Fix time-add rounding bug

Without this fix, time arithmetic yielded results that were not
mathematically accurate, even though the exact results were
representable; for example, (time-add 0 1e-13) yielded a timestamp
equal to 0 instead of to 1e-13.
* lisp/timezone.el (timezone-time-from-absolute):
Let time-add do its thing rather than using floating point
internally, which has rounding errors.  We now have bignums and so
don’t need floating point to avoid overflow issues.
* src/timefns.c (timeform_sub_ps_p): New function.
(time_arith): If either argument is a float, represent the
result exactly instead of discarding sub-ps info.
* test/lisp/timezone-tests.el (timezone-tests-time-from-absolute):
Don’t assume (HI LO US PS) timestamp format.
* test/src/emacs-module-tests.el (mod-test-add-nanosecond/valid):
Don’t assume that time-add discards sub-ns info.
* test/src/timefns-tests.el (time-rounding-tests):
Add regression test to detect time-add rounding bug.

1 files changed, 5 insertions, 2 deletions

diff --git a/test/src/emacs-module-tests.el b/test/src/emacs-module-tests.el
index c44c386d30b..c5107847318 100644
--- a/test/src/emacs-module-tests.el
+++ b/test/src/emacs-module-tests.el
@@ -335,12 +335,15 @@ Interactively, you can try hitting \\[keyboard-quit] to quit."
                   ;; New (TICKS . HZ) format.
                   '(123456789 . 1000000000)))
     (ert-info ((format "input: %s" input))
-      (let ((result (mod-test-add-nanosecond input)))
+      (let ((result (mod-test-add-nanosecond input))
+	    (desired-result
+	     (let ((hz 1000000000))
+	       (time-add (time-convert input hz) (cons 1 hz)))))
         (should (consp result))
         (should (integerp (car result)))
         (should (integerp (cdr result)))
         (should (cl-plusp (cdr result)))
-        (should (time-equal-p result (time-add input '(0 0 0 1000))))))))
+        (should (time-equal-p result desired-result))))))
 
 (ert-deftest mod-test-add-nanosecond/nil ()
   (should (<= (float-time (mod-test-add-nanosecond nil))