[shufpd version faster than movhlps for size=4, same rest of the time
peter@cordes.ca**20080315060439] {
hunk ./rshift.asm 147
-	pxor	%xmm6, %xmm6		C break the partial-reg dependency.  2.0 cycles/limb instead of 2.5 on conroe and penryn
-	movhlps	%xmm3, %xmm6
+C	pxor	%xmm6, %xmm6		C break the partial-reg dependency.  2.0 cycles/limb instead of 2.5 on conroe and penryn
+C	movhlps	%xmm3, %xmm6
hunk ./rshift.asm 150
-	punpcklqdq %xmm7, %xmm6		C dest=dest[0],src[0]
+C	punpcklqdq %xmm7, %xmm6		C dest=dest[0],src[0]
hunk ./rshift.asm 153
-C	movdqa	%xmm3, %xmm6
+	movdqa	%xmm3, %xmm6
hunk ./rshift.asm 155
-C	shufpd	$1, %xmm7, %xmm6		C %6=limbs(1,2).  take dest[1],src[0], so op=1+0<<1
+	shufpd	$1, %xmm7, %xmm6		C %6=limbs(1,2).  take dest[1],src[0], so op=1+0<<1
}