Extract the high and low words together. With gcc-3.4 on uniformly

distributed non-large args, this saves about 14 of 134 cycles for
Athlon64s and about 5 of 199 cycles for AthlonXPs.

Moved the check for x == 0 inside the check for subnormals.  With
gcc-3.4 on uniformly distributed non-large args, this saves another
5 cycles on Athlon64s and loses 1 cycle on AthlonXPs.

Use INSERT_WORDS() and not SET_HIGH_WORD() when converting the first
approximation from bits to double.  With gcc-3.4 on uniformly distributed
non-large args, this saves another 4 cycles on both Athlon64s and and
AthlonXPs.

Accessing doubles as 2 words may be an optimization on old CPUs, but on
current CPUs it tends to cause extra operations and pipeline stalls,
especially for writes, even when only 1 of the words needs to be accessed.

Removed an unused variable.

1 files changed, 6 insertions, 8 deletions

diff --git a/lib/msun/src/s_cbrt.c b/lib/msun/src/s_cbrt.c
index 29e53e2..545b0e7 100644
--- a/lib/msun/src/s_cbrt.c
+++ b/lib/msun/src/s_cbrt.c
@@ -43,17 +43,13 @@ cbrt(double x)
 	    uint64_t bits;
 	} u;
 	double r,s,t=0.0,w;
-	uint64_t bits;
 	u_int32_t sign;
 	u_int32_t high,low;
 
-	GET_HIGH_WORD(hx,x);
+	EXTRACT_WORDS(hx,low,x);
 	sign=hx&0x80000000; 		/* sign= sign(x) */
 	hx  ^=sign;
 	if(hx>=0x7ff00000) return(x+x); /* cbrt(NaN,INF) is itself */
-	GET_LOW_WORD(low,x);
-	if((hx|low)==0)
-	    return(x);			/* cbrt(0) is itself */
 
     /*
      * Rough cbrt to 5 bits:
@@ -70,13 +66,15 @@ cbrt(double x)
      * subtraction virtually to keep e >= 0 so that ordinary integer
      * division rounds towards minus infinity; this is also efficient.
      */
-	if(hx<0x00100000) { 		/* subnormal number */
+	if(hx<0x00100000) { 		/* zero or subnormal? */
+	    if((hx|low)==0)
+		return(x);		/* cbrt(0) is itself */
 	    SET_HIGH_WORD(t,0x43500000); /* set t= 2**54 */
 	    t*=x;
 	    GET_HIGH_WORD(high,t);
-	    SET_HIGH_WORD(t,sign|((high&0x7fffffff)/3+B2));
+	    INSERT_WORDS(t,sign|((high&0x7fffffff)/3+B2),0);
 	} else
-	    SET_HIGH_WORD(t,sign|(hx/3+B1));
+	    INSERT_WORDS(t,sign|(hx/3+B1),0);
 
     /*
      * New cbrt to 23 bits: