xref: /titanic_41/usr/src/common/crypto/sha1/sha1.c (revision d4660949aa62dd6a963f4913b7120b383cf473c4)
1 /*
2  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
3  * Use is subject to license terms.
4  */
5 
6 #pragma ident	"%Z%%M%	%I%	%E% SMI"
7 
8 /*
9  * The basic framework for this code came from the reference
10  * implementation for MD5.  That implementation is Copyright (C)
11  * 1991-2, RSA Data Security, Inc. Created 1991. All rights reserved.
12  *
13  * License to copy and use this software is granted provided that it
14  * is identified as the "RSA Data Security, Inc. MD5 Message-Digest
15  * Algorithm" in all material mentioning or referencing this software
16  * or this function.
17  *
18  * License is also granted to make and use derivative works provided
19  * that such works are identified as "derived from the RSA Data
20  * Security, Inc. MD5 Message-Digest Algorithm" in all material
21  * mentioning or referencing the derived work.
22  *
23  * RSA Data Security, Inc. makes no representations concerning either
24  * the merchantability of this software or the suitability of this
25  * software for any particular purpose. It is provided "as is"
26  * without express or implied warranty of any kind.
27  *
28  * These notices must be retained in any copies of any part of this
29  * documentation and/or software.
30  *
31  * NOTE: Cleaned-up and optimized, version of SHA1, based on the FIPS 180-1
32  * standard, available at http://www.itl.nist.gov/div897/pubs/fip180-1.htm
33  * Not as fast as one would like -- further optimizations are encouraged
34  * and appreciated.
35  */
36 
37 #include <sys/types.h>
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/sysmacros.h>
41 #include <sys/sha1.h>
42 #include <sys/sha1_consts.h>
43 
44 #ifndef	_KERNEL
45 #include <strings.h>
46 #include <stdlib.h>
47 #include <errno.h>
48 #include <sys/systeminfo.h>
49 #endif  /* !_KERNEL */
50 
51 static void Encode(uint8_t *, const uint32_t *, size_t);
52 
53 #if	defined(__sparc)
54 
55 #define	SHA1_TRANSFORM(ctx, in) \
56 	SHA1Transform((ctx)->state[0], (ctx)->state[1], (ctx)->state[2], \
57 		(ctx)->state[3], (ctx)->state[4], (ctx), (in))
58 
59 static void SHA1Transform(uint32_t, uint32_t, uint32_t, uint32_t, uint32_t,
60     SHA1_CTX *, const uint8_t *);
61 
62 #elif	defined(__amd64)
63 
64 #define	SHA1_TRANSFORM(ctx, in) sha1_block_data_order((ctx), (in), 1)
65 #define	SHA1_TRANSFORM_BLOCKS(ctx, in, num) sha1_block_data_order((ctx), \
66 		(in), (num))
67 
68 void sha1_block_data_order(SHA1_CTX *ctx, const void *inpp, size_t num_blocks);
69 
70 #else
71 
72 #define	SHA1_TRANSFORM(ctx, in) SHA1Transform((ctx), (in))
73 
74 static void SHA1Transform(SHA1_CTX *, const uint8_t *);
75 
76 #endif
77 
78 
79 static uint8_t PADDING[64] = { 0x80, /* all zeros */ };
80 
81 /*
82  * F, G, and H are the basic SHA1 functions.
83  */
84 #define	F(b, c, d)	(((b) & (c)) | ((~b) & (d)))
85 #define	G(b, c, d)	((b) ^ (c) ^ (d))
86 #define	H(b, c, d)	(((b) & (c)) | (((b)|(c)) & (d)))
87 
88 /*
89  * ROTATE_LEFT rotates x left n bits.
90  */
91 
92 #if	defined(__GNUC__) && defined(_LP64)
93 static __inline__ uint64_t
94 ROTATE_LEFT(uint64_t value, uint32_t n)
95 {
96 	uint32_t t32;
97 
98 	t32 = (uint32_t)value;
99 	return ((t32 << n) | (t32 >> (32 - n)));
100 }
101 
102 #else
103 
104 #define	ROTATE_LEFT(x, n)	\
105 	(((x) << (n)) | ((x) >> ((sizeof (x) * NBBY)-(n))))
106 
107 #endif
108 
109 #if	defined(__GNUC__) && (defined(__i386) || defined(__amd64))
110 
111 #define	HAVE_BSWAP
112 
113 extern __inline__ uint32_t bswap(uint32_t value)
114 {
115 	__asm__("bswap %0" : "+r" (value));
116 	return (value);
117 }
118 
119 #endif
120 
121 /*
122  * SHA1Init()
123  *
124  * purpose: initializes the sha1 context and begins and sha1 digest operation
125  *   input: SHA1_CTX *	: the context to initializes.
126  *  output: void
127  */
128 
129 void
130 SHA1Init(SHA1_CTX *ctx)
131 {
132 	ctx->count[0] = ctx->count[1] = 0;
133 
134 	/*
135 	 * load magic initialization constants. Tell lint
136 	 * that these constants are unsigned by using U.
137 	 */
138 
139 	ctx->state[0] = 0x67452301U;
140 	ctx->state[1] = 0xefcdab89U;
141 	ctx->state[2] = 0x98badcfeU;
142 	ctx->state[3] = 0x10325476U;
143 	ctx->state[4] = 0xc3d2e1f0U;
144 }
145 
146 #ifdef VIS_SHA1
147 #ifdef _KERNEL
148 
149 #include <sys/regset.h>
150 #include <sys/vis.h>
151 #include <sys/fpu/fpusystm.h>
152 
153 /* the alignment for block stores to save fp registers */
154 #define	VIS_ALIGN	(64)
155 
156 extern int sha1_savefp(kfpu_t *, int);
157 extern void sha1_restorefp(kfpu_t *);
158 
159 uint32_t	vis_sha1_svfp_threshold = 128;
160 
161 #endif /* _KERNEL */
162 
163 /*
164  * VIS SHA-1 consts.
165  */
166 static uint64_t VIS[] = {
167 	0x8000000080000000ULL,
168 	0x0002000200020002ULL,
169 	0x5a8279996ed9eba1ULL,
170 	0x8f1bbcdcca62c1d6ULL,
171 	0x012389ab456789abULL};
172 
173 extern void SHA1TransformVIS(uint64_t *, uint32_t *, uint32_t *, uint64_t *);
174 
175 
176 /*
177  * SHA1Update()
178  *
179  * purpose: continues an sha1 digest operation, using the message block
180  *          to update the context.
