With the new benchmark in place, I measured the run time of
- the glibc 2.35 implementation of totalorder,
- the gnulib implementation (picked by configuring with
gl_cv_func_totalorder_in_libm=no gl_cv_func_totalorder_no_libm=no \
gl_cv_func_totalorderf_in_libm=no gl_cv_func_totalorderf_no_libm=no \
gl_cv_func_totalorderl_in_libm=no gl_cv_func_totalorderl_no_libm=no \
- the gnulib implementation with some disabled NaN tests.
This change (see attached patch) is correct: it still passes the unit
tests.
Here are the running times (on x86_64) of "./bench-totalorder fdl 1000000":
f d l
glibc 1.816 1.671 2.078
gnulib 1.445 1.425 8.690
gnulib with patch 1.798 1.974 14.032
Conclusion:
* My patch is a slowdown. It apparently "optimized" the fast path away. :-D
* The gnulib implementation is significantly faster than glibc, except for
the long-double case. I'll redo the measurements on various CPU types and
then tell the glibc people...
Kudos to you, Paul, for an implementation that is not only standards-compliant
and portable, but also faster than glibc!
Bruno
diff --git a/lib/totalorder.c b/lib/totalorder.c
index ee90ebb1f9..212e549305 100644
--- a/lib/totalorder.c
+++ b/lib/totalorder.c
@@ -21,6 +21,8 @@
/* Specification. */
#include <math.h>
+#include <float.h>
+
int
totalorder (double const *x, double const *y)
{
@@ -31,19 +33,27 @@ totalorder (double const *x, double const *y)
if (!xs != !ys)
return xs;
- /* If one of *X, *Y is a NaN and the other isn't, the answer is easy
- as well: the negative NaN is "smaller", the positive NaN is "greater"
- than the other argument. */
- int xn = isnand (*x);
- int yn = isnand (*y);
- if (!xn != !yn)
- return !xn == !xs;
- /* If none of *X, *Y is a NaN, the '<=' operator does the job, including
- for -Infinity and +Infinity. */
- if (!xn)
- return *x <= *y;
-
- /* At this point, *X and *Y are NaNs with the same sign bit. */
+#if !(defined __hppa || defined __mips__ || defined __sh__)
+ /* If the 'double' type is as described by IEEE 754:2008 = ISO/IEC 60559:2011,
+ we can skip the 'isnand' tests.
+ Note that while gcc 13 defines __DBL_IS_IEC_60559__, clang 16 doesn't. */
+ if (!(DBL_MAX_EXP == 1024 && DBL_MIN_EXP == -1021))
+#endif
+ {
+ /* If one of *X, *Y is a NaN and the other isn't, the answer is easy
+ as well: the negative NaN is "smaller", the positive NaN is "greater"
+ than the other argument. */
+ int xn = isnand (*x);
+ int yn = isnand (*y);
+ if (!xn != !yn)
+ return !xn == !xs;
+ /* If none of *X, *Y is a NaN, the '<=' operator does the job, including
+ for -Infinity and +Infinity. */
+ if (!xn)
+ return *x <= *y;
+
+ /* At this point, *X and *Y are NaNs with the same sign bit. */
+ }
unsigned long long extended_sign = -!!xs;
#if defined __hppa || defined __mips__ || defined __sh__
diff --git a/lib/totalorderf.c b/lib/totalorderf.c
index fa2a1245de..c1137d5150 100644
--- a/lib/totalorderf.c
+++ b/lib/totalorderf.c
@@ -21,6 +21,8 @@
/* Specification. */
#include <math.h>
+#include <float.h>
+
int
totalorderf (float const *x, float const *y)
{
@@ -31,19 +33,27 @@ totalorderf (float const *x, float const *y)
if (!xs != !ys)
return xs;
- /* If one of *X, *Y is a NaN and the other isn't, the answer is easy
- as well: the negative NaN is "smaller", the positive NaN is "greater"
- than the other argument. */
- int xn = isnanf (*x);
- int yn = isnanf (*y);
- if (!xn != !yn)
- return !xn == !xs;
- /* If none of *X, *Y is a NaN, the '<=' operator does the job, including
- for -Infinity and +Infinity. */
- if (!xn)
- return *x <= *y;
-
- /* At this point, *X and *Y are NaNs with the same sign bit. */
+#if !(defined __hppa || defined __mips__ || defined __sh__)
+ /* If the 'float' type is as described by IEEE 754:2008 = ISO/IEC 60559:2011,
+ we can skip the 'isnanf' tests.
