Fix bug with eql etc. on NaNs

Fix a bug where eql, sxhash-eql, memql, and make-hash-table
were not consistent on NaNs.  Likewise for equal,
sxhash-equal, member, and make-hash-table.  Some of these
functions ignored NaN significands, whereas others treated
them as significant.  It's more logical to treat significands
as significant, and this typically makes eql a bit more
efficient on floats, with just one integer comparison instead
of one to three floating-point comparisons.
* doc/lispref/numbers.texi (Float Basics): Document that
NaNs are never numerically equal, but might be eql.
* src/fns.c (WORDS_PER_DOUBLE): Move to top level of this file.
(union double_and_words): Now named, and at the top level of this file.
(same_float): New function.
(Fmemql, Feql, internal_equal, cmpfn_eql): Use it, so that
the corresponding functions treat NaNs consistently.
(sxhash_float): Simplify based on above-mentioned changes.

* test/src/fns-tests.el (fns-tests-equality-nan): New test.

1 files changed, 11 insertions, 0 deletions

diff --git a/test/src/fns-tests.el b/test/src/fns-tests.el
index d9cca557cf2..e4b9cbe25a4 100644
--- a/test/src/fns-tests.el
+++ b/test/src/fns-tests.el
@@ -23,6 +23,17 @@
 
 (require 'cl-lib)
 
+;; Test that equality predicates work correctly on NaNs when combined
+;; with hash tables based on those predicates.  This was not the case
+;; for eql in Emacs 26.
+(ert-deftest fns-tests-equality-nan ()
+  (dolist (test (list #'eq #'eql #'equal))
+    (let* ((h (make-hash-table :test test))
+           (nan 0.0e+NaN)
+           (-nan (- nan)))
+      (puthash nan t h)
+      (should (eq (funcall test nan -nan) (gethash -nan h))))))
+
 (ert-deftest fns-tests-reverse ()
   (should-error (reverse))
   (should-error (reverse 1))