181  *   input: SHA1_CTX *	: the context to update
182  *          void *	: the message block
183  *          size_t    : the length of the message block in bytes
184  *  output: void
185  */
186 
187 void
188 SHA1Update(SHA1_CTX *ctx, const void *inptr, size_t input_len)
189 {
190 	uint32_t i, buf_index, buf_len;
191 	uint64_t X0[40], input64[8];
192 	const uint8_t *input = inptr;
193 #ifdef _KERNEL
194 	int usevis = 0;
195 #else
196 	int usevis = 1;
197 #endif /* _KERNEL */
198 
199 	/* check for noop */
200 	if (input_len == 0)
201 		return;
202 
203 	/* compute number of bytes mod 64 */
204 	buf_index = (ctx->count[1] >> 3) & 0x3F;
205 
206 	/* update number of bits */
207 	if ((ctx->count[1] += (input_len << 3)) < (input_len << 3))
208 		ctx->count[0]++;
209 
210 	ctx->count[0] += (input_len >> 29);
211 
212 	buf_len = 64 - buf_index;
213 
214 	/* transform as many times as possible */
215 	i = 0;
216 	if (input_len >= buf_len) {
217 #ifdef _KERNEL
218 		kfpu_t *fpu;
219 		if (fpu_exists) {
220 			uint8_t fpua[sizeof (kfpu_t) + GSR_SIZE + VIS_ALIGN];
221 			uint32_t len = (input_len + buf_index) & ~0x3f;
222 			int svfp_ok;
223 
224 			fpu = (kfpu_t *)P2ROUNDUP((uintptr_t)fpua, 64);
225 			svfp_ok = ((len >= vis_sha1_svfp_threshold) ? 1 : 0);
226 			usevis = fpu_exists && sha1_savefp(fpu, svfp_ok);
227 		} else {
228 			usevis = 0;
229 		}
230 #endif /* _KERNEL */
231 
232 		/*
233 		 * general optimization:
234 		 *
235 		 * only do initial bcopy() and SHA1Transform() if
236 		 * buf_index != 0.  if buf_index == 0, we're just
237 		 * wasting our time doing the bcopy() since there
238 		 * wasn't any data left over from a previous call to
239 		 * SHA1Update().
240 		 */
241 
242 		if (buf_index) {
243 			bcopy(input, &ctx->buf_un.buf8[buf_index], buf_len);
244 			if (usevis) {
245 				SHA1TransformVIS(X0,
246 				    ctx->buf_un.buf32,
247 				    &ctx->state[0], VIS);
248 			} else {
249 				SHA1_TRANSFORM(ctx, ctx->buf_un.buf8);
250 			}
251 			i = buf_len;
252 		}
253 
254 		/*
255 		 * VIS SHA-1: uses the VIS 1.0 instructions to accelerate
256 		 * SHA-1 processing. This is achieved by "offloading" the
257 		 * computation of the message schedule (MS) to the VIS units.
258 		 * This allows the VIS computation of the message schedule
259 		 * to be performed in parallel with the standard integer
260 		 * processing of the remainder of the SHA-1 computation.
261 		 * performance by up to around 1.37X, compared to an optimized
262 		 * integer-only implementation.
263 		 *
264 		 * The VIS implementation of SHA1Transform has a different API
265 		 * to the standard integer version:
266 		 *
267 		 * void SHA1TransformVIS(
268 		 *	 uint64_t *, // Pointer to MS for ith block
269 		 *	 uint32_t *, // Pointer to ith block of message data
270 		 *	 uint32_t *, // Pointer to SHA state i.e ctx->state
271 		 *	 uint64_t *, // Pointer to various VIS constants
272 		 * )
273 		 *
274 		 * Note: the message data must by 4-byte aligned.
275 		 *
276 		 * Function requires VIS 1.0 support.
277 		 *
278 		 * Handling is provided to deal with arbitrary byte alingment
279 		 * of the input data but the performance gains are reduced
280 		 * for alignments other than 4-bytes.
281 		 */
282 		if (usevis) {
283 			if (!IS_P2ALIGNED(&input[i], sizeof (uint32_t))) {
284 				/*
285 				 * Main processing loop - input misaligned
286 				 */
287 				for (; i + 63 < input_len; i += 64) {
288 					bcopy(&input[i], input64, 64);
289 					SHA1TransformVIS(X0,
290 					    (uint32_t *)input64,
291 					    &ctx->state[0], VIS);
292 				}
293 			} else {
294 				/*
295 				 * Main processing loop - input 8-byte aligned
296 				 */
297 				for (; i + 63 < input_len; i += 64) {
298 					SHA1TransformVIS(X0,
299 					/* LINTED E_BAD_PTR_CAST_ALIGN */
300 					    (uint32_t *)&input[i],
301 					    &ctx->state[0], VIS);
302 				}
303 
304 			}
305 #ifdef _KERNEL
306 			sha1_restorefp(fpu);
307 #endif /* _KERNEL */
308 		} else {
309 			for (; i + 63 < input_len; i += 64) {
310 				SHA1_TRANSFORM(ctx, &input[i]);
311 			}
312 		}
313 
314 		/*
315 		 * general optimization:
316 		 *
317 		 * if i and input_len are the same, return now instead
318 		 * of calling bcopy(), since the bcopy() in this case
319 		 * will be an expensive nop.
320 		 */
321 
322 		if (input_len == i)
323 			return;
324 
325 		buf_index = 0;
326 	}
327 
328 	/* buffer remaining input */
329 	bcopy(&input[i], &ctx->buf_un.buf8[buf_index], input_len - i);
330 }
331 
332 #else /* VIS_SHA1 */
333 
334 void
335 SHA1Update(SHA1_CTX *ctx, const void *inptr, size_t input_len)
336 {
337 	uint32_t i, buf_index, buf_len;
338 	const uint8_t *input = inptr;
339 #if defined(__amd64)
340 	uint32_t	block_count;
341 #endif	/* __amd64 */
342 
343 	/* check for noop */
344 	if (input_len == 0)
345 		return;
346 
347 	/* compute number of bytes mod 64 */
348 	buf_index = (ctx->count[1] >> 3) & 0x3F;
349 
350 	/* update number of bits */
351 	if ((ctx->count[1] += (input_len << 3)) < (input_len << 3))
352 		ctx->count[0]++;
353 
354 	ctx->count[0] += (input_len >> 29);
355 
356 	buf_len = 64 - buf_index;
357 
358 	/* transform as many times as possible */
359 	i = 0;
360 	if (input_len >= buf_len) {
361 
362 		/*
363 		 * general optimization:
364 		 *
365 		 * only do initial bcopy() and SHA1Transform() if
366 		 * buf_index != 0.  if buf_index == 0, we're just
367 		 * wasting our time doing the bcopy() since there
368 		 * wasn't any data left over from a previous call to
369 		 * SHA1Update().
