[order instructions to minimize ROB read port and other stalls.  No measurable speedup, though
peter@cordes.ca**20080314212019] {
hunk ./rshift.asm 109
- 	movd	%ecx, %xmm0		C %0=64-cnt=left count=lc; %1=cnt;
-
-	pxor	%xmm2, %xmm2		C we need this for later, in L(out)
-
hunk ./rshift.asm 110
-	shlq	%cl, %rax		C return value=limb0<<lc.  shift count (64-cnt) in %cl from function setup
+	shlq	%cl, %rax		C return value=limb0<<lc. shift count has to be in %cl.	 C modifies flags, so it has to go before the loop
+ 	movd	%ecx, %xmm0		C %0=64-cnt=left count=lc; %1=cnt;  C this can go anywhere before the loop.
hunk ./rshift.asm 117
-C rsi=last srclimb+1; rdi=last destlimb+1; rdx=-n; cl=%0=64-cnt; %1=cnt;
+C rsi=last srclimb+1; rdi=last destlimb+1; rdx=-n; %0=64-cnt; %1=cnt;
hunk ./rshift.asm 145
-C %3=%6=%7=limb0,limb1 (since it was loaded before rdx was incremented)
-	C pxor	%xmm2, %xmm2  C moved to middle of integer work at beginning of function. not that it saved any cycles.
+C %3=%6=limb0,limb1 (since it was loaded before rdx was incremented).  %7=%3 on for n>2
+	C seems to make no diff where we put pxor, so move it to function start if that helps alignment
+	pxor	%xmm2, %xmm2		C we need this for later, in L(out).
}