Fix powf inaccuracy (bug 18956).

The flt-32 version of powf can be inaccurate because of bugs in the
extra-precision calculation of (x-1)/(x+1) or (x-1.5)/(x+1.5) as part
of calculating log(x) with extra precision: a constant used (as part
of adding 1 or 1.5 through integer arithmetic) is incorrect, and then
the code fails to mask a computed high part before using it in
arithmetic that relies on s_h*t_h being exactly representable.  This
patch fixes these bugs.

Tested for x86_64 and x86.  x86_64 ulps for powf removed and
regenerated to reflect reduced ulps from the increased accuracy for
existing tests.

	[BZ #18956]
	* sysdeps/ieee754/flt-32/e_powf.c (__ieee754_powf): Add 0x00400000
	not 0x0040000 for high bit of mantissa.  Mask with 0xfffff000 when
	extracting high part.
	* math/auto-libm-test-in: Add another test of pow.
	* math/auto-libm-test-out: Regenerated.
	* sysdeps/x86_64/fpu/libm-test-ulps: Update.

1 files changed, 1 insertions, 0 deletions

diff --git a/math/auto-libm-test-in b/math/auto-libm-test-in
index 4b93857efb..e86be23fc8 100644
--- a/math/auto-libm-test-in
+++ b/math/auto-libm-test-in
@@ -3219,6 +3219,7 @@ pow 0x1.ce78f2p+0 -0x2.7f1f78p+4
 pow 0xf.fffffp+124 -0x5.b5b648p+0
 pow 0x1.430d4cp+0 0x5.0e462p+4
 pow 0x9.8b82ap-4 -0x1.99907ap+12
+pow 0xd.73035p-4 -0x1.47bb8p+8
 
 sin 0
 sin -0