370 		 */
371 
372 		if (buf_index) {
373 			bcopy(input, &ctx->buf_un.buf8[buf_index], buf_len);
374 			SHA1_TRANSFORM(ctx, ctx->buf_un.buf8);
375 			i = buf_len;
376 		}
377 
378 #if !defined(__amd64)
379 		for (; i + 63 < input_len; i += 64)
380 			SHA1_TRANSFORM(ctx, &input[i]);
381 #else
382 		block_count = (input_len - i) >> 6;
383 		if (block_count > 0) {
384 			SHA1_TRANSFORM_BLOCKS(ctx, &input[i], block_count);
385 			i += block_count << 6;
386 		}
387 #endif	/* !__amd64 */
388 
389 		/*
390 		 * general optimization:
391 		 *
392 		 * if i and input_len are the same, return now instead
393 		 * of calling bcopy(), since the bcopy() in this case
394 		 * will be an expensive nop.
395 		 */
396 
397 		if (input_len == i)
398 			return;
399 
400 		buf_index = 0;
401 	}
402 
403 	/* buffer remaining input */
404 	bcopy(&input[i], &ctx->buf_un.buf8[buf_index], input_len - i);
405 }
406 
407 #endif /* VIS_SHA1 */
408 
409 /*
410  * SHA1Final()
411  *
412  * purpose: ends an sha1 digest operation, finalizing the message digest and
413  *          zeroing the context.
414  *   input: uchar_t *	: A buffer to store the digest.
415  *			: The function actually uses void* because many
416  *			: callers pass things other than uchar_t here.
417  *          SHA1_CTX *  : the context to finalize, save, and zero
418  *  output: void
419  */
420 
421 void
422 SHA1Final(void *digest, SHA1_CTX *ctx)
423 {
424 	uint8_t		bitcount_be[sizeof (ctx->count)];
425 	uint32_t	index = (ctx->count[1] >> 3) & 0x3f;
426 
427 	/* store bit count, big endian */
428 	Encode(bitcount_be, ctx->count, sizeof (bitcount_be));
429 
430 	/* pad out to 56 mod 64 */
431 	SHA1Update(ctx, PADDING, ((index < 56) ? 56 : 120) - index);
432 
433 	/* append length (before padding) */
434 	SHA1Update(ctx, bitcount_be, sizeof (bitcount_be));
435 
436 	/* store state in digest */
437 	Encode(digest, ctx->state, sizeof (ctx->state));
438 
439 	/* zeroize sensitive information */
440 	bzero(ctx, sizeof (*ctx));
441 }
442 
443 
444 #if !defined(__amd64)
445 
446 typedef uint32_t sha1word;
447 
448 /*
449  * sparc optimization:
450  *
451  * on the sparc, we can load big endian 32-bit data easily.  note that
452  * special care must be taken to ensure the address is 32-bit aligned.
453  * in the interest of speed, we don't check to make sure, since
454  * careful programming can guarantee this for us.
455  */
456 
457 #if	defined(_BIG_ENDIAN)
458 
459 #define	LOAD_BIG_32(addr)	(*(uint32_t *)(addr))
460 
461 #else	/* !defined(_BIG_ENDIAN) */
462 
463 #if	defined(HAVE_BSWAP)
464 
465 #define	LOAD_BIG_32(addr) bswap(*((uint32_t *)(addr)))
466 
467 #else	/* !defined(HAVE_BSWAP) */
468 
469 /* little endian -- will work on big endian, but slowly */
470 #define	LOAD_BIG_32(addr)	\
471 	(((addr)[0] << 24) | ((addr)[1] << 16) | ((addr)[2] << 8) | (addr)[3])
472 
473 #endif	/* !defined(HAVE_BSWAP) */
474 
475 #endif	/* !defined(_BIG_ENDIAN) */
476 
477 /*
478  * SHA1Transform()
479  */
480 #if	defined(W_ARRAY)
481 #define	W(n) w[n]
482 #else	/* !defined(W_ARRAY) */
483 #define	W(n) w_ ## n
484 #endif	/* !defined(W_ARRAY) */
485 
486 
487 #if	defined(__sparc)
488 
489 /*
490  * sparc register window optimization:
491  *
492  * `a', `b', `c', `d', and `e' are passed into SHA1Transform
493  * explicitly since it increases the number of registers available to
494  * the compiler.  under this scheme, these variables can be held in
495  * %i0 - %i4, which leaves more local and out registers available.
496  *
497  * purpose: sha1 transformation -- updates the digest based on `block'
498  *   input: uint32_t	: bytes  1 -  4 of the digest
499  *          uint32_t	: bytes  5 -  8 of the digest
500  *          uint32_t	: bytes  9 - 12 of the digest
501  *          uint32_t	: bytes 12 - 16 of the digest
502  *          uint32_t	: bytes 16 - 20 of the digest
503  *          SHA1_CTX *	: the context to update
504  *          uint8_t [64]: the block to use to update the digest
505  *  output: void
506  */
507 
508 void
509 SHA1Transform(uint32_t a, uint32_t b, uint32_t c, uint32_t d, uint32_t e,
510     SHA1_CTX *ctx, const uint8_t blk[64])
511 {
512 	/*
513 	 * sparc optimization:
514 	 *
515 	 * while it is somewhat counter-intuitive, on sparc, it is
516 	 * more efficient to place all the constants used in this
517 	 * function in an array and load the values out of the array
518 	 * than to manually load the constants.  this is because
519 	 * setting a register to a 32-bit value takes two ops in most
520 	 * cases: a `sethi' and an `or', but loading a 32-bit value
521 	 * from memory only takes one `ld' (or `lduw' on v9).  while
522 	 * this increases memory usage, the compiler can find enough
523 	 * other things to do while waiting to keep the pipeline does
524 	 * not stall.  additionally, it is likely that many of these
525 	 * constants are cached so that later accesses do not even go
526 	 * out to the bus.
527 	 *
528 	 * this array is declared `static' to keep the compiler from
529 	 * having to bcopy() this array onto the stack frame of
530 	 * SHA1Transform() each time it is called -- which is
531 	 * unacceptably expensive.
532 	 *
533 	 * the `const' is to ensure that callers are good citizens and
534 	 * do not try to munge the array.  since these routines are
535 	 * going to be called from inside multithreaded kernelland,
536 	 * this is a good safety check. -- `sha1_consts' will end up in
537 	 * .rodata.
538 	 *
539 	 * unfortunately, loading from an array in this manner hurts
540 	 * performance under intel.  so, there is a macro,
541 	 * SHA1_CONST(), used in SHA1Transform(), that either expands to
542 	 * a reference to this array, or to the actual constant,
543 	 * depending on what platform this code is compiled for.
