xref: /titanic_41/usr/src/common/crypto/md5/md5.c (revision 0bb073995ac5a95bd35f2dd790df1ea3d8c2d507)
1 /*
2  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
3  * Use is subject to license terms.
4  */
5 
6 /*
7  * Cleaned-up and optimized version of MD5, based on the reference
8  * implementation provided in RFC 1321.  See RSA Copyright information
9  * below.
10  */
11 
12 #pragma ident	"%Z%%M%	%I%	%E% SMI"
13 
14 /*
15  * MD5C.C - RSA Data Security, Inc., MD5 message-digest algorithm
16  */
17 
18 /*
19  * Copyright (C) 1991-2, RSA Data Security, Inc. Created 1991. All
20  * rights reserved.
21  *
22  * License to copy and use this software is granted provided that it
23  * is identified as the "RSA Data Security, Inc. MD5 Message-Digest
24  * Algorithm" in all material mentioning or referencing this software
25  * or this function.
26  *
27  * License is also granted to make and use derivative works provided
28  * that such works are identified as "derived from the RSA Data
29  * Security, Inc. MD5 Message-Digest Algorithm" in all material
30  * mentioning or referencing the derived work.
31  *
32  * RSA Data Security, Inc. makes no representations concerning either
33  * the merchantability of this software or the suitability of this
34  * software for any particular purpose. It is provided "as is"
35  * without express or implied warranty of any kind.
36  *
37  * These notices must be retained in any copies of any part of this
38  * documentation and/or software.
39  */
40 
41 #include <sys/types.h>
42 #include <sys/md5.h>
43 #include <sys/md5_consts.h>	/* MD5_CONST() optimization */
44 #include "md5_byteswap.h"
45 #if	!defined(_KERNEL) || defined(_BOOT)
46 #include <strings.h>
47 #endif /* !_KERNEL || _BOOT */
48 
49 #ifdef _KERNEL
50 #include <sys/systm.h>
51 #endif /* _KERNEL */
52 
53 static void Encode(uint8_t *, const uint32_t *, size_t);
54 
55 #if !defined(__amd64)
56 static void MD5Transform(uint32_t, uint32_t, uint32_t, uint32_t, MD5_CTX *,
57     const uint8_t [64]);
58 #else
59 void md5_block_asm_host_order(MD5_CTX *ctx, const void *inpp,
60     unsigned int input_length_in_blocks);
61 #endif /* !defined(__amd64) */
62 
63 static uint8_t PADDING[64] = { 0x80, /* all zeros */ };
64 
65 /*
66  * F, G, H and I are the basic MD5 functions.
67  */
68 #define	F(b, c, d)	(((b) & (c)) | ((~b) & (d)))
69 #define	G(b, c, d)	(((b) & (d)) | ((c) & (~d)))
70 #define	H(b, c, d)	((b) ^ (c) ^ (d))
71 #define	I(b, c, d)	((c) ^ ((b) | (~d)))
72 
73 /*
74  * ROTATE_LEFT rotates x left n bits.
75  */
76 #define	ROTATE_LEFT(x, n)	\
77 	(((x) << (n)) | ((x) >> ((sizeof (x) << 3) - (n))))
78 
79 /*
80  * FF, GG, HH, and II transformations for rounds 1, 2, 3, and 4.
81  * Rotation is separate from addition to prevent recomputation.
82  */
83 
84 #define	FF(a, b, c, d, x, s, ac) { \
85 	(a) += F((b), (c), (d)) + (x) + ((unsigned long long)(ac)); \
86 	(a) = ROTATE_LEFT((a), (s)); \
87 	(a) += (b); \
88 	}
89 
90 #define	GG(a, b, c, d, x, s, ac) { \
91 	(a) += G((b), (c), (d)) + (x) + ((unsigned long long)(ac)); \
92 	(a) = ROTATE_LEFT((a), (s)); \
93 	(a) += (b); \
94 	}
95 
96 #define	HH(a, b, c, d, x, s, ac) { \
97 	(a) += H((b), (c), (d)) + (x) + ((unsigned long long)(ac)); \
98 	(a) = ROTATE_LEFT((a), (s)); \
99 	(a) += (b); \
100 	}
101 
102 #define	II(a, b, c, d, x, s, ac) { \
103 	(a) += I((b), (c), (d)) + (x) + ((unsigned long long)(ac)); \
104 	(a) = ROTATE_LEFT((a), (s)); \
105 	(a) += (b); \
106 	}
107 
108 /*
109  * Loading 32-bit constants on a RISC is expensive since it involves both a
110  * `sethi' and an `or'.  thus, we instead have the compiler generate `ld's to
111  * load the constants from an array called `md5_consts'.  however, on intel
112  * (and other CISC processors), it is cheaper to load the constant
113  * directly.  thus, the c code in MD5Transform() uses the macro MD5_CONST()
114  * which either expands to a constant or an array reference, depending on the
115  * architecture the code is being compiled for.
