[use little-endian notation for register contents
peter@cordes.ca**20080318230601] {
hunk ./rshift.asm 124
-C note: %6 = limb0, limb1  means the _high_ quad of %xmm6 = limb1.  Intel is little-endian, so this is a bit bogus.  It actually matters if you use psrldq or something.
+C notation: %6 = limb1:limb0  means the _high_ quad of %xmm6 = limb1.
hunk ./rshift.asm 128
-	movdqa	(%rsi), %xmm3		C %3 = limb0, limb1
+	movdqa	(%rsi), %xmm3		C %3 = limb1:limb0
hunk ./rshift.asm 140
-C %3=limb0,limb1;
+C %3=limb1:limb0;
hunk ./rshift.asm 148
-	movdqa	(%rsi,%rdx,8), %xmm7	C %7=limb2,limb3.  If src is 16byte aligned, this can't cross a page boundary and segfault.  we might harmlessly read past the end of the array of limbs if its length is odd.
+	movdqa	(%rsi,%rdx,8), %xmm7	C %7=limbs(3:2).  If src is 16byte aligned, this can't cross a page boundary and segfault.  we might harmlessly read past the end of the array of limbs if its length is odd.
+
+C require: %3=limbs(1:0); %7=limbs(3:2)
hunk ./rshift.asm 153
-C	pshufd  $14, %xmm3, %xmm6	C %6=limb1,  14 = (2 + 3<<2)  C latency=2 on pre-penryn
-C	punpcklqdq %xmm7, %xmm6		C dest=dest[0],src[0]
+C	pshufd  $14, %xmm3, %xmm6	C %6=xxx:limb1  14 = (2 + 3<<2)  C latency=2 on pre-penryn
+C	punpcklqdq %xmm7, %xmm6		C dest=src[0]:dest[0]
hunk ./rshift.asm 158
-C require %6=limb1,xxx; %7=limb
-C	punpcklqdq %xmm7, %xmm6		C dest=dest[0],src[0]
+C require %6=xxx:limb1; %7=limbs(3:2)
+C	punpcklqdq %xmm7, %xmm6		C dest=src[0]:dest[0]
hunk ./rshift.asm 163
+C shufpd version ran 2.0 cycles/limb on conroe and penryn, but 3.7 instead of 3.0 for movhlps on K8
hunk ./rshift.asm 165
-C require: %3=limbs(0,1); %6=xxx,limb1; %7=limbs(2,3)  C shufpd version ran 2.0 cycles/limb on conroe and penryn, but 3.7 instead of 3.0 on K8
-	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(2:1).  take dest[1],src[0], so op=1+0<<1
hunk ./rshift.asm 169
-C require: %3=limb0,limb1; %6 = limb1,limb2
+C require: %3=limb1:limb0; %6 = limb2:limb1
hunk ./rshift.asm 171
-	psrlq	%xmm1, %xmm3		C %3=limbs(0,1)>>cnt; %6=limbs(1,2)<<lc
-	por	%xmm6, %xmm3		C %3=result limbs(0,1)
+	psrlq	%xmm1, %xmm3		C %3=limbs(1:0)>>cnt; %6=limbs(2:1)<<lc
+	por	%xmm6, %xmm3		C %3=result limbs(1:0)
hunk ./rshift.asm 175
-C require %7=limbs(2,3) (next iteration's limbs(0,1))
+C require %7=limbs(3:2) (next iteration's limbs(1:0))
hunk ./rshift.asm 182
-C %3=limb0,limb1 (since it was loaded before rdx was incremented).
+C %3=limb1:limb0 (since it was loaded before rdx was incremented).
hunk ./rshift.asm 187
-	punpckhqdq %xmm6, %xmm3		C %6=0,0;  %3=limb1,0
-	psllq	%xmm0, %xmm3		C %3=limb1<<lc,0
-	psrlq	%xmm1, %xmm2		C %2=(limb0,1)>>c
-C require: %3=limb1<<lc, 0; %2=limb0>>c,limb1>>c
-	por	%xmm2, %xmm3		C %3=result limbs 0,1
+	punpckhqdq %xmm6, %xmm3		C %6=0:0;  %3=limb0:1
+	psllq	%xmm0, %xmm3		C %3=0:limb1<<lc
+	psrlq	%xmm1, %xmm2		C %2=(limbs(1:0))>>c
+C require: %3=0:limb1<<lc; %2=limb1>>c:limb0>>c
+	por	%xmm2, %xmm3		C %3=result limbs 1:0
}