mere presence of `throw` in function body makes function much slower

[matthias314]

I have noticed that the run-time of a function may significantly depend on whether the function body contains throw, even if no exception is thrown. The performance tips in the manual don't mention this. I therefore wonder whether it is a bug. The following MWE is adapted from the method typed_hvcat(::Type{T}, rows::Tuple{Vararg{Int}}, xs::Number...) where T in abstractarray.jl.

@noinline throw_arg_err(msg) = throw(ArgumentError(msg))

function hv_cat(rows::Tuple{Vararg{Int}}, xs::Int...)
    nr = length(rows)
    nc = rows[1]
    for i = 2:nr
        if nc != rows[i]
            throw(ArgumentError("mismatch"))  # SLOW
            # throw_arg_err("mismatch")  # FAST
        end
    end
    # Base.hvcat_fill(Matrix{Int}(undef, nr, nc), xs)  # v1.6.1
    Base.hvcat_fill!(Matrix{Int}(undef, nr, nc), xs)  # master
end

The SLOW variant gives

julia> @btime hv_cat((1,1,1), 1,2,3);
  107.800 ns (2 allocations: 144 bytes)

The FAST variant is 3x faster (and makes fewer allocations):

julia> @btime hv_cat((1,1,1), 1,2,3);
  34.530 ns (1 allocation: 112 bytes)

Note that the code differs only in a branch that is not executed. I've tried to replace the call to Base.hvcat_fill! at the end with some other code, but then the difference disappeared. It also disappears if one changes the arguments to hv_cat((1,1), 1,2).

Julia Version 1.8.0-DEV.61
Commit 7553ca13cc (2021-06-21 17:18 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: Intel(R) Core(TM) i3-10110U CPU @ 2.10GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.0 (ORCJIT, skylake)

The example also works for Julia 1.6.1 with the slight change indicated in the code.

[KristofferC]

Ref #23281