116  *
117  * Right now, i386 and amd64 are the CISC exceptions.
118  * If we get another CISC ISA, we'll have to change the ifdef.
119  */
120 
121 #if defined(__i386) || defined(__amd64)
122 
123 #define	MD5_CONST(x)		(MD5_CONST_ ## x)
124 #define	MD5_CONST_e(x)		MD5_CONST(x)
125 #define	MD5_CONST_o(x)		MD5_CONST(x)
126 
127 #else
128 /*
129  * sparc/RISC optimization:
130  *
131  * while it is somewhat counter-intuitive, on sparc (and presumably other RISC
132  * machines), it is more efficient to place all the constants used in this
133  * function in an array and load the values out of the array than to manually
134  * load the constants.  this is because setting a register to a 32-bit value
135  * takes two ops in most cases: a `sethi' and an `or', but loading a 32-bit
136  * value from memory only takes one `ld' (or `lduw' on v9).  while this
137  * increases memory usage, the compiler can find enough other things to do
138  * while waiting to keep the pipeline does not stall.  additionally, it is
139  * likely that many of these constants are cached so that later accesses do
140  * not even go out to the bus.
141  *
142  * this array is declared `static' to keep the compiler from having to
143  * bcopy() this array onto the stack frame of MD5Transform() each time it is
144  * called -- which is unacceptably expensive.
145  *
146  * the `const' is to ensure that callers are good citizens and do not try to
147  * munge the array.  since these routines are going to be called from inside
148  * multithreaded kernelland, this is a good safety check. -- `constants' will
149  * end up in .rodata.
150  *
151  * unfortunately, loading from an array in this manner hurts performance under
152  * intel (and presumably other CISC machines).  so, there is a macro,
153  * MD5_CONST(), used in MD5Transform(), that either expands to a reference to
154  * this array, or to the actual constant, depending on what platform this code
155  * is compiled for.
156  */
157 
158 #ifdef sun4v
159 
160 /*
161  * Going to load these consts in 8B chunks, so need to enforce 8B alignment
162  */
163 
164 /* CSTYLED */
165 #pragma align 64 (md5_consts)
166 #define	_MD5_CHECK_ALIGNMENT
167 
168 #endif /* sun4v */
169 
170 static const uint32_t md5_consts[] = {
171 	MD5_CONST_0,	MD5_CONST_1,	MD5_CONST_2,	MD5_CONST_3,
172 	MD5_CONST_4,	MD5_CONST_5,	MD5_CONST_6,	MD5_CONST_7,
173 	MD5_CONST_8,	MD5_CONST_9,	MD5_CONST_10,	MD5_CONST_11,
174 	MD5_CONST_12,	MD5_CONST_13,	MD5_CONST_14,	MD5_CONST_15,
175 	MD5_CONST_16,	MD5_CONST_17,	MD5_CONST_18,	MD5_CONST_19,
176 	MD5_CONST_20,	MD5_CONST_21,	MD5_CONST_22,	MD5_CONST_23,
177 	MD5_CONST_24,	MD5_CONST_25,	MD5_CONST_26,	MD5_CONST_27,
178 	MD5_CONST_28,	MD5_CONST_29,	MD5_CONST_30,	MD5_CONST_31,
179 	MD5_CONST_32,	MD5_CONST_33,	MD5_CONST_34,	MD5_CONST_35,
180 	MD5_CONST_36,	MD5_CONST_37,	MD5_CONST_38,	MD5_CONST_39,
181 	MD5_CONST_40,	MD5_CONST_41,	MD5_CONST_42,	MD5_CONST_43,
182 	MD5_CONST_44,	MD5_CONST_45,	MD5_CONST_46,	MD5_CONST_47,
183 	MD5_CONST_48,	MD5_CONST_49,	MD5_CONST_50,	MD5_CONST_51,
184 	MD5_CONST_52,	MD5_CONST_53,	MD5_CONST_54,	MD5_CONST_55,
185 	MD5_CONST_56,	MD5_CONST_57,	MD5_CONST_58,	MD5_CONST_59,
186 	MD5_CONST_60,	MD5_CONST_61,	MD5_CONST_62,	MD5_CONST_63
187 };
188 
189 
190 #ifdef sun4v
191 /*
192  * To reduce the number of loads, load consts in 64-bit
193  * chunks and then split.
