hypot performance poor for more than 3 arguments

[sethaxen]

When 4 or more arguments are pass to hypot, its performance suddenly becomes very poor.

julia> using BenchmarkTools

julia> f(x) = hypot(x...)
f (generic function with 1 method)

julia> x = (1.0, 2.0, 3.0)
(1.0, 2.0, 3.0)

julia> @btime f($x)   # fine
  11.415 ns (0 allocations: 0 bytes)
3.741657386773941

julia> x = (1.0, 2.0, 3.0, 4.0)
(1.0, 2.0, 3.0, 4.0)

julia> @btime f($x)  # so slow!!
  708.965 ns (13 allocations: 336 bytes)
5.477225575051661

Calling hypot pairwise is much faster here

julia> myhypot(x...) = reduce(hypot, x)
myhypot (generic function with 1 method)

julia> g(x) = myhypot(x...)
g (generic function with 1 method)

julia> @btime g($x)
  41.038 ns (0 allocations: 0 bytes)
5.477225575051661

The trend continues even for large tuples:

julia> x = (randn(100)...,);

julia> @btime f($x)
  23.303 μs (623 allocations: 19.14 KiB)
9.54225022090292

julia> @btime g($x)
  9.713 μs (207 allocations: 7.14 KiB)
9.54225022090292

Is there a good reason we shouldn't adopt this default for hypot(x1, x2, x3, x...)?

[oscardssmith]

yes. Your version doesn't deal with overflow as well. That said, we clearly could do better than we currently do.

[N5N3]

Looks like a code_gen(Edit:invoke) problem?

julia> @code_typed hypot(1.0,1.0,1.0,1.0,1.0)
CodeInfo(
1 ─ %1 = Core.getfield(xs, 1)::Float64
│   %2 = Core.getfield(xs, 2)::Float64
│   %3 = Core.getfield(xs, 3)::Float64
│   %4 = invoke Base.Math._hypot(x::Float64, y::Float64, %1::Float64, %2::Vararg{Float64}, %3)::Float64
└──      return %4
) => Float64

julia> @code_typed hypot(1.0,1.0,1.0,1.0,1.0,1.0)
CodeInfo(
1 ─ %1 = Core.getfield(xs, 1)::Float64
│   %2 = Core.getfield(xs, 2)::Float64
│   %3 = Core.getfield(xs, 3)::Float64
│   %4 = Core.getfield(xs, 4)::Float64
│   %5 = invoke Base.Math._hypot(x::Float64, y::Float64, %1::Float64, %2::Vararg{Float64}, %3, %4)::Float64
└──      return %5
) => Float64

The above Vararg should be Float64

[simeonschaub]

That's just invoke statements being printed a little weird. This only shows that the call to _hypot was not inlined, I don't see why this would have anything to do with codegen.

[N5N3]

I just wonder why afoldl does much better with a similar patten
Edit: afoldl does much better as we add typeof(afoldl).name.mt.max_args = 34
If this is expected, we should pass a tuple rather than use invoke:
Or just increase the _hypot's max_args.