Allocation-free code in julia v1.10 shows allocations in julia v1.11

[fhagemann]

This code used to be allocation-free in julia-1.10 but is not anymore in julia-1.11:

using BenchmarkTools

fmap(fs, x...) = _fmap(x, fs...)
@inline _fmap(x, f, fs...) = (f(x...), _fmap(x, fs...)...)
@inline _fmap(x) = ()

fs = (hypot, atand)
x = 3
y = 4
@btime fmap($fs, $x, $y);

Julia v1.10.8: 51.022 ns (0 allocations: 0 bytes)
Julia v1.11.4: 191.837 ns (3 allocations: 64 bytes)
(The allocations are also observed in julia v1.12.0-beta1: 199.714 ns (3 allocations: 64 bytes))

Is this something that would need to be addressed in the julia source code or can this be avoided by small tweaks to the code? Any help is very much appreciated.

[adienes]

interestingly, AllocCheck.jl doesn't seem to yell about this, despite the 64b reported from @allocated

[nsajko]

can this be avoided by small tweaks to the code?

Yeah. The allocation should be prevented by forcing specialization on the Vararg length for fmap, see:

• Be aware of when Julia avoids specializing

So write fmap(fs, x::Vararg{Any,N}) where {N} instead of fmap(fs, x...).

Whether this workaround is required in the actual code is a different matter, depending on your goals. Perhaps ask on Discourse or Zulip if you have not already.