194  *
195  * No need to mask upper 32-bits, as just interested in
196  * low 32-bits (saves an & operation and means that this
197  * optimization doesn't increases the icount.
198  */
199 #define	MD5_CONST_e(x)		(md5_consts64[x/2] >> 32)
200 #define	MD5_CONST_o(x)		(md5_consts64[x/2])
201 
202 #else
203 
204 #define	MD5_CONST_e(x)		(md5_consts[x])
205 #define	MD5_CONST_o(x)		(md5_consts[x])
206 
207 #endif /* sun4v */
208 
209 #endif
210 
211 /*
212  * MD5Init()
213  *
214  * purpose: initializes the md5 context and begins and md5 digest operation
215  *   input: MD5_CTX *	: the context to initialize.
216  *  output: void
217  */
218 
219 void
220 MD5Init(MD5_CTX *ctx)
221 {
222 	ctx->count[0] = ctx->count[1] = 0;
223 
224 	/* load magic initialization constants */
225 	ctx->state[0] = MD5_INIT_CONST_1;
226 	ctx->state[1] = MD5_INIT_CONST_2;
227 	ctx->state[2] = MD5_INIT_CONST_3;
228 	ctx->state[3] = MD5_INIT_CONST_4;
229 }
230 
231 /*
232  * MD5Update()
233  *
234  * purpose: continues an md5 digest operation, using the message block
235  *          to update the context.
236  *   input: MD5_CTX *	: the context to update
237  *          uint8_t *	: the message block
238  *          uint32_t    : the length of the message block in bytes
239  *  output: void
240  *
241  * MD5 crunches in 64-byte blocks.  All numeric constants here are related to
242  * that property of MD5.
243  */
244 
245 void
246 MD5Update(MD5_CTX *ctx, const void *inpp, unsigned int input_len)
247 {
248 	uint32_t		i, buf_index, buf_len;
249 #ifdef	sun4v
250 	uint32_t		old_asi;
251 #endif	/* sun4v */
252 #if defined(__amd64)
253 	uint32_t		block_count;
254 #endif /* !defined(__amd64) */
255 	const unsigned char 	*input = (const unsigned char *)inpp;
256 
257 	/* compute (number of bytes computed so far) mod 64 */
258 	buf_index = (ctx->count[0] >> 3) & 0x3F;
259 
260 	/* update number of bits hashed into this MD5 computation so far */
261 	if ((ctx->count[0] += (input_len << 3)) < (input_len << 3))
262 		ctx->count[1]++;
263 	ctx->count[1] += (input_len >> 29);
264 
265 	buf_len = 64 - buf_index;
266 
267 	/* transform as many times as possible */
268 	i = 0;
269 	if (input_len >= buf_len) {
270 
271 		/*
272 		 * general optimization:
273 		 *
274 		 * only do initial bcopy() and MD5Transform() if
275 		 * buf_index != 0.  if buf_index == 0, we're just
276 		 * wasting our time doing the bcopy() since there
277 		 * wasn't any data left over from a previous call to
278 		 * MD5Update().
279 		 */
280 
281 #ifdef sun4v
282 		/*
283 		 * For N1 use %asi register. However, costly to repeatedly set
284 		 * in MD5Transform. Therefore, set once here.
285 		 * Should probably restore the old value afterwards...
286 		 */
287 		old_asi = get_little();
288 		set_little(0x88);
289 #endif /* sun4v */
290 
291 		if (buf_index) {
292 			bcopy(input, &ctx->buf_un.buf8[buf_index], buf_len);
293 
294 #if !defined(__amd64)
295 			MD5Transform(ctx->state[0], ctx->state[1],
296 			    ctx->state[2], ctx->state[3], ctx,
297 			    ctx->buf_un.buf8);
298 #else
299 			md5_block_asm_host_order(ctx, ctx->buf_un.buf8, 1);
300 #endif /* !defined(__amd64) */
301 
302 			i = buf_len;
303 		}
304 
305 #if !defined(__amd64)
306 		for (; i + 63 < input_len; i += 64)
307 			MD5Transform(ctx->state[0], ctx->state[1],
308 			    ctx->state[2], ctx->state[3], ctx, &input[i]);
309 
310 #else
311 		block_count = (input_len - i) >> 6;
312 		if (block_count > 0) {
313 			md5_block_asm_host_order(ctx, &input[i], block_count);
314 			i += block_count << 6;
315 		}
316 #endif /* !defined(__amd64) */
317 
318 
319 #ifdef sun4v
320 		/*
321 		 * Restore old %ASI value
322 		 */
323 		set_little(old_asi);
324 #endif /* sun4v */
325 
326 		/*
327 		 * general optimization:
328 		 *
329 		 * if i and input_len are the same, return now instead
330 		 * of calling bcopy(), since the bcopy() in this
331 		 * case will be an expensive nop.
