math: change the formula used for acos.s

old: 2*atan2(sqrt(1-x),sqrt(1+x))
new: atan2(fabs(sqrt((1-x)*(1+x))),x)
improvements:
* all edge cases are fixed (sign of zero in downward rounding)
* a bit faster (here a single call is about 131ns vs 162ns)
* a bit more precise (at most 1ulp error on 1M uniform random
samples in [0,1), the old formula gave some 2ulp errors as well)

diff --git a/src/math/i386/acos.s b/src/math/i386/acos.s
index bfff0c5c9a89d3332ea7df6a8e5d3cd0a68a01f3..47f365efb1ea7493c34f7d64b7db33761e4337bd 100644 (file)
--- a/src/math/i386/acos.s
+++ b/src/math/i386/acos.s
@@ -1,3 +1,5 @@
+# use acos(x) = atan2(fabs(sqrt((1-x)*(1+x))), x)
+
 .global acosf
 .type acosf,@function
 acosf:
@@ -14,17 +16,13 @@ acosl:
 .type acos,@function
 acos:
        fldl 4(%esp)
-1:     fld1
-       fld %st(1)
+1:     fld %st(0)
        fld1
-       fsubp
-       fsqrt
-       fxch %st(2)
-       faddp
+       fsub %st(0),%st(1)
+       fadd %st(2)
+       fmulp
        fsqrt
+       fabs         # fix sign of zero (matters in downward rounding mode)
+       fxch %st(1)
        fpatan
-       fld1
-       fld1
-       faddp
-       fmulp
        ret

diff --git a/src/math/x86_64/acosl.s b/src/math/x86_64/acosl.s
index db68d2de268cd83498e0ad43bfababdb2790608d..88e01b49a2bbe72cbb3332e15fa71a9134a4a99c 100644 (file)
--- a/src/math/x86_64/acosl.s
+++ b/src/math/x86_64/acosl.s
@@ -1,18 +1,16 @@
+# see ../i386/acos.s
+
 .global acosl
 .type acosl,@function
 acosl:
        fldt 8(%rsp)
+1:     fld %st(0)
        fld1
-       fld %st(1)
-       fld1
-       fsubp
-       fsqrt
-       fxch %st(2)
-       faddp
+       fsub %st(0),%st(1)
+       fadd %st(2)
+       fmulp
        fsqrt
+       fabs
+       fxch %st(1)
        fpatan
-       fld1
-       fld1
-       faddp
-       fmulp
        ret