544 	 */
545 
546 	static const uint32_t sha1_consts[] = {
547 		SHA1_CONST_0,	SHA1_CONST_1,	SHA1_CONST_2,	SHA1_CONST_3,
548 	};
549 
550 	/*
551 	 * general optimization:
552 	 *
553 	 * use individual integers instead of using an array.  this is a
554 	 * win, although the amount it wins by seems to vary quite a bit.
555 	 */
556 
557 	uint32_t	w_0, w_1, w_2,  w_3,  w_4,  w_5,  w_6,  w_7;
558 	uint32_t	w_8, w_9, w_10, w_11, w_12, w_13, w_14, w_15;
559 
560 	/*
561 	 * sparc optimization:
562 	 *
563 	 * if `block' is already aligned on a 4-byte boundary, use
564 	 * LOAD_BIG_32() directly.  otherwise, bcopy() into a
565 	 * buffer that *is* aligned on a 4-byte boundary and then do
566 	 * the LOAD_BIG_32() on that buffer.  benchmarks have shown
567 	 * that using the bcopy() is better than loading the bytes
568 	 * individually and doing the endian-swap by hand.
569 	 *
570 	 * even though it's quite tempting to assign to do:
571 	 *
572 	 * blk = bcopy(ctx->buf_un.buf32, blk, sizeof (ctx->buf_un.buf32));
573 	 *
574 	 * and only have one set of LOAD_BIG_32()'s, the compiler
575 	 * *does not* like that, so please resist the urge.
576 	 */
577 
578 	if ((uintptr_t)blk & 0x3) {		/* not 4-byte aligned? */
579 		bcopy(blk, ctx->buf_un.buf32,  sizeof (ctx->buf_un.buf32));
580 		w_15 = LOAD_BIG_32(ctx->buf_un.buf32 + 15);
581 		w_14 = LOAD_BIG_32(ctx->buf_un.buf32 + 14);
582 		w_13 = LOAD_BIG_32(ctx->buf_un.buf32 + 13);
583 		w_12 = LOAD_BIG_32(ctx->buf_un.buf32 + 12);
584 		w_11 = LOAD_BIG_32(ctx->buf_un.buf32 + 11);
585 		w_10 = LOAD_BIG_32(ctx->buf_un.buf32 + 10);
586 		w_9  = LOAD_BIG_32(ctx->buf_un.buf32 +  9);
587 		w_8  = LOAD_BIG_32(ctx->buf_un.buf32 +  8);
588 		w_7  = LOAD_BIG_32(ctx->buf_un.buf32 +  7);
589 		w_6  = LOAD_BIG_32(ctx->buf_un.buf32 +  6);
590 		w_5  = LOAD_BIG_32(ctx->buf_un.buf32 +  5);
591 		w_4  = LOAD_BIG_32(ctx->buf_un.buf32 +  4);
592 		w_3  = LOAD_BIG_32(ctx->buf_un.buf32 +  3);
593 		w_2  = LOAD_BIG_32(ctx->buf_un.buf32 +  2);
594 		w_1  = LOAD_BIG_32(ctx->buf_un.buf32 +  1);
595 		w_0  = LOAD_BIG_32(ctx->buf_un.buf32 +  0);
596 	} else {
597 		/*LINTED*/
598 		w_15 = LOAD_BIG_32(blk + 60);
599 		/*LINTED*/
600 		w_14 = LOAD_BIG_32(blk + 56);
601 		/*LINTED*/
602 		w_13 = LOAD_BIG_32(blk + 52);
603 		/*LINTED*/
604 		w_12 = LOAD_BIG_32(blk + 48);
605 		/*LINTED*/
606 		w_11 = LOAD_BIG_32(blk + 44);
607 		/*LINTED*/
608 		w_10 = LOAD_BIG_32(blk + 40);
609 		/*LINTED*/
610 		w_9  = LOAD_BIG_32(blk + 36);
611 		/*LINTED*/
612 		w_8  = LOAD_BIG_32(blk + 32);
613 		/*LINTED*/
614 		w_7  = LOAD_BIG_32(blk + 28);
615 		/*LINTED*/
616 		w_6  = LOAD_BIG_32(blk + 24);
617 		/*LINTED*/
618 		w_5  = LOAD_BIG_32(blk + 20);
619 		/*LINTED*/
620 		w_4  = LOAD_BIG_32(blk + 16);
621 		/*LINTED*/
622 		w_3  = LOAD_BIG_32(blk + 12);
623 		/*LINTED*/
624 		w_2  = LOAD_BIG_32(blk +  8);
625 		/*LINTED*/
626 		w_1  = LOAD_BIG_32(blk +  4);
627 		/*LINTED*/
628 		w_0  = LOAD_BIG_32(blk +  0);
629 	}
630 #else	/* !defined(__sparc) */
631 
632 void
633 SHA1Transform(SHA1_CTX *ctx, const uint8_t blk[64])
634 {
635 	sha1word a = ctx->state[0];
636 	sha1word b = ctx->state[1];
637 	sha1word c = ctx->state[2];
638 	sha1word d = ctx->state[3];
639 	sha1word e = ctx->state[4];
640 
641 #if	defined(W_ARRAY)
642 	sha1word	w[16];
643 #else	/* !defined(W_ARRAY) */
644 	sha1word	w_0, w_1, w_2,  w_3,  w_4,  w_5,  w_6,  w_7;
645 	sha1word	w_8, w_9, w_10, w_11, w_12, w_13, w_14, w_15;
646 #endif	/* !defined(W_ARRAY) */
647 
648 	W(0)  = LOAD_BIG_32(blk +  0);
649 	W(1)  = LOAD_BIG_32(blk +  4);
650 	W(2)  = LOAD_BIG_32(blk +  8);
651 	W(3)  = LOAD_BIG_32(blk + 12);
652 	W(4)  = LOAD_BIG_32(blk + 16);
653 	W(5)  = LOAD_BIG_32(blk + 20);
654 	W(6)  = LOAD_BIG_32(blk + 24);
655 	W(7)  = LOAD_BIG_32(blk + 28);
656 	W(8)  = LOAD_BIG_32(blk + 32);
657 	W(9)  = LOAD_BIG_32(blk + 36);
658 	W(10) = LOAD_BIG_32(blk + 40);
659 	W(11) = LOAD_BIG_32(blk + 44);
660 	W(12) = LOAD_BIG_32(blk + 48);
661 	W(13) = LOAD_BIG_32(blk + 52);
662 	W(14) = LOAD_BIG_32(blk + 56);
663 	W(15) = LOAD_BIG_32(blk + 60);
664 
665 #endif	/* !defined(__sparc) */
666 
667 	/*
668 	 * general optimization:
669 	 *
670 	 * even though this approach is described in the standard as
671 	 * being slower algorithmically, it is 30-40% faster than the
672 	 * "faster" version under SPARC, because this version has more
673 	 * of the constraints specified at compile-time and uses fewer
674 	 * variables (and therefore has better register utilization)
675 	 * than its "speedier" brother.  (i've tried both, trust me)
676 	 *
677 	 * for either method given in the spec, there is an "assignment"
678 	 * phase where the following takes place:
679 	 *
680 	 *	tmp = (main_computation);
681 	 *	e = d; d = c; c = rotate_left(b, 30); b = a; a = tmp;
682 	 *
683 	 * we can make the algorithm go faster by not doing this work,
684 	 * but just pretending that `d' is now `e', etc. this works
685 	 * really well and obviates the need for a temporary variable.