332 		 */
333 
334 		if (input_len == i)
335 			return;
336 
337 		buf_index = 0;
338 	}
339 
340 	/* buffer remaining input */
341 	bcopy(&input[i], &ctx->buf_un.buf8[buf_index], input_len - i);
342 }
343 
344 /*
345  * MD5Final()
346  *
347  * purpose: ends an md5 digest operation, finalizing the message digest and
348  *          zeroing the context.
349  *   input: uchar_t *	: a buffer to store the digest in
350  *			: The function actually uses void* because many
351  *			: callers pass things other than uchar_t here.
352  *          MD5_CTX *   : the context to finalize, save, and zero
353  *  output: void
354  */
355 
356 void
357 MD5Final(void *digest, MD5_CTX *ctx)
358 {
359 	uint8_t		bitcount_le[sizeof (ctx->count)];
360 	uint32_t	index = (ctx->count[0] >> 3) & 0x3f;
361 
362 	/* store bit count, little endian */
363 	Encode(bitcount_le, ctx->count, sizeof (bitcount_le));
364 
365 	/* pad out to 56 mod 64 */
366 	MD5Update(ctx, PADDING, ((index < 56) ? 56 : 120) - index);
367 
368 	/* append length (before padding) */
369 	MD5Update(ctx, bitcount_le, sizeof (bitcount_le));
370 
371 	/* store state in digest */
372 	Encode(digest, ctx->state, sizeof (ctx->state));
373 
374 	/* zeroize sensitive information */
375 	bzero(ctx, sizeof (*ctx));
376 }
377 
378 #ifndef	_KERNEL
379 
380 void
381 md5_calc(unsigned char *output, unsigned char *input, unsigned int inlen)
382 {
383 	MD5_CTX context;
384 
385 	MD5Init(&context);
386 	MD5Update(&context, input, inlen);
387 	MD5Final(output, &context);
388 }
389 
390 #endif	/* !_KERNEL */
391 
392 #if !defined(__amd64)
393 /*
394  * sparc register window optimization:
395  *
396  * `a', `b', `c', and `d' are passed into MD5Transform explicitly
397  * since it increases the number of registers available to the
398  * compiler.  under this scheme, these variables can be held in
399  * %i0 - %i3, which leaves more local and out registers available.
400  */
401 
402 /*
403  * MD5Transform()
404  *
405  * purpose: md5 transformation -- updates the digest based on `block'
406  *   input: uint32_t	: bytes  1 -  4 of the digest
407  *          uint32_t	: bytes  5 -  8 of the digest
408  *          uint32_t	: bytes  9 - 12 of the digest
409  *          uint32_t	: bytes 12 - 16 of the digest
410  *          MD5_CTX *   : the context to update
411  *          uint8_t [64]: the block to use to update the digest
412  *  output: void
413  */
414 
415 static void
416 MD5Transform(uint32_t a, uint32_t b, uint32_t c, uint32_t d,
417     MD5_CTX *ctx, const uint8_t block[64])
418 {
419 	/*
420 	 * general optimization:
421 	 *
422 	 * use individual integers instead of using an array.  this is a
423 	 * win, although the amount it wins by seems to vary quite a bit.
424 	 */
425 
426 	register uint32_t	x_0, x_1, x_2,  x_3,  x_4,  x_5,  x_6,  x_7;
427 	register uint32_t	x_8, x_9, x_10, x_11, x_12, x_13, x_14, x_15;
428 #ifdef sun4v
429 	unsigned long long 	*md5_consts64;
430 
431 		/* LINTED E_BAD_PTR_CAST_ALIGN */
432 	md5_consts64 = (unsigned long long *) md5_consts;
433 #endif	/* sun4v */
434 
435 	/*
436 	 * general optimization:
437 	 *
438 	 * the compiler (at least SC4.2/5.x) generates better code if
439 	 * variable use is localized.  in this case, swapping the integers in
440 	 * this order allows `x_0 'to be swapped nearest to its first use in
441 	 * FF(), and likewise for `x_1' and up.  note that the compiler
442 	 * prefers this to doing each swap right before the FF() that
443 	 * uses it.
