, 100000000);

julia> @time map(A, x);
elapsed time: 0.8371671 seconds (800000128 bytes allocated)

On master:

julia> immutable A; x::Int; end

julia> x = rand(Int, 100000000);

julia> @time map(A, x);
elapsed time: 4.653523433 seconds (3814 MB allocated, 5.85% gc time in 140 pauses with 1 full sweep)

It doesn't look like map_to! has changed, so this may be a codegen regression.

[JeffBezanson]

Oh drat. This used to work, but removing the type from the first argument to map and map_to! causes us not to specialize it for every type. This is one of those heuristics.

[timholy]

Thanks for catching and plugging this.