This patch makes ARM sqrt and sqrtf use the VSQRT VFP square root
instruction when available, instead of much larger generic code for
computing square roots.
Now, GCC will normally inline sqrt calls except for negative arguments
where errno needs to be set, and because the benchtests fail to use
-fno-builtin that means no significant difference in benchmark results
for sqrt (note, however, there are lots of __ieee754_sqrt calls
internally in libm, which are *not* inlined - although some
architectures define __ieee754_sqrt in their math_private.h for that
purpose, ARM doesn't - so improving out-of-line sqrt performance is
still relevant to those other functions, if not for most ordinary
direct users of sqrt). With the benchtests changed to use
-fno-builtin for sqrt tests, typical performance results before the
change are ("max" is wildly varying in any case):
"duration": 9.88358e+09,
"iterations": 4.8783e+07,
"max": 457.764,
"min": 183.105,
"mean": 202.603
and after it are:
"duration": 9.45663e+09,
"iterations": 2.24385e+08,
"max": 274.659,
"min": 30.517,
"mean": 42.1447
Tested for ARM (hard-float and soft-float).
[BZ #20660]
* sysdeps/arm/e_sqrt.c: New file.
* sysdeps/arm/e_sqrtf.c: Likewise.
