suspicious-general-float-equality
Referencebit-exact equality on floats
pedantic · warning · stable · since 1.0 · fix: never · off by default
= and <> on floats ask for bit-exact equality of values that
arithmetic rarely reproduces bit-exactly: 0.1 +. 0.2 <> 0.3.
(* bad *) weight = 0.1(* good *) Float.abs (weight -. 0.1) < epsilonFires when = or <> resolves to its Stdlib declaration and at
least one operand’s type head is the predefined float. Comparisons
against exactly representable anchors are excluded: a float literal
denoting zero (0., -0.0, 0x0p0) and the infinity and
neg_infinity constants. A nan constant operand is
invalid-nan-comparison’s finding, never this rule’s — the two rules
partition float =/<> exactly. Orderings are meaningful on floats
and out of scope; type abbreviations over float are conservatively
clean (no expansion); shadowed operators do not fire. Code that means
bit-exact equality writes Float.equal x y — a distinct declaration
this rule never matches — or carries [@litany.allow "suspicious-general-float-equality: exact comparison intended"] where
the operator form must stay. No fix: a margin-of-error rewrite is
domain-specific.

