The type used within e_sqrt.c(__slow_ieee754_sqrtf) was, unnecessarily and
likely inadvertently, double. float is not only appropriate, but also
more efficient, avoiding the need for the compiler to emit a
round-to-single-precision instruction.
This is the difference in compiled code:
0000000000000000 <__ieee754_sqrtf>:
0: 2c 08 20 ec fsqrts f1,f1
- 4: 18 08 20 fc frsp f1,f1
- 8: 20 00 80 4e blr
+ 4: 20 00 80 4e blr
(Found by Anton Blanchard.)
+2019-02-11 Paul A. Clarke <pc@us.ibm.com>
+
+ * sysdeps/powerpc/fpu/e_sqrt.c (__slow_ieee754_sqrtf):
+ Use float instead of double.
+
2019-02-11 TAMUKI Shoichi <tamuki@linet.gr.jp>
* time/strftime_l.c: Fix a few whitespace arrangement inconsistencies.
float
__ieee754_sqrtf (float x)
{
- double z;
+ float z;
#ifdef _ARCH_PPCSQ
asm ("fsqrts %0,%1\n" :"=f" (z):"f" (x));