Strange Allocation in `@generated`



I use a generated function for fast evaluation of monomials and their gradients. The following example could be easily rewritten without a generated function, but I don’t see how to easily implement its gradient without @generated. But here the problem I have is easier to see:

using StaticArrays, BenchmarkTools

@generated function monomial(α, x::SVector{K}) where {K}
   # @assert length(α) >= K
   ex = "@fastmath "
   for i = 1:K
      ex *= "x[$i]^α[$i] * "
   ex = ex[1:end-3]

α3 = (1,2,3)
x3 = @SVector rand(3)
@btime monomial($α3, $x3)   # 15.546 ns (0 allocations: 0 bytes)

α6 = (1,2,3,2,5,2)
x6 = @SVector rand(6)
@btime monomial($α6, $x6)   # 2.037 μs (29 allocations: 528 bytes)

is the length of x is 1,2,3,4,5 then there is no allocation, if it is 6, then there is an allocation. Does anybody have an idea what is going on here?


  • if I turn off the @fastmath macro then the allocation goes to zero! But I really need the @fastmath, it gives me a factor 5-10 performance gain here.
  • If I apply the @fastmath to each individual x[I]^a[I] then the allocation goes down to zero as well, but when the length of the vector becomes 8, then it start allocating again


I now have a work-around:

@generated function monomial(α, x::SVector{K, T}) where {K, T}
   # @assert length(α) >= K
   ex_f = "f = SVector{$K, $T}("
   for i = 1:K
      ex_f  *= "x[$i]^α[$i], "
   ex_f  =  ex_f[1:end-2] * ")"

      return prod(f)

but it is still strange and I’d love to find out what goes wrong?



FWIW I tested your examples on Julia master and it is definitely better there.


Thank you - that seems to match my problem