686 	 * however, we still explicitly perform the rotate action,
687 	 * since it is cheaper on SPARC to do it once than to have to
688 	 * do it over and over again.
689 	 */
690 
691 	/* round 1 */
692 	e = ROTATE_LEFT(a, 5) + F(b, c, d) + e + W(0) + SHA1_CONST(0); /* 0 */
693 	b = ROTATE_LEFT(b, 30);
694 
695 	d = ROTATE_LEFT(e, 5) + F(a, b, c) + d + W(1) + SHA1_CONST(0); /* 1 */
696 	a = ROTATE_LEFT(a, 30);
697 
698 	c = ROTATE_LEFT(d, 5) + F(e, a, b) + c + W(2) + SHA1_CONST(0); /* 2 */
699 	e = ROTATE_LEFT(e, 30);
700 
701 	b = ROTATE_LEFT(c, 5) + F(d, e, a) + b + W(3) + SHA1_CONST(0); /* 3 */
702 	d = ROTATE_LEFT(d, 30);
703 
704 	a = ROTATE_LEFT(b, 5) + F(c, d, e) + a + W(4) + SHA1_CONST(0); /* 4 */
705 	c = ROTATE_LEFT(c, 30);
706 
707 	e = ROTATE_LEFT(a, 5) + F(b, c, d) + e + W(5) + SHA1_CONST(0); /* 5 */
708 	b = ROTATE_LEFT(b, 30);
709 
710 	d = ROTATE_LEFT(e, 5) + F(a, b, c) + d + W(6) + SHA1_CONST(0); /* 6 */
711 	a = ROTATE_LEFT(a, 30);
712 
713 	c = ROTATE_LEFT(d, 5) + F(e, a, b) + c + W(7) + SHA1_CONST(0); /* 7 */
714 	e = ROTATE_LEFT(e, 30);
715 
716 	b = ROTATE_LEFT(c, 5) + F(d, e, a) + b + W(8) + SHA1_CONST(0); /* 8 */
717 	d = ROTATE_LEFT(d, 30);
718 
719 	a = ROTATE_LEFT(b, 5) + F(c, d, e) + a + W(9) + SHA1_CONST(0); /* 9 */
720 	c = ROTATE_LEFT(c, 30);
721 
722 	e = ROTATE_LEFT(a, 5) + F(b, c, d) + e + W(10) + SHA1_CONST(0); /* 10 */
723 	b = ROTATE_LEFT(b, 30);
724 
725 	d = ROTATE_LEFT(e, 5) + F(a, b, c) + d + W(11) + SHA1_CONST(0); /* 11 */
726 	a = ROTATE_LEFT(a, 30);
727 
728 	c = ROTATE_LEFT(d, 5) + F(e, a, b) + c + W(12) + SHA1_CONST(0); /* 12 */
729 	e = ROTATE_LEFT(e, 30);
730 
731 	b = ROTATE_LEFT(c, 5) + F(d, e, a) + b + W(13) + SHA1_CONST(0); /* 13 */
732 	d = ROTATE_LEFT(d, 30);
733 
734 	a = ROTATE_LEFT(b, 5) + F(c, d, e) + a + W(14) + SHA1_CONST(0); /* 14 */
735 	c = ROTATE_LEFT(c, 30);
736 
737 	e = ROTATE_LEFT(a, 5) + F(b, c, d) + e + W(15) + SHA1_CONST(0); /* 15 */
738 	b = ROTATE_LEFT(b, 30);
739 
740 	W(0) = ROTATE_LEFT((W(13) ^ W(8) ^ W(2) ^ W(0)), 1);		/* 16 */
741 	d = ROTATE_LEFT(e, 5) + F(a, b, c) + d + W(0) + SHA1_CONST(0);
742 	a = ROTATE_LEFT(a, 30);
743 
744 	W(1) = ROTATE_LEFT((W(14) ^ W(9) ^ W(3) ^ W(1)), 1);		/* 17 */
745 	c = ROTATE_LEFT(d, 5) + F(e, a, b) + c + W(1) + SHA1_CONST(0);
746 	e = ROTATE_LEFT(e, 30);
747 
748 	W(2) = ROTATE_LEFT((W(15) ^ W(10) ^ W(4) ^ W(2)), 1);	/* 18 */
749 	b = ROTATE_LEFT(c, 5) + F(d, e, a) + b + W(2) + SHA1_CONST(0);
750 	d = ROTATE_LEFT(d, 30);
751 
752 	W(3) = ROTATE_LEFT((W(0) ^ W(11) ^ W(5) ^ W(3)), 1);		/* 19 */
753 	a = ROTATE_LEFT(b, 5) + F(c, d, e) + a + W(3) + SHA1_CONST(0);
754 	c = ROTATE_LEFT(c, 30);
755 
756 	/* round 2 */
757 	W(4) = ROTATE_LEFT((W(1) ^ W(12) ^ W(6) ^ W(4)), 1);		/* 20 */
758 	e = ROTATE_LEFT(a, 5) + G(b, c, d) + e + W(4) + SHA1_CONST(1);
759 	b = ROTATE_LEFT(b, 30);
760 
761 	W(5) = ROTATE_LEFT((W(2) ^ W(13) ^ W(7) ^ W(5)), 1);		/* 21 */
762 	d = ROTATE_LEFT(e, 5) + G(a, b, c) + d + W(5) + SHA1_CONST(1);
763 	a = ROTATE_LEFT(a, 30);
764 
765 	W(6) = ROTATE_LEFT((W(3) ^ W(14) ^ W(8) ^ W(6)), 1);		/* 22 */
766 	c = ROTATE_LEFT(d, 5) + G(e, a, b) + c + W(6) + SHA1_CONST(1);
767 	e = ROTATE_LEFT(e, 30);
768 
769 	W(7) = ROTATE_LEFT((W(4) ^ W(15) ^ W(9) ^ W(7)), 1);		/* 23 */
770 	b = ROTATE_LEFT(c, 5) + G(d, e, a) + b + W(7) + SHA1_CONST(1);
771 	d = ROTATE_LEFT(d, 30);
772 
773 	W(8) = ROTATE_LEFT((W(5) ^ W(0) ^ W(10) ^ W(8)), 1);		/* 24 */
774 	a = ROTATE_LEFT(b, 5) + G(c, d, e) + a + W(8) + SHA1_CONST(1);
775 	c = ROTATE_LEFT(c, 30);
776 
777 	W(9) = ROTATE_LEFT((W(6) ^ W(1) ^ W(11) ^ W(9)), 1);		/* 25 */
