Fix bug when rounding large numbers to floating point types (#54314)

- fix #52355 using option 4 (round to nearest representable integer)
- update docstrings *including documenting convert to Inf behavior even
though Inf is not the "closest" floating point value*
- add some assorted tests

---------

Co-authored-by: mikmoore <95002244+mikmoore@users.noreply.github.com>
(cherry picked from commit e7a1def)

base/essentials.jl

@@ -397,8 +397,9 @@ Stacktrace:
 [...]
 ```
 
-If `T` is a [`AbstractFloat`](@ref) type,
-then it will return the closest value to `x` representable by `T`.
+If `T` is a [`AbstractFloat`](@ref) type, then it will return the
+closest value to `x` representable by `T`. Inf is treated as one
+ulp greater than `floatmax(T)` for purposes of determining nearest.
 
 ```jldoctest
 julia> x = 1/3

base/float.jl

@@ -464,6 +464,19 @@ round(x::IEEEFloat, ::RoundingMode{:Down})    = floor_llvm(x)
 round(x::IEEEFloat, ::RoundingMode{:Up})      = ceil_llvm(x)
 round(x::IEEEFloat, ::RoundingMode{:Nearest}) = rint_llvm(x)
 
+rounds_up(x, ::RoundingMode{:Down}) = false
+rounds_up(x, ::RoundingMode{:Up}) = true
+rounds_up(x, ::RoundingMode{:ToZero}) = signbit(x)
+rounds_up(x, ::RoundingMode{:FromZero}) = !signbit(x)
+function _round_convert(::Type{T}, x_integer, x, r::Union{RoundingMode{:ToZero}, RoundingMode{:FromZero}, RoundingMode{:Up}, RoundingMode{:Down}}) where {T<:AbstractFloat}
+    x_t = convert(T, x_integer)
+    if rounds_up(x, r)
+        x_t < x ? nextfloat(x_t) : x_t
+    else
+        x_t > x ? prevfloat(x_t) : x_t
+    end
+end
+
 ## floating point promotions ##
 promote_rule(::Type{Float32}, ::Type{Float16}) = Float32
 promote_rule(::Type{Float64}, ::Type{Float16}) = Float64

base/rounding.jl

@@ -338,6 +338,10 @@ The [`RoundingMode`](@ref) `r` controls the direction of the rounding; the defau
 of 0.5) being rounded to the nearest even integer. Note that `round` may give incorrect
 results if the global rounding mode is changed (see [`rounding`](@ref)).
 
+When rounding to a floating point type, will round to integers representable by that type
+(and Inf) rather than true integers. Inf is treated as one ulp greater than the
+`floatmax(T)` for purposes of determining "nearest", similar to [`convert`](@ref).
+
 # Examples
 ```jldoctest
 julia> round(1.7)
@@ -363,6 +367,12 @@ julia> round(123.456; sigdigits=2)
 
 julia> round(357.913; sigdigits=4, base=2)
 352.0
+
+julia> round(Float16, typemax(UInt128))
+Inf16
+
+julia> floor(Float16, typemax(UInt128))
+Float16(6.55e4)
 ```
 
 !!! note
@@ -466,6 +476,7 @@ floor(::Type{T}, x) where T = round(T, x, RoundDown)
  ceil(::Type{T}, x) where T = round(T, x, RoundUp)
 round(::Type{T}, x) where T = round(T, x, RoundNearest)
 
-round(::Type{T}, x, r::RoundingMode) where T = convert(T, round(x, r))
+round(::Type{T}, x, r::RoundingMode) where T = _round_convert(T, round(x, r), x, r)
+_round_convert(::Type{T}, x_integer, x, r) where T = convert(T, x_integer)
 
 round(x::Integer, r::RoundingMode) = x

test/rounding.jl

@@ -458,3 +458,15 @@ end
     @test_throws InexactError round(Int128, -Inf16)
     # More comprehensive testing is present in test/floatfuncs.jl
 end
+
+@testset "floor(<:AbstractFloat, large_number) (#52355)" begin
+    @test floor(Float32, 0xffff_ffff) == prevfloat(2f0^32) <= 0xffff_ffff
+    @test trunc(Float16, typemax(UInt128)) == floatmax(Float16)
+    @test round(Float16, typemax(UInt128)) == Inf16
+    for i in [-BigInt(floatmax(Float64)), -BigInt(floatmax(Float64))*100, BigInt(floatmax(Float64)), BigInt(floatmax(Float64))*100]
+        f = ceil(Float64, i)
+        @test f >= i
+        @test isinteger(f) || isinf(f)
+        @test prevfloat(f) < i
+    end
+end