[FFmpeg-devel] [PATCH] lavu/common: use fabs(f) for FFABS/FFNABS

[Rémi Denis-Courmont]

The absolute value of a floating point value is easier to calculate than
that of an integer: it is obtained by either clearing the sign or xoring it
with itself. Accordingly, Arm, LoongArch, RISC-V have dedicated instructions
to manipulate float sign bits. x86-64 has ANDPS.

As per C11 appendix F, FFABS() is not quite the same as fabs() however,
due to differing NaN handling. GCC is able to optimise the current FFABS
definition to match fabs() anyway with just -fno-signed-zeros. But Clang is
evidently not doing so and inserts a comparison with 0 and a conditional
select or branch. Now that we have C11, this is easy to fix properly.
(Another option is to manually audit which of the ~880 FFABS/FFNABS use
sites involve floating point values.)

Note that this still preserves the old definitions for C++ and pre-C11
compilers since the 2 macros are public (even though they start with FF).
---
 libavutil/common.h | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/libavutil/common.h b/libavutil/common.h
index 3b830daf30..48761885fb 100644
--- a/libavutil/common.h
+++ b/libavutil/common.h
@@ -71,7 +71,14 @@
  * as with *abs()
  * @see FFNABS()
  */
+#if !defined (__cplusplus) && __STDC_VERSION__ >= 201112L
+#define FFABS(a) _Generic(a, \
+                          float: fabsf(a), \
+                          double: fabs(a), \
+                          default: ((a) >= 0) ? (a) : (-(a)))
+#else
 #define FFABS(a) ((a) >= 0 ? (a) : (-(a)))
+#endif
 #define FFSIGN(a) ((a) > 0 ? 1 : -1)
 
 /**
@@ -80,7 +87,14 @@
  * As with many macros, this evaluates its argument twice, it thus must not have
  * a sideeffect, that is FFNABS(x++) has undefined behavior.
  */
+#if !defined (__cplusplus) && __STDC_VERSION__ >= 201112L
+#define FFNABS(a) _Generic(a, \
+                           float: -fabsf(a), \
+                           double: -fabs(a), \
+                           default: ((a) <= 0) ? (a) : (-(a)))
+#else
 #define FFNABS(a) ((a) <= 0 ? (a) : (-(a)))
+#endif
 
 /**
  * Unsigned Absolute value.
-- 
2.45.2