778 	e = ROTATE_LEFT(a, 5) + G(b, c, d) + e + W(9) + SHA1_CONST(1);
779 	b = ROTATE_LEFT(b, 30);
780 
781 	W(10) = ROTATE_LEFT((W(7) ^ W(2) ^ W(12) ^ W(10)), 1);	/* 26 */
782 	d = ROTATE_LEFT(e, 5) + G(a, b, c) + d + W(10) + SHA1_CONST(1);
783 	a = ROTATE_LEFT(a, 30);
784 
785 	W(11) = ROTATE_LEFT((W(8) ^ W(3) ^ W(13) ^ W(11)), 1);	/* 27 */
786 	c = ROTATE_LEFT(d, 5) + G(e, a, b) + c + W(11) + SHA1_CONST(1);
787 	e = ROTATE_LEFT(e, 30);
788 
789 	W(12) = ROTATE_LEFT((W(9) ^ W(4) ^ W(14) ^ W(12)), 1);	/* 28 */
790 	b = ROTATE_LEFT(c, 5) + G(d, e, a) + b + W(12) + SHA1_CONST(1);
791 	d = ROTATE_LEFT(d, 30);
792 
793 	W(13) = ROTATE_LEFT((W(10) ^ W(5) ^ W(15) ^ W(13)), 1);	/* 29 */
794 	a = ROTATE_LEFT(b, 5) + G(c, d, e) + a + W(13) + SHA1_CONST(1);
795 	c = ROTATE_LEFT(c, 30);
796 
797 	W(14) = ROTATE_LEFT((W(11) ^ W(6) ^ W(0) ^ W(14)), 1);	/* 30 */
798 	e = ROTATE_LEFT(a, 5) + G(b, c, d) + e + W(14) + SHA1_CONST(1);
799 	b = ROTATE_LEFT(b, 30);
800 
801 	W(15) = ROTATE_LEFT((W(12) ^ W(7) ^ W(1) ^ W(15)), 1);	/* 31 */
802 	d = ROTATE_LEFT(e, 5) + G(a, b, c) + d + W(15) + SHA1_CONST(1);
803 	a = ROTATE_LEFT(a, 30);
804 
805 	W(0) = ROTATE_LEFT((W(13) ^ W(8) ^ W(2) ^ W(0)), 1);		/* 32 */
806 	c = ROTATE_LEFT(d, 5) + G(e, a, b) + c + W(0) + SHA1_CONST(1);
807 	e = ROTATE_LEFT(e, 30);
808 
809 	W(1) = ROTATE_LEFT((W(14) ^ W(9) ^ W(3) ^ W(1)), 1);		/* 33 */
810 	b = ROTATE_LEFT(c, 5) + G(d, e, a) + b + W(1) + SHA1_CONST(1);
811 	d = ROTATE_LEFT(d, 30);
812 
813 	W(2) = ROTATE_LEFT((W(15) ^ W(10) ^ W(4) ^ W(2)), 1);	/* 34 */
814 	a = ROTATE_LEFT(b, 5) + G(c, d, e) + a + W(2) + SHA1_CONST(1);
815 	c = ROTATE_LEFT(c, 30);
816 
817 	W(3) = ROTATE_LEFT((W(0) ^ W(11) ^ W(5) ^ W(3)), 1);		/* 35 */
818 	e = ROTATE_LEFT(a, 5) + G(b, c, d) + e + W(3) + SHA1_CONST(1);
819 	b = ROTATE_LEFT(b, 30);
820 
821 	W(4) = ROTATE_LEFT((W(1) ^ W(12) ^ W(6) ^ W(4)), 1);		/* 36 */
822 	d = ROTATE_LEFT(e, 5) + G(a, b, c) + d + W(4) + SHA1_CONST(1);
823 	a = ROTATE_LEFT(a, 30);
824 
825 	W(5) = ROTATE_LEFT((W(2) ^ W(13) ^ W(7) ^ W(5)), 1);		/* 37 */
826 	c = ROTATE_LEFT(d, 5) + G(e, a, b) + c + W(5) + SHA1_CONST(1);
827 	e = ROTATE_LEFT(e, 30);
828 
829 	W(6) = ROTATE_LEFT((W(3) ^ W(14) ^ W(8) ^ W(6)), 1);		/* 38 */
830 	b = ROTATE_LEFT(c, 5) + G(d, e, a) + b + W(6) + SHA1_CONST(1);
831 	d = ROTATE_LEFT(d, 30);
832 
833 	W(7) = ROTATE_LEFT((W(4) ^ W(15) ^ W(9) ^ W(7)), 1);		/* 39 */
834 	a = ROTATE_LEFT(b, 5) + G(c, d, e) + a + W(7) + SHA1_CONST(1);
835 	c = ROTATE_LEFT(c, 30);
836 
837 	/* round 3 */
838 	W(8) = ROTATE_LEFT((W(5) ^ W(0) ^ W(10) ^ W(8)), 1);		/* 40 */
839 	e = ROTATE_LEFT(a, 5) + H(b, c, d) + e + W(8) + SHA1_CONST(2);
840 	b = ROTATE_LEFT(b, 30);
841 
842 	W(9) = ROTATE_LEFT((W(6) ^ W(1) ^ W(11) ^ W(9)), 1);		/* 41 */
843 	d = ROTATE_LEFT(e, 5) + H(a, b, c) + d + W(9) + SHA1_CONST(2);
844 	a = ROTATE_LEFT(a, 30);
845 
846 	W(10) = ROTATE_LEFT((W(7) ^ W(2) ^ W(12) ^ W(10)), 1);	/* 42 */
847 	c = ROTATE_LEFT(d, 5) + H(e, a, b) + c + W(10) + SHA1_CONST(2);
848 	e = ROTATE_LEFT(e, 30);
849 
850 	W(11) = ROTATE_LEFT((W(8) ^ W(3) ^ W(13) ^ W(11)), 1);	/* 43 */
851 	b = ROTATE_LEFT(c, 5) + H(d, e, a) + b + W(11) + SHA1_CONST(2);
852 	d = ROTATE_LEFT(d, 30);
853 
854 	W(12) = ROTATE_LEFT((W(9) ^ W(4) ^ W(14) ^ W(12)), 1);	/* 44 */
855 	a = ROTATE_LEFT(b, 5) + H(c, d, e) + a + W(12) + SHA1_CONST(2);
856 	c = ROTATE_LEFT(c, 30);
857 
858 	W(13) = ROTATE_LEFT((W(10) ^ W(5) ^ W(15) ^ W(13)), 1);	/* 45 */
859 	e = ROTATE_LEFT(a, 5) + H(b, c, d) + e + W(13) + SHA1_CONST(2);
860 	b = ROTATE_LEFT(b, 30);
861 
862 	W(14) = ROTATE_LEFT((W(11) ^ W(6) ^ W(0) ^ W(14)), 1);	/* 46 */
