Inconsistent behavior of Base.floatrange

[metab0t]

When digging into a bug in StatsBase.fit:

• StatsBase.fit(StatsBase.Histogram) produces wrong output JuliaStats/StatsBase.jl#667
• Incorrect Histogram fit for arrays of equal Float64 values JuliaStats/StatsBase.jl#616

I find that the core issue happens in usage of Base.floatrange

I try the following code:

julia> Base.floatrange(1.05 , 1.0 , 2, 1.0)
2.05:2.0:4.05

julia> Base.floatrange(1.0  , 1.0 , 2, 1.0)
1.0:1.0:2.0

In twiceprecision.jl, ref in two steprangelen_hp definitions has different meanings.

• In the integer case, it means TwicePrecision{Float64}(ref)
• In the float case, it means TwicePrecision{Float64}(ref...) (it should align with integer case)

However, TwicePrecision{Float64}(1.0, 1.0) and TwicePrecision{Float64}((1.0, 1.0))are totally different.

Therefore, two floatrange definitions return different results

• julia/base/twiceprecision.jl

  Line 381 in 45ab664

  function floatrange(::Type{T}, start_n::Integer, step_n::Integer, len::Integer, den::Integer) where T

  (correct)

• julia/base/twiceprecision.jl

  Line 395 in 45ab664

  function floatrange(a::AbstractFloat, st::AbstractFloat, len::Real, divisor::AbstractFloat)

  (if it hits steprangelen_hp of the float case, result is wrong)

Code examples:

julia> Base.floatrange(Float64, 11, 10, 2, 1)
11.0:10.0:21.0

julia> Base.floatrange(11.0, 10.0, 2, 1.0) # fallback to use floatrange(Float64, ::Integer, ::Integer, ::Integer, ::Integer)
11.0:10.0:21.0

julia> Base.floatrange(11.1, 10.0, 2, 1.0) # use float case of steprangelen_hp
12.1:11.0:23.1

Version info:

julia> versioninfo()
Julia Version 1.7.2
Commit bf53498635 (2022-02-06 15:21 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: AMD Ryzen 7 5800H with Radeon Graphics
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, znver3)
Environment:
  JULIA_NUM_THREADS = auto
  JULIA_PKG_USE_CLI_GIT = true

[oscardssmith]

wow. That is awful.

[N5N3]

Do we have a MWE to reproduce this bug with public API? (Base.floatrange is not documented)

I build the latest master with L395 commented. ranges test only show one error

Test Summary: |     Pass  Error  Broken     Total     Time
  Overall     | 12110865      1  327682  12438548  1m11.8s
    ranges    | 12110865      1  327682  12438548  1m10.3s
    FAILURE

The global RNG seed was 0xd80800aebdb44d7a3931f2e34a3be386.

Error in testset ranges:
Error During Test at C:\Users\XXX\Documents\Github\julia\test\ranges.jl:877
  Test threw exception
  Expression: length(Base.floatrange(-3.0e9, 1.0, 1, 1.0)) == 1
  MethodError: no method matching floatrange(::Float64, ::Float64, ::Int64, ::Float64)
  Closest candidates are:
    floatrange(::Type{T}, ::Integer, ::Integer, ::Integer, ::Integer) where T at twiceprecision.jl:381

So I think we should just remove it.

[metab0t]

https://github.com/search?l=Julia&p=1&q=Base.floatrange&type=Code

After code search on GitHub, we can almost conclude that the only widely user of Base.floatrange is StatsBase/src/hist.jl.
So maybe we should remove it and fix StatsBase?

[metab0t]

Maybe we can remove it because StatsBase.jl does not depend on it anymore.

[Moelf]

@metab0t feel free to make a PR implementing it and we can PkgEval it

[metab0t]

@Moelf I have opened a PR to remove the inconsistent instance of Base.floatrange.