bug-gnulib
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[PATCH 2/2] Optimize a few more three-valued comparisons


From: Paul Eggert
Subject: [PATCH 2/2] Optimize a few more three-valued comparisons
Date: Fri, 24 Jul 2020 10:54:59 -0700

* lib/timespec.h (timespec_cmp, timespec_sign):
* lib/utimecmp.c (utimecmpat):
Avoid conditional branches by using _GL_CMP.
---
 ChangeLog      |  5 +++++
 lib/timespec.h | 40 +++-------------------------------------
 lib/utimecmp.c |  6 ++----
 3 files changed, 10 insertions(+), 41 deletions(-)

diff --git a/ChangeLog b/ChangeLog
index 01577c9bb..b922dca63 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,10 @@
 2020-07-24  Paul Eggert  <eggert@cs.ucla.edu>
 
+       Optimize a few more three-valued comparisons
+       * lib/timespec.h (timespec_cmp, timespec_sign):
+       * lib/utimecmp.c (utimecmpat):
+       Avoid conditional branches by using _GL_CMP.
+
        Fix _GL_CMP parenthesization typo
        * m4/gnulib-common.m4 (_GL_CMP): Properly parenthesize.
 
diff --git a/lib/timespec.h b/lib/timespec.h
index 02684ce6e..f88f1e47a 100644
--- a/lib/timespec.h
+++ b/lib/timespec.h
@@ -59,46 +59,12 @@ make_timespec (time_t s, long int ns)
   return r;
 }
 
-/* Return negative, zero, positive if A < B, A == B, A > B, respectively.
-
-   For each timestamp T, this code assumes that either:
-
-     * T.tv_nsec is in the range 0..999999999; or
-     * T.tv_sec corresponds to a valid leap second on a host that supports
-       leap seconds, and T.tv_nsec is in the range 1000000000..1999999999; or
-     * T.tv_sec is the minimum time_t value and T.tv_nsec is -1; or
-       T.tv_sec is the maximum time_t value and T.tv_nsec is 2000000000.
-       This allows for special struct timespec values that are less or
-       greater than all possible valid timestamps.
-
-   In all these cases, it is safe to subtract two tv_nsec values and
-   convert the result to integer without worrying about overflow on
-   any platform of interest to the GNU project, since all such
-   platforms have 32-bit int or wider.
-
-   Replacing "a.tv_nsec - b.tv_nsec" with something like
-   "a.tv_nsec < b.tv_nsec ? -1 : a.tv_nsec > b.tv_nsec" would cause
-   this function to work in some cases where the above assumption is
-   violated, but not in all cases (e.g., a.tv_sec==1, a.tv_nsec==-2,
-   b.tv_sec==0, b.tv_nsec==999999999) and is arguably not worth the
-   extra instructions.  Using a subtraction has the advantage of
-   detecting some invalid cases on platforms that detect integer
-   overflow.  */
+/* Return negative, zero, positive if A < B, A == B, A > B, respectively.  */
 
 _GL_TIMESPEC_INLINE int _GL_ATTRIBUTE_PURE
 timespec_cmp (struct timespec a, struct timespec b)
 {
-  if (a.tv_sec < b.tv_sec)
-    return -1;
-  if (a.tv_sec > b.tv_sec)
-    return 1;
-
-  /* Pacify gcc -Wstrict-overflow (bleeding-edge circa 2017-10-02).  See:
-     https://lists.gnu.org/r/bug-gnulib/2017-10/msg00006.html  */
-  assume (-1 <= a.tv_nsec && a.tv_nsec <= 2 * TIMESPEC_HZ);
-  assume (-1 <= b.tv_nsec && b.tv_nsec <= 2 * TIMESPEC_HZ);
-
-  return a.tv_nsec - b.tv_nsec;
+  return 2 * _GL_CMP (a.tv_sec, b.tv_sec) + _GL_CMP (a.tv_nsec, b.tv_nsec);
 }
 
 /* Return -1, 0, 1, depending on the sign of A.  A.tv_nsec must be
@@ -106,7 +72,7 @@ timespec_cmp (struct timespec a, struct timespec b)
 _GL_TIMESPEC_INLINE int _GL_ATTRIBUTE_PURE
 timespec_sign (struct timespec a)
 {
-  return a.tv_sec < 0 ? -1 : a.tv_sec || a.tv_nsec;
+  return _GL_CMP (a.tv_sec, 0) + (!a.tv_sec & !!a.tv_nsec);
 }
 
 struct timespec timespec_add (struct timespec, struct timespec)
diff --git a/lib/utimecmp.c b/lib/utimecmp.c
index 600d76d90..341c68508 100644
--- a/lib/utimecmp.c
+++ b/lib/utimecmp.c
@@ -396,8 +396,6 @@ utimecmpat (int dfd, char const *dst_name,
     }
 
   /* Compare the timestamps and return -1, 0, 1 accordingly.  */
-  return (dst_s < src_s ? -1
-          : dst_s > src_s ? 1
-          : dst_ns < src_ns ? -1
-          : dst_ns > src_ns);
+  return (_GL_CMP (dst_s, src_s)
+          + ((dst_s == src_s ? ~0 : 0) & _GL_CMP (dst_ns, src_ns)));
 }
-- 
2.25.4




reply via email to

[Prev in Thread] Current Thread [Next in Thread]