863 	d = ROTATE_LEFT(e, 5) + H(a, b, c) + d + W(14) + SHA1_CONST(2);
864 	a = ROTATE_LEFT(a, 30);
865 
866 	W(15) = ROTATE_LEFT((W(12) ^ W(7) ^ W(1) ^ W(15)), 1);	/* 47 */
867 	c = ROTATE_LEFT(d, 5) + H(e, a, b) + c + W(15) + SHA1_CONST(2);
868 	e = ROTATE_LEFT(e, 30);
869 
870 	W(0) = ROTATE_LEFT((W(13) ^ W(8) ^ W(2) ^ W(0)), 1);		/* 48 */
871 	b = ROTATE_LEFT(c, 5) + H(d, e, a) + b + W(0) + SHA1_CONST(2);
872 	d = ROTATE_LEFT(d, 30);
873 
874 	W(1) = ROTATE_LEFT((W(14) ^ W(9) ^ W(3) ^ W(1)), 1);		/* 49 */
875 	a = ROTATE_LEFT(b, 5) + H(c, d, e) + a + W(1) + SHA1_CONST(2);
876 	c = ROTATE_LEFT(c, 30);
877 
878 	W(2) = ROTATE_LEFT((W(15) ^ W(10) ^ W(4) ^ W(2)), 1);	/* 50 */
879 	e = ROTATE_LEFT(a, 5) + H(b, c, d) + e + W(2) + SHA1_CONST(2);
880 	b = ROTATE_LEFT(b, 30);
881 
882 	W(3) = ROTATE_LEFT((W(0) ^ W(11) ^ W(5) ^ W(3)), 1);		/* 51 */
883 	d = ROTATE_LEFT(e, 5) + H(a, b, c) + d + W(3) + SHA1_CONST(2);
884 	a = ROTATE_LEFT(a, 30);
885 
886 	W(4) = ROTATE_LEFT((W(1) ^ W(12) ^ W(6) ^ W(4)), 1);		/* 52 */
887 	c = ROTATE_LEFT(d, 5) + H(e, a, b) + c + W(4) + SHA1_CONST(2);
888 	e = ROTATE_LEFT(e, 30);
889 
890 	W(5) = ROTATE_LEFT((W(2) ^ W(13) ^ W(7) ^ W(5)), 1);		/* 53 */
891 	b = ROTATE_LEFT(c, 5) + H(d, e, a) + b + W(5) + SHA1_CONST(2);
892 	d = ROTATE_LEFT(d, 30);
893 
894 	W(6) = ROTATE_LEFT((W(3) ^ W(14) ^ W(8) ^ W(6)), 1);		/* 54 */
895 	a = ROTATE_LEFT(b, 5) + H(c, d, e) + a + W(6) + SHA1_CONST(2);
896 	c = ROTATE_LEFT(c, 30);
897 
898 	W(7) = ROTATE_LEFT((W(4) ^ W(15) ^ W(9) ^ W(7)), 1);		/* 55 */
899 	e = ROTATE_LEFT(a, 5) + H(b, c, d) + e + W(7) + SHA1_CONST(2);
900 	b = ROTATE_LEFT(b, 30);
901 
902 	W(8) = ROTATE_LEFT((W(5) ^ W(0) ^ W(10) ^ W(8)), 1);		/* 56 */
903 	d = ROTATE_LEFT(e, 5) + H(a, b, c) + d + W(8) + SHA1_CONST(2);
904 	a = ROTATE_LEFT(a, 30);
905 
906 	W(9) = ROTATE_LEFT((W(6) ^ W(1) ^ W(11) ^ W(9)), 1);		/* 57 */
907 	c = ROTATE_LEFT(d, 5) + H(e, a, b) + c + W(9) + SHA1_CONST(2);
908 	e = ROTATE_LEFT(e, 30);
909 
910 	W(10) = ROTATE_LEFT((W(7) ^ W(2) ^ W(12) ^ W(10)), 1);	/* 58 */
911 	b = ROTATE_LEFT(c, 5) + H(d, e, a) + b + W(10) + SHA1_CONST(2);
912 	d = ROTATE_LEFT(d, 30);
913 
914 	W(11) = ROTATE_LEFT((W(8) ^ W(3) ^ W(13) ^ W(11)), 1);	/* 59 */
915 	a = ROTATE_LEFT(b, 5) + H(c, d, e) + a + W(11) + SHA1_CONST(2);
916 	c = ROTATE_LEFT(c, 30);
917 
918 	/* round 4 */
919 	W(12) = ROTATE_LEFT((W(9) ^ W(4) ^ W(14) ^ W(12)), 1);	/* 60 */
920 	e = ROTATE_LEFT(a, 5) + G(b, c, d) + e + W(12) + SHA1_CONST(3);
921 	b = ROTATE_LEFT(b, 30);
922 
923 	W(13) = ROTATE_LEFT((W(10) ^ W(5) ^ W(15) ^ W(13)), 1);	/* 61 */
924 	d = ROTATE_LEFT(e, 5) + G(a, b, c) + d + W(13) + SHA1_CONST(3);
925 	a = ROTATE_LEFT(a, 30);
926 
927 	W(14) = ROTATE_LEFT((W(11) ^ W(6) ^ W(0) ^ W(14)), 1);	/* 62 */
928 	c = ROTATE_LEFT(d, 5) + G(e, a, b) + c + W(14) + SHA1_CONST(3);
929 	e = ROTATE_LEFT(e, 30);
930 
931 	W(15) = ROTATE_LEFT((W(12) ^ W(7) ^ W(1) ^ W(15)), 1);	/* 63 */
932 	b = ROTATE_LEFT(c, 5) + G(d, e, a) + b + W(15) + SHA1_CONST(3);
933 	d = ROTATE_LEFT(d, 30);
934 
935 	W(0) = ROTATE_LEFT((W(13) ^ W(8) ^ W(2) ^ W(0)), 1);		/* 64 */
936 	a = ROTATE_LEFT(b, 5) + G(c, d, e) + a + W(0) + SHA1_CONST(3);
937 	c = ROTATE_LEFT(c, 30);
938 
939 	W(1) = ROTATE_LEFT((W(14) ^ W(9) ^ W(3) ^ W(1)), 1);		/* 65 */
940 	e = ROTATE_LEFT(a, 5) + G(b, c, d) + e + W(1) + SHA1_CONST(3);
941 	b = ROTATE_LEFT(b, 30);
942 
943 	W(2) = ROTATE_LEFT((W(15) ^ W(10) ^ W(4) ^ W(2)), 1);	/* 66 */
944 	d = ROTATE_LEFT(e, 5) + G(a, b, c) + d + W(2) + SHA1_CONST(3);
945 	a = ROTATE_LEFT(a, 30);
946 
947 	W(3) = ROTATE_LEFT((W(0) ^ W(11) ^ W(5) ^ W(3)), 1);		/* 67 */
