From 2937a4d323722a4c14824c3ab4fbc737e382bca8 Mon Sep 17 00:00:00 2001
From: Andreas Noack <andreas@noack.dk>
Date: Tue, 8 Sep 2026 10:52:47 +0200
Subject: [PATCH] Use IEEE comparison in binindex

Pass `lt = <` to `searchsortedfirst`/`searchsortedlast` in `_edge_binindex`
instead of relying on `isless`. Since `-0.0 < 0.0` is false, -0.0 and 0.0 are
binned identically without normalizing the inputs, so `_normalize_zero`, the
separate `AbstractRange` method, and the constructor check rejecting ranges
containing -0.0 can all go. `<` is also cheaper than `isless`: `fit(Histogram)`
is about 20% faster with range edges and about 2x faster with vector edges.
---
 src/hist.jl  | 31 +++++++------------------------
 test/hist.jl | 16 ++++++++++++++--
 2 files changed, 21 insertions(+), 26 deletions(-)

diff --git a/src/hist.jl b/src/hist.jl
index 08f9e3920..57cbb7e32 100644
--- a/src/hist.jl
+++ b/src/hist.jl
@@ -197,14 +197,6 @@ mutable struct Histogram{T<:Real,N,E} <: AbstractHistogram{T,N,E}
         closed == :right || closed == :left || error("closed must :left or :right")
         isdensity && !(T <: AbstractFloat) && error("Density histogram must have float-type weights")
         _edges_nbins(edges) == size(weights) || error("Histogram edge vectors must be 1 longer than corresponding weight dimensions")
-        # We do not handle -0.0 in ranges correctly in `binindex` for performance
-        # Constructing ranges starting or ending with -0.0 is very hard,
-        # and ranges containing -0.0 elsewhere virtually impossible,
-        # but check this just in case as it is cheap
-        foreach(edges) do e
-            e isa AbstractRange && any(isequal(-0.0), e) &&
-                throw(ArgumentError("ranges containing -0.0 not allowed in edges"))
-        end
         new{T,N,E}(edges,weights,closed,isdensity)
     end
 end
@@ -240,25 +232,16 @@ binindex(h::AbstractHistogram{T,1}, x::Real) where {T} = binindex(h, (x,))[1]
 binindex(h::Histogram{T,N}, xs::NTuple{N,Real}) where {T,N} =
     map((edge, x) -> _edge_binindex(edge, h.closed, x), h.edges, xs)
 
-_normalize_zero(x::AbstractFloat) = isequal(x, -0.0) ? zero(x) : x
-_normalize_zero(x::Any) = x
-
-# Always treat -0.0 like 0.0
+# Compare with `<` rather than the default `isless`: `isless(-0.0, 0.0)` is true, so -0.0
+# would be binned differently from 0.0, whereas `-0.0 < 0.0` is false and the two are
+# treated as equal. NaN compares false with everything under `<`, so it ends up outside
+# the edges (bin index 0 or `length(edge)`) and is dropped by `push!` as before. `<` is also
+# cheaper than `isless` and keeps the arithmetic fast path for ranges.
 @inline function _edge_binindex(edge::AbstractVector, closed::Symbol, x::Real)
     if closed === :right
-        return searchsortedfirst(edge, _normalize_zero(x), by=_normalize_zero) - 1
-    else
-        return searchsortedlast(edge, _normalize_zero(x), by=_normalize_zero)
-    end
-end
-# Passing by=_normalize_zero for ranges would have a large performance hit
-# as it would force using the AbstractVector fallback
-# This is not worth it given that it is very difficult to construct a range containing -0.0
-@inline function _edge_binindex(edge::AbstractRange, closed::Symbol, x::Real)
-    if closed === :right
-        return searchsortedfirst(edge, _normalize_zero(x)) - 1
+        return searchsortedfirst(edge, x, lt = <) - 1
     else
-        return searchsortedlast(edge, _normalize_zero(x))
+        return searchsortedlast(edge, x, lt = <)
     end
 end
 
diff --git a/test/hist.jl b/test/hist.jl
index 1fe57237f..038367e3a 100644
--- a/test/hist.jl
+++ b/test/hist.jl
@@ -295,8 +295,20 @@ end
         fit(Histogram, [0.0, 1.0], -0.5:0.5:0.0, closed=:right) ==
         fit(Histogram, [-0.0, 1.0], -0.5:0.5:0.0, closed=:right)
 
-    @test_throws ArgumentError fit(Histogram, [-0.5], LinRange(-1.0, -0.0, 3))
-    @test_throws ArgumentError fit(Histogram, [-0.5], UnitRange(-0.0, 1.0))
+    # ranges containing -0.0 behave like the corresponding ranges with 0.0
+    @test fit(Histogram, [-0.5, 0.0], LinRange(-1.0, -0.0, 3)) ==
+        fit(Histogram, [-0.5, 0.0], LinRange(-1.0, 0.0, 3))
+    @test fit(Histogram, [-0.5, 0.0], LinRange(-1.0, -0.0, 3), closed=:right) ==
+        fit(Histogram, [-0.5, -0.0], LinRange(-1.0, 0.0, 3), closed=:right)
+    @test fit(Histogram, [0.0, 0.5], UnitRange(-0.0, 1.0)) ==
+        fit(Histogram, [-0.0, 0.5], UnitRange(0.0, 1.0))
+    @test fit(Histogram, [0.0, 0.5], UnitRange(-0.0, 1.0), closed=:right) ==
+        fit(Histogram, [-0.0, 0.5], UnitRange(0.0, 1.0), closed=:right)
+    # NaN is dropped
+    @test fit(Histogram, [NaN, 0.5], 0.0:0.5:1.0).weights == [0, 1]
+    @test fit(Histogram, [NaN, 0.5], 0.0:0.5:1.0, closed=:right).weights == [1, 0]
+    @test fit(Histogram, [NaN, 0.5], [0.0, 0.5, 1.0]).weights == [0, 1]
+    @test fit(Histogram, [NaN, 0.5], [0.0, 0.5, 1.0], closed=:right).weights == [1, 0]
 end
 
 end # @testset "StatsBase.Histogram"