[update benchmark comments.
peter@cordes.ca**20080316233505] {
hunk ./rshift.asm 74
-C ++++++++++++++++++++ fast intro (%ecx), movhlps version +++++++++++++++++++
+C ++++++++++++++++++++ fast intro (%ecx), new loop, shufpd (and sometimes movhlps) version +++++++++++++++++++
hunk ./rshift.asm 76
-C size 1		15.938		cycles
-C size 496		3.044		cycles/limb
+C size 1		15.938		cycles ( movhlps)
+C size 4		5.947		cycles/limb (shufpd)
+C size 496		3.791		cycles/limb (3.044 movhlps, still slower than MMX)
hunk ./rshift.asm 82
-C size 4		6.420		cycles  (4.0 with movdqa ; shufpd version uncommented instead.)
-C size 496		2.036-2.068	cycles/limb
+C size 4		4.0		cycles/limb  (6.420 with movhlps)
+C size 496		2.052		cycles/limb (2.036-2.068 movhlps)
hunk ./rshift.asm 87
-C size 496		2.062		cycles/limb
+C size 4		4.013		cycles/limb (shufpd)
+C size 496		2.062		cycles/limb (shufpd/movhlps)
}