sinpi fails on bigfloats

[fph]

sinpi fails on BigFloats:

julia> CRlibm.sinpi(BigFloat(0.5))
ERROR: StackOverflowError:
Stacktrace:
 [1] sinpi(::BigFloat, ::RoundingMode{:Nearest}) at /home/sad/.julia/packages/CRlibm/NFCH5/src/CRlibm.jl:137 (repeats 79984 times)

[dpsanders]

Sorry I didn't see this before.
CRlibm is only designed to work with arguments of type Float64.

[dpsanders]

So I guess that should just error.

[fph]

Thanks! Then I guess the issue is with IntervalArithmetic.jl, because sin(@biginterval(0.5)), which appears in their test test/interval_tests/trig.jl, seems to use CRlibm.sin on BigFloat arguments internally. I guess I should report it there then, do you confirm?

[dpsanders]

Where do you see that? It seems to work for me:

julia> sinpi(@biginterval(0.5))
[0.999999, 1]₂₅₆

[fph]

I confirm I can replicate it on my machine. Ubuntu Linux 20.04, JuliaPro v1.5.0, CRlibm v0.8.0.

 Activating environment at `~/.julia/environments/JuliaPro_v1.5.0-1/Project.toml`
Starting Julia...
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.5.0 (2020-08-01)
 _/ |\__'_|_|_|\__'_|  |  Official https://julialang.org/ release
|__/                   |

julia> using CRlibm

julia> Pkg.status("CRlibm")
Status `~/.julia/environments/JuliaPro_v1.5.0-1/Project.toml`
  [96374032] CRlibm v0.8.0

julia> CRlibm.sinpi(BigFloat(0.5))
ERROR: StackOverflowError:
Stacktrace:
 [1] sinpi(::BigFloat, ::RoundingMode{:Nearest}) at /home/sad/.julia/packages/CRlibm/NFCH5/src/CRlibm.jl:137 (repeats 79984 times)

[dpsanders]

I meant where do you see that sinpi(x) for x::Interval{BigFloat} calls CRlibm.sin on a BigFloat.
sinpi(x) for a BigFloat Interval works, so I don't see how it can be calling this method.

But now I see that CRlibm.sin(y) does work when y is a BigFloat, so CRlibm.sinpi should too, I guess. (Or neither should work, preferably!)

[dpsanders]

Ah I see, it was to define methods with arguments specifying the rounding mode, like
CRlibm.sin(big(0.5), RoundDown)

[fph]

sinpi(@biginterval(0.5)) works on my machine, too, though.

[orkolorko]

with @fph we were debugging an issue with our package InvariantMeasures.jl.
It seems like the fallback to MPFR raises this StackOverFlow error.

In other words, my installation of CRlibm was shadowing MPFR:

julia> CRlibm.sinpi(8.0, RoundDown)
ERROR: StackOverflowError:
Stacktrace:
[1] sinpi(::BigFloat, ::RoundingMode{:Down}) at /home/isaia/.julia/packages/CRlibm/NFCH5/src/CRlibm.jl:137 (repeats 79984 times)

julia> CRlibm.setup(false)
┌ Warning: CRlibm is falling back to use MPFR; it will have
│         the same functionality, but with slower execution.
│         This is currently the only option on Windows.
└ @ CRlibm ~/.julia/packages/CRlibm/NFCH5/src/CRlibm.jl:24
[ Info: CRlibm is shadowing MPFR.
true

Once I ran:

pkg> build CRlibm

julia> using CRlibm

julia> CRlibm.setup(false)
false

julia> CRlibm.sinpi(8.0, RoundUp)
0.0

things started working.

[dpsanders]

I don't understand what is going on here.
Which version of CRlibm do you have?

] st CRlibm

And which OS?

Did you try in a new Julia session?

[orkolorko]

Hi, I'm on Ubuntu 20.10, CRlibm 1.0.1.

I think I found the reason: in CRlibm.jl line 152 we have

const MPFR_function_names = split("""exp expm1 log log1p log2 log10
sin cos tan asin acos atan
sinh cosh 
""")

so, no fallback MPFR function is defined for sinpi and cospi. Adding them to the list gives rise to an error in the test, seems like the output of MPFR is not tight enough.

[dpsanders]

Thanks.

I think the problem is that MPFR does not seem to have a sinpi function implemented?

[orkolorko]

Yes, it is true; when calling sinpi on BigFloat an helper function is called

@which sinpi(BigFloat(1))
sinpi(x::T) where T<:AbstractFloat in Base.Math at special/trig.jl:751

The function above assumes RoundToNearest when reducing x to x/pi and x/pi to [-1, 1]. I think this may be the reason why

sinpi(@biginterval(1))

is defined as

function sinpi(a::Interval{T}) where T
    isempty(a) && return a
    w = a * Interval{T}(π)
    return(sin(w))
end