+ Note that while gcc 13 defines __FLT_IS_IEC_60559__, clang 16 doesn't. */
+ if (!(FLT_MAX_EXP == 128 && FLT_MIN_EXP == -125))
+#endif
+ {
+ /* If one of *X, *Y is a NaN and the other isn't, the answer is easy
+ as well: the negative NaN is "smaller", the positive NaN is "greater"
+ than the other argument. */
+ int xn = isnanf (*x);
+ int yn = isnanf (*y);
+ if (!xn != !yn)
+ return !xn == !xs;
+ /* If none of *X, *Y is a NaN, the '<=' operator does the job, including
+ for -Infinity and +Infinity. */
+ if (!xn)
+ return *x <= *y;
+
+ /* At this point, *X and *Y are NaNs with the same sign bit. */
+ }
unsigned int extended_sign = -!!xs;
#if defined __hppa || defined __mips__ || defined __sh__
diff --git a/lib/totalorderl.c b/lib/totalorderl.c
index a06e70d55f..e00ca3c68a 100644
--- a/lib/totalorderl.c
+++ b/lib/totalorderl.c
@@ -37,19 +37,27 @@ totalorderl (long double const *x, long double const *y)
if (!xs != !ys)
return xs;
- /* If one of *X, *Y is a NaN and the other isn't, the answer is easy
- as well: the negative NaN is "smaller", the positive NaN is "greater"
- than the other argument. */
- int xn = isnanl (*x);
- int yn = isnanl (*y);
- if (!xn != !yn)
- return !xn == !xs;
- /* If none of *X, *Y is a NaN, the '<=' operator does the job, including
- for -Infinity and +Infinity. */
- if (!xn)
- return *x <= *y;
-
- /* At this point, *X and *Y are NaNs with the same sign bit. */
+#if !(defined __hppa || defined __mips__ || defined __sh__)
+ /* If the 'long double' type is as described by IEEE 754:2008 =
+ ISO/IEC 60559:2011, we can skip the 'isnanl' tests.
+ Note that while gcc 13 defines __LDBL_IS_IEC_60559__, clang 16 doesn't. */
+ if (!(LDBL_MAX_EXP == 16384 && LDBL_MIN_EXP == -16381))
+#endif
+ {
+ /* If one of *X, *Y is a NaN and the other isn't, the answer is easy
+ as well: the negative NaN is "smaller", the positive NaN is "greater"
+ than the other argument. */
+ int xn = isnanl (*x);
+ int yn = isnanl (*y);
+ if (!xn != !yn)
+ return !xn == !xs;
+ /* If none of *X, *Y is a NaN, the '<=' operator does the job, including
+ for -Infinity and +Infinity. */
+ if (!xn)
+ return *x <= *y;
+
+ /* At this point, *X and *Y are NaNs with the same sign bit. */
+ }
unsigned long long extended_sign = -!!xs;
diff --git a/modules/totalorder b/modules/totalorder
index 00f190c42e..5d04d02fd4 100644
--- a/modules/totalorder
+++ b/modules/totalorder
@@ -9,6 +9,7 @@ m4/totalorder.m4
Depends-on:
math
extensions
+float [test $HAVE_TOTALORDER = 0 || test $REPLACE_TOTALORDER = 1]
isnand [test $HAVE_TOTALORDER = 0 || test $REPLACE_TOTALORDER = 1]
signbit [test $HAVE_TOTALORDER = 0 || test $REPLACE_TOTALORDER = 1]
diff --git a/modules/totalorderf b/modules/totalorderf
index ab1569c9cc..07c881c9a4 100644
--- a/modules/totalorderf
+++ b/modules/totalorderf
@@ -9,6 +9,7 @@ m4/totalorder.m4
Depends-on:
math
extensions
+float [test $HAVE_TOTALORDERF = 0 || test $REPLACE_TOTALORDERF = 1]
isnanf [test $HAVE_TOTALORDERF = 0 || test $REPLACE_TOTALORDERF = 1]
signbit [test $HAVE_TOTALORDERF = 0 || test $REPLACE_TOTALORDERF = 1]
