115 	mov	 4(HASH_PTR), B
128 	RR F1,A,B,C,D,E,0
129 	RR F1,D,E,A,B,C,2
130 	RR F1,B,C,D,E,A,4
131 	RR F1,E,A,B,C,D,6
132 	RR F1,C,D,E,A,B,8
134 	RR F1,A,B,C,D,E,10
135 	RR F1,D,E,A,B,C,12
136 	RR F1,B,C,D,E,A,14
137 	RR F1,E,A,B,C,D,16
138 	RR F1,C,D,E,A,B,18
140 	RR F2,A,B,C,D,E,20
141 	RR F2,D,E,A,B,C,22
142 	RR F2,B,C,D,E,A,24
143 	RR F2,E,A,B,C,D,26
144 	RR F2,C,D,E,A,B,28
146 	RR F2,A,B,C,D,E,30
147 	RR F2,D,E,A,B,C,32
148 	RR F2,B,C,D,E,A,34
149 	RR F2,E,A,B,C,D,36
150 	RR F2,C,D,E,A,B,38
152 	RR F3,A,B,C,D,E,40
153 	RR F3,D,E,A,B,C,42
154 	RR F3,B,C,D,E,A,44
155 	RR F3,E,A,B,C,D,46
156 	RR F3,C,D,E,A,B,48
158 	RR F3,A,B,C,D,E,50
159 	RR F3,D,E,A,B,C,52
160 	RR F3,B,C,D,E,A,54
161 	RR F3,E,A,B,C,D,56
162 	RR F3,C,D,E,A,B,58
168 	RR F4,A,B,C,D,E,60
169 	RR F4,D,E,A,B,C,62
170 	RR F4,B,C,D,E,A,64
171 	RR F4,E,A,B,C,D,66
172 	RR F4,C,D,E,A,B,68
174 	RR F4,A,B,C,D,E,70
175 	RR F4,D,E,A,B,C,72
176 	RR F4,B,C,D,E,A,74
177 	RR F4,E,A,B,C,D,76
178 	RR F4,C,D,E,A,B,78
181 	UPDATE_HASH  4(HASH_PTR), B
188 	jne	1b
193   .set B, REG_B  define
203 	mov	B, REG_B
209 .macro SWAP_REG_NAMES  a, b  argument
211   .set \a, \b
212   .set \b, _T
215 .macro F1  b, c, d
219 	and	\b, T1
223 .macro F2  b, c, d
227 	xor	\b, T1
230 .macro F3  b, c ,d
233 	mov	\b, T2
234 	or	\b, T1
240 .macro F4  b, c, d
241 	F2 \b, \c, \d
251  *   t1 = F(b, c, d);   e += w(i)
252  *   e += t1;           b <<= 30;   d  += w(i+1);
253  *   t1 = F(a, b, c);
260 .macro RR  F, a, b, c, d, e, round
262 	\F   \b, \c, \d		# t1 = F(b, c, d);
264 	rol	$30, \b
268 	\F   \a, \b, \c
281 	# write:  \a, \b
282 	# rotate: \a<=\d, \b<=\e, \c<=\a, \d<=\b, \e<=\c
349   .if ((i & 3) == 0)
351   .elseif ((i & 3) == 1)
354   .elseif ((i & 3) == 2)
356   .elseif ((i & 3) == 3)
357 	movdqa  W_TMP1, WK(i&~3)
368  * some "heavy-lifting" vectorization for rounds 16-31 due to w[i]->w[i-3]
374   .if ((i & 3) == 0)
378 	psrldq	$4, W_TMP1		# w[i-3]
380   .elseif ((i & 3) == 1)
386   .elseif ((i & 3) == 2)
393   .elseif ((i & 3) == 3)
398 	movdqa	W_TMP1, WK(i&~3)
405  * in SHA-1 specification: w[i] = (w[i-3] ^ w[i-8]  ^ w[i-14] ^ w[i-16]) rol 1
407  * allows more efficient vectorization since w[i]=>w[i-3] dependency is broken
410   .if ((i & 3) == 0)
414   .elseif ((i & 3) == 1)
418   .elseif ((i & 3) == 2)
422   .elseif ((i & 3) == 3)
425 	movdqa	W_TMP1, WK(i&~3)
457 .macro xmm_mov a, b  argument
458 	movdqu	\a,\b
487   .if ((i & 3) == 0)
489   .elseif ((i & 3) == 1)
491   .elseif ((i & 3) == 2)
493   .elseif ((i & 3) == 3)
494 	vmovdqa	W_TMP1, WK(i&~3)
500   .if ((i & 3) == 0)
502 	vpsrldq	$4, W_minus_04, W_TMP1		# w[i-3]
505   .elseif ((i & 3) == 1)
509   .elseif ((i & 3) == 2)
514   .elseif ((i & 3) == 3)
518 	vmovdqu	W_TMP1, WK(i&~3)
524   .if ((i & 3) == 0)
527   .elseif ((i & 3) == 1)
530   .elseif ((i & 3) == 2)
534   .elseif ((i & 3) == 3)
536 	vmovdqu	W_TMP1, WK(i&~3)
545 .macro xmm_mov a, b  argument
546 	vmovdqu	\a,\b