Correct exponent_max to conform to IEEE definition (#38929)

* Exponent exponent_max returns actual maximum exponent, not the sentinel. Co-authored-by: mschauer <[email protected]>
JuliaLang · Dec 20, 2020 · 58d6684 · 58d6684 · nanosoldier · Dec 21, 2020
1 parent 1887550
commit 58d6684
Show file tree

Hide file tree

Showing 4 changed files with 30 additions and 6 deletions.
diff --git a/base/float.jl b/base/float.jl
@@ -78,11 +78,35 @@ for T in (Float16, Float32, Float64)
  @eval exponent_bits(::Type{$T}) = $(sizeof(T)*8 - significand_bits(T) - 1)
  @eval exponent_bias(::Type{$T}) = $(Int(exponent_one(T) >> significand_bits(T)))
  # maximum float exponent
- @eval exponent_max(::Type{$T}) = $(Int(exponent_mask(T) >> significand_bits(T)) - exponent_bias(T))
+ @eval exponent_max(::Type{$T}) = $(Int(exponent_mask(T) >> significand_bits(T)) - exponent_bias(T) - 1)
  # maximum float exponent without bias
  @eval exponent_raw_max(::Type{$T}) = $(Int(exponent_mask(T) >> significand_bits(T)))
 end
 
+"""
+ exponent_max(T)
+
+Maximum [`exponent`](@ref) value for a floating point number of type `T`.
+
+# Examples
+```jldoctest
+julia> Base.exponent_max(Float64)
+1023
+```
+
+Note, `exponent_max(T) + 1` is a possible value of the exponent field
+with bias, which might be used as sentinel value for `Inf` or `NaN`.
+"""
+function exponent_max end
+
+"""
+ exponent_raw_max(T)
+
+Maximum value of the [`exponent`](@ref) field for a floating point number of type `T` without bias,
+i.e. the maximum integer value representable by [`exponent_bits(T)`](@ref) bits.
+"""
+function exponent_raw_max end
+
 ## conversions to floating-point ##
 
 # TODO: deprecate in 2.0

diff --git a/base/ryu/utils.jl b/base/ryu/utils.jl
@@ -353,7 +353,7 @@ end
 """
 function pow5invsplit_lookup end
 for T in (Float64, Float32, Float16)
- e2_max = exponent_max(T) - precision(T) - 2
+ e2_max = exponent_max(T) - precision(T) - 1
  i_max = log10pow2(e2_max)
  table_sym = Symbol("pow5invsplit_table_", string(T))
  @eval const $table_sym = Tuple(Any[pow5invsplit($T, i) for i = 0:$i_max])

diff --git a/base/special/exp.jl b/base/special/exp.jl
@@ -239,7 +239,7 @@ for (func, base) in (:exp2=>Val(2), :exp=>Val(:ℯ), :exp10=>Val(10))
  twopk = reinterpret(T, (N+Int32(151)) << Int32(23))
  return (twopk*small_part)*(2f0^(-24))
  end
- N == exponent_max(T) && return small_part * T(2.0) * T(2.0)^(exponent_max(T) - 1)
+ N == (exponent_max(T)+1) && return small_part * T(2.0) * T(2.0)^exponent_max(T)
  end
  twopk = reinterpret(T, (N+Int32(127)) << Int32(23))
  return twopk*small_part

diff --git a/base/special/ldexp_exp.jl b/base/special/ldexp_exp.jl
@@ -97,9 +97,9 @@ function _frexp_exp(x::T) where T<:Union{Float32, Float64}
  exp_x = exp(x - kr*log(T(2))) # exp_x*2^k = exp(x)
 
  # Calculate the ks in exp_x*2^ks
- ks = exponent_rshift(T, highword(exp_x)) - (exponent_bias(T) + (exponent_max(T) - 1)) + kr
+ ks = exponent_rshift(T, highword(exp_x)) - (exponent_bias(T) + exponent_max(T)) + kr
 
- # Rescale exp_x to have exponent k2 = exponent_max(T) - 1
- exp_x = modify_exponent(exp_x, UInt32(exponent_bias(T) + (exponent_max(T) - 1)))
+ # Rescale exp_x to have exponent k2 = exponent_max(T)
+ exp_x = modify_exponent(exp_x, UInt32(exponent_bias(T) + exponent_max(T)))
  return exp_x, ks
 end