inference problem with broadcasting and /

[tpapp]

This is an MWE isolated from a larger piece of code:

using ForwardDiff, StaticArrays, Test
_dual(x) = ForwardDiff.Dual(x, 1 + x) # for testing
f(η, B, R, 𝑢, σ, C) = (η .* (B .- R) ./ 𝑢 .- C) ./ σ
A = fill(1.0, SMatrix{2, 2})
b = 1.0
@inferred f(A, b, A, b, b, A)        # infers fine
@inferred f(A, b, A, _dual(b), b, A) # does not infer

Version information:

julia> versioninfo()
Julia Version 1.3.0-rc4.1
Commit 8c4656b97a (2019-10-15 14:08 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: Intel(R) Core(TM) i5-8250U CPU @ 1.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, skylake)
Environment:
  JULIA_CMDSTAN_HOME = /home/tamas/src/cmdstan/

(v1) pkg> st StaticArrays ForwardDiff
    Status `~/.julia/environments/v1/Project.toml`
  [163ba53b] DiffResults v0.0.4
  [f6369f11] ForwardDiff v0.10.3
  [276daf66] SpecialFunctions v0.7.2
  [90137ffa] StaticArrays v0.11.0 #master (https://github.com/JuliaArrays/StaticArrays.jl.git)
  [b77e0a4c] InteractiveUtils 
  [37e2e46d] LinearAlgebra 

[tpapp]

And

f(η, B, R, 𝑢, σ, C) = (η .* (B .- R) / 𝑢 .- C) ./ σ

(note / before 𝑢) works fine.

[c42f]

Thanks for the report. I do wonder whether there's anything we can realistically do about this, or whether a fix is required in Base. I guess a deep dive will be required to figure out what's going on.

[tpapp]

I tried to investigate but could not figure it out, I was kind of hoping that someone here with more knowledge about debugging these issues would take it up.

[c42f]

Understood. I haven't had time to do the requisite digging and TBH Julia's type inference implementation is not something I know a lot about yet.

Any thoughts @martinholters? You've solved a few of these type of problems before :-)

[schmrlng]

This might be related to #609 and JuliaLang/julia#32552 (and linked history within) -- this is definitely my biggest gripe with julia, that it's essentially impossible to write high performance code combining StaticArrays and broadcasting because type inference gets "tired" halfway through drilling down to the bottom of a huge stack of highly-parameterized Broadcasted types. Trying to make things easier for inference (below, asking it to do a bit less dot fusion) can be effective as a workaround but it's definitely kind of annoying:

  | | |_| | | | (_| |  |  Version 1.3.0-rc4.1 (2019-10-15)
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org/ release
|__/                   |

julia> using ForwardDiff, StaticArrays, Test

julia> begin
           _dual(x) = ForwardDiff.Dual(x, 1 + x) # for testing
           broadcasted_division(x, y) = x ./ y
           f(η, B, R, 𝑢, σ, C) = (η .* broadcasted_division(B .- R, 𝑢) .- C) ./ σ
           A = fill(1.0, SMatrix{2, 2})
           b = 1.0
       end
1.0

julia> @inferred f(A, b, A, b, b, A)        # infers fine
2×2 SArray{Tuple{2,2},Float64,2,4} with indices SOneTo(2)×SOneTo(2):
 -1.0  -1.0
 -1.0  -1.0

julia> @inferred f(A, b, A, _dual(b), b, A) # now also infers fine
2×2 SArray{Tuple{2,2},ForwardDiff.Dual{Nothing,Float64,1},2,4} with indices SOneTo(2)×SOneTo(2):
 Dual{Nothing}(-1.0,-0.0)  Dual{Nothing}(-1.0,-0.0)
 Dual{Nothing}(-1.0,-0.0)  Dual{Nothing}(-1.0,-0.0)

(v1.3) pkg> st
    Status `~/.julia/environments/v1.3/Project.toml`
  [f6369f11] ForwardDiff v0.10.6
  [90137ffa] StaticArrays v0.12.1 #master (https://github.com/JuliaArrays/StaticArrays.jl.git)

[mateuszbaran]

Small update: I've proposed a solution here: JuliaLang/julia#41090 .