Some broadcast performance tweaks

[pabloferz]

This allows dot-calls to fully specialize on the first argument when it is a type. The solution here does not extend to type arguments in other positions. For that, something like #19829 for type arguments would be needed, but I'm not sure that is possible in general (except for built-in types).

This PR also allows to remove shape checking in broadcast! when applying an @inbounds to it.

Should fix #19849

[pabloferz]

With this patch, on my machine, I get

julia> A = collect(0.0:0.1:100.0);

julia> @benchmark (x->trunc(Int,x)).($A)
BenchmarkTools.Trial: 
  memory estimate:  8.00 kb
  allocs estimate:  1
  --------------
  minimum time:     2.635 μs (0.00% GC)
  median time:      2.738 μs (0.00% GC)
  mean time:        3.113 μs (6.93% GC)
  maximum time:     124.929 μs (92.14% GC)
  --------------
  samples:          10000
  evals/sample:     9
  time tolerance:   5.00%
  memory tolerance: 1.00%

julia> @benchmark trunc.(Int, $A)
BenchmarkTools.Trial: 
  memory estimate:  8.00 kb
  allocs estimate:  1
  --------------
  minimum time:     2.662 μs (0.00% GC)
  median time:      2.798 μs (0.00% GC)
  mean time:        3.061 μs (6.39% GC)
  maximum time:     69.634 μs (90.93% GC)
  --------------
  samples:          10000
  evals/sample:     9
  time tolerance:   5.00%
  memory tolerance: 1.00%

[tkelman]

Nice. A little off-topic, but I sent you an email (at the address you use in your gitconfig) on new year's eve, did it go to spam?

[pabloferz]

I just had a look at the email you sent me (it got lost among a bunch of other emails). That's a really nice new year's surprise. Thank you!

[pabloferz]

Let's give my new granted powers a run: @nanosoldier runbenchmarks(ALL, vs = ":master")

[tkelman]

I think you need to click a confirm button somewhere, github likes to hide these things

[pabloferz]

Found it. Let's try again. @nanosoldier runbenchmarks(ALL, vs = ":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @jrevels

[martinholters]

I had another idea the other day: To make the outer layer broadcast[!] inspect its argument types, and if any of them are scalar, inline them in the function argument, so that e.g. with T=Int and xs=rand(5), broadcast(f, T, xs) would call broadcast(a -> f(T, a), xs). I'm not sure whether that is at all possible without loosing inferability. The problem boils down to being able to write something like

function fixarg{i}(f, ::Type{Val{i}}, v)
    (args...) -> begin
        f(args[1:i-1]..., v, args[i:end]...)
    end
end

so that foo() = fixarg(^, Val{2}, 2)(3) could be inferred. I think this is impossible as it would need to instantiate a new closure type based on the Val{i} type, i.e. during inference. Can someone confirm this or is the general idea worth pursuing?

[tkelman]

like #19724?

[martinholters]

Yes, of course, thanks for the pointer! Let me see whether I can make something useful out of that...

[pabloferz]

There's also some discussion of other possible approaches here

[stevengj]

We can't do #19829 for type objects, because there's no "type literal" in Julia... something like Int is just an identifier, and it is only resolved to be a type much later in the compilation process (long after lowering).