948 	c = ROTATE_LEFT(d, 5) + G(e, a, b) + c + W(3) + SHA1_CONST(3);
949 	e = ROTATE_LEFT(e, 30);
950 
951 	W(4) = ROTATE_LEFT((W(1) ^ W(12) ^ W(6) ^ W(4)), 1);		/* 68 */
952 	b = ROTATE_LEFT(c, 5) + G(d, e, a) + b + W(4) + SHA1_CONST(3);
953 	d = ROTATE_LEFT(d, 30);
954 
955 	W(5) = ROTATE_LEFT((W(2) ^ W(13) ^ W(7) ^ W(5)), 1);		/* 69 */
956 	a = ROTATE_LEFT(b, 5) + G(c, d, e) + a + W(5) + SHA1_CONST(3);
957 	c = ROTATE_LEFT(c, 30);
958 
959 	W(6) = ROTATE_LEFT((W(3) ^ W(14) ^ W(8) ^ W(6)), 1);		/* 70 */
960 	e = ROTATE_LEFT(a, 5) + G(b, c, d) + e + W(6) + SHA1_CONST(3);
961 	b = ROTATE_LEFT(b, 30);
962 
963 	W(7) = ROTATE_LEFT((W(4) ^ W(15) ^ W(9) ^ W(7)), 1);		/* 71 */
964 	d = ROTATE_LEFT(e, 5) + G(a, b, c) + d + W(7) + SHA1_CONST(3);
965 	a = ROTATE_LEFT(a, 30);
966 
967 	W(8) = ROTATE_LEFT((W(5) ^ W(0) ^ W(10) ^ W(8)), 1);		/* 72 */
968 	c = ROTATE_LEFT(d, 5) + G(e, a, b) + c + W(8) + SHA1_CONST(3);
969 	e = ROTATE_LEFT(e, 30);
970 
971 	W(9) = ROTATE_LEFT((W(6) ^ W(1) ^ W(11) ^ W(9)), 1);		/* 73 */
972 	b = ROTATE_LEFT(c, 5) + G(d, e, a) + b + W(9) + SHA1_CONST(3);
973 	d = ROTATE_LEFT(d, 30);
974 
975 	W(10) = ROTATE_LEFT((W(7) ^ W(2) ^ W(12) ^ W(10)), 1);	/* 74 */
976 	a = ROTATE_LEFT(b, 5) + G(c, d, e) + a + W(10) + SHA1_CONST(3);
977 	c = ROTATE_LEFT(c, 30);
978 
979 	W(11) = ROTATE_LEFT((W(8) ^ W(3) ^ W(13) ^ W(11)), 1);	/* 75 */
980 	e = ROTATE_LEFT(a, 5) + G(b, c, d) + e + W(11) + SHA1_CONST(3);
981 	b = ROTATE_LEFT(b, 30);
982 
983 	W(12) = ROTATE_LEFT((W(9) ^ W(4) ^ W(14) ^ W(12)), 1);	/* 76 */
984 	d = ROTATE_LEFT(e, 5) + G(a, b, c) + d + W(12) + SHA1_CONST(3);
985 	a = ROTATE_LEFT(a, 30);
986 
987 	W(13) = ROTATE_LEFT((W(10) ^ W(5) ^ W(15) ^ W(13)), 1);	/* 77 */
988 	c = ROTATE_LEFT(d, 5) + G(e, a, b) + c + W(13) + SHA1_CONST(3);
989 	e = ROTATE_LEFT(e, 30);
990 
991 	W(14) = ROTATE_LEFT((W(11) ^ W(6) ^ W(0) ^ W(14)), 1);	/* 78 */
992 	b = ROTATE_LEFT(c, 5) + G(d, e, a) + b + W(14) + SHA1_CONST(3);
993 	d = ROTATE_LEFT(d, 30);
994 
995 	W(15) = ROTATE_LEFT((W(12) ^ W(7) ^ W(1) ^ W(15)), 1);	/* 79 */
996 
997 	ctx->state[0] += ROTATE_LEFT(b, 5) + G(c, d, e) + a + W(15) +
998 	    SHA1_CONST(3);
999 	ctx->state[1] += b;
1000 	ctx->state[2] += ROTATE_LEFT(c, 30);
1001 	ctx->state[3] += d;
1002 	ctx->state[4] += e;
1003 
1004 	/* zeroize sensitive information */
1005 	W(0) = W(1) = W(2) = W(3) = W(4) = W(5) = W(6) = W(7) = W(8) = 0;
1006 	W(9) = W(10) = W(11) = W(12) = W(13) = W(14) = W(15) = 0;
1007 }
1008 #endif	/* !__amd64 */
1009 
1010 
1011 /*
1012  * Encode()
1013  *
1014  * purpose: to convert a list of numbers from little endian to big endian
1015  *   input: uint8_t *	: place to store the converted big endian numbers
1016  *	    uint32_t *	: place to get numbers to convert from
1017  *          size_t	: the length of the input in bytes
1018  *  output: void
1019  */
1020 
1021 static void
1022 Encode(uint8_t *_RESTRICT_KYWD output, const uint32_t *_RESTRICT_KYWD input,
1023     size_t len)
1024 {
1025 	size_t		i, j;
1026 
1027 #if	defined(__sparc)
1028 	if (IS_P2ALIGNED(output, sizeof (uint32_t))) {
1029 		for (i = 0, j = 0; j < len; i++, j += 4) {
1030 			/* LINTED: pointer alignment */
1031 			*((uint32_t *)(output + j)) = input[i];
1032 		}
1033 	} else {
1034 #endif	/* little endian -- will work on big endian, but slowly */
1035 		for (i = 0, j = 0; j < len; i++, j += 4) {
1036 			output[j]	= (input[i] >> 24) & 0xff;
1037 			output[j + 1]	= (input[i] >> 16) & 0xff;
1038 			output[j + 2]	= (input[i] >>  8) & 0xff;
1039 			output[j + 3]	= input[i] & 0xff;
1040 		}
1041 #if	defined(__sparc)
1042 	}
1043 #endif
1044 }
1045