Add general fallback for rem2pi

[11happy]

Overview:

• This PR fixes rem2pi(x::Integer, r::RoundingMode) is missing. #44715 by adding a general fallback.

Testing:

• Tested the updated code.
• Verified that other functionalities remain unaffected.

Dependencies:

• No dependencies on other pull requests.

CC:

• @N5N3

[barucden]

Are the other methods (for Int32 and Int64) any better than this fallback? I see that the one for Int32 omits the condition, but the one for Int64 seems exactly like the newly-added, general method.

[11happy]

Are the other methods (for Int32 and Int64) any better than this fallback? I see that the one for Int32 omits the condition, but the one for Int64 seems exactly like the newly-added, general method.

like there's not much significant difference, Int32 just omits the condition.

[barucden]

The method for Int32 avoids the check x == Float64(x) likely because Int32 is always representable as Float64 exactly. The same holds even for Int8 and Int16, right? We might extend that method to accept Union{Int8, Int16, Int32}.

[ViralBShah]

cc @oscardssmith

[11happy]

The method for Int32 avoids the check x == Float64(x) likely because Int32 is always representable as Float64 exactly. The same holds even for Int8 and Int16, right? We might extend that method to accept Union{Int8, Int16, Int32}.

sure could do that , should I?

[N5N3]

sure could do that , should I?

I guess this is not a must. LLVM is clever enough to eliminate the error path. (So perhaps there's no need to keep the Int32 dispatch)

julia> function myrem2pi(x::Integer, r::RoundingMode)                                          
           fx = float(x)                                                                       
               fx == x || throw(ArgumentError(LazyString(typeof(x), " argument to rem2pi is too large: ", x)))                                                                                
           rem2pi(fx, r)                                                                       
       end
myrem2pi (generic function with 1 method)

julia> @code_llvm myrem2pi(Int32(1), RoundUp)
; Function Signature: myrem2pi(Int32, Base.Rounding.RoundingMode{:Up})
;  @ REPL[1]:1 within `myrem2pi`
; Function Attrs: uwtable
define double @julia_myrem2pi_6707(i32 signext %"x::Int32") #0 {
L5:
;  @ REPL[1]:2 within `myrem2pi`
; ┌ @ float.jl:356 within `float`
; │┌ @ float.jl:329 within `AbstractFloat`
; ││┌ @ float.jl:221 within `Float64`
     %0 = sitofp i32 %"x::Int32" to double
; └└└
;  @ REPL[1]:4 within `myrem2pi`
  %1 = call double @j_rem2pi_6713(double %0)
  ret double %1
}

[barucden]

What about removing the Int32 method then, as you suggested @N5N3?

[N5N3]

@barucden That make sense to me. Perhaps we need some test to ensure there's no regression.