444 	 */
445 
446 	/*
447 	 * sparc v9/v8plus optimization:
448 	 *
449 	 * if `block' is already aligned on a 4-byte boundary, use the
450 	 * optimized load_little_32() directly.  otherwise, bcopy()
451 	 * into a buffer that *is* aligned on a 4-byte boundary and
452 	 * then do the load_little_32() on that buffer.  benchmarks
453 	 * have shown that using the bcopy() is better than loading
454 	 * the bytes individually and doing the endian-swap by hand.
455 	 *
456 	 * even though it's quite tempting to assign to do:
457 	 *
458 	 * blk = bcopy(blk, ctx->buf_un.buf32, sizeof (ctx->buf_un.buf32));
459 	 *
460 	 * and only have one set of LOAD_LITTLE_32()'s, the compiler (at least
461 	 * SC4.2/5.x) *does not* like that, so please resist the urge.
462 	 */
463 
464 #ifdef _MD5_CHECK_ALIGNMENT
465 	if ((uintptr_t)block & 0x3) {		/* not 4-byte aligned? */
466 		bcopy(block, ctx->buf_un.buf32, sizeof (ctx->buf_un.buf32));
467 
468 #ifdef sun4v
469 		x_15 = LOAD_LITTLE_32_f(ctx->buf_un.buf32);
470 		x_14 = LOAD_LITTLE_32_e(ctx->buf_un.buf32);
471 		x_13 = LOAD_LITTLE_32_d(ctx->buf_un.buf32);
472 		x_12 = LOAD_LITTLE_32_c(ctx->buf_un.buf32);
473 		x_11 = LOAD_LITTLE_32_b(ctx->buf_un.buf32);
474 		x_10 = LOAD_LITTLE_32_a(ctx->buf_un.buf32);
475 		x_9  = LOAD_LITTLE_32_9(ctx->buf_un.buf32);
476 		x_8  = LOAD_LITTLE_32_8(ctx->buf_un.buf32);
477 		x_7  = LOAD_LITTLE_32_7(ctx->buf_un.buf32);
478 		x_6  = LOAD_LITTLE_32_6(ctx->buf_un.buf32);
479 		x_5  = LOAD_LITTLE_32_5(ctx->buf_un.buf32);
480 		x_4  = LOAD_LITTLE_32_4(ctx->buf_un.buf32);
481 		x_3  = LOAD_LITTLE_32_3(ctx->buf_un.buf32);
482 		x_2  = LOAD_LITTLE_32_2(ctx->buf_un.buf32);
483 		x_1  = LOAD_LITTLE_32_1(ctx->buf_un.buf32);
484 		x_0  = LOAD_LITTLE_32_0(ctx->buf_un.buf32);
485 #else
486 		x_15 = LOAD_LITTLE_32(ctx->buf_un.buf32 + 15);
487 		x_14 = LOAD_LITTLE_32(ctx->buf_un.buf32 + 14);
488 		x_13 = LOAD_LITTLE_32(ctx->buf_un.buf32 + 13);
489 		x_12 = LOAD_LITTLE_32(ctx->buf_un.buf32 + 12);
490 		x_11 = LOAD_LITTLE_32(ctx->buf_un.buf32 + 11);
491 		x_10 = LOAD_LITTLE_32(ctx->buf_un.buf32 + 10);
492 		x_9  = LOAD_LITTLE_32(ctx->buf_un.buf32 +  9);
493 		x_8  = LOAD_LITTLE_32(ctx->buf_un.buf32 +  8);
494 		x_7  = LOAD_LITTLE_32(ctx->buf_un.buf32 +  7);
495 		x_6  = LOAD_LITTLE_32(ctx->buf_un.buf32 +  6);
496 		x_5  = LOAD_LITTLE_32(ctx->buf_un.buf32 +  5);
497 		x_4  = LOAD_LITTLE_32(ctx->buf_un.buf32 +  4);
498 		x_3  = LOAD_LITTLE_32(ctx->buf_un.buf32 +  3);
499 		x_2  = LOAD_LITTLE_32(ctx->buf_un.buf32 +  2);
500 		x_1  = LOAD_LITTLE_32(ctx->buf_un.buf32 +  1);
501 		x_0  = LOAD_LITTLE_32(ctx->buf_un.buf32 +  0);
502 #endif /* sun4v */
503 	} else
504 #endif
505 	{
506 
507 #ifdef sun4v
508 		/* LINTED E_BAD_PTR_CAST_ALIGN */
509 		x_15 = LOAD_LITTLE_32_f(block);
510 		/* LINTED E_BAD_PTR_CAST_ALIGN */
511 		x_14 = LOAD_LITTLE_32_e(block);
512 		/* LINTED E_BAD_PTR_CAST_ALIGN */
513 		x_13 = LOAD_LITTLE_32_d(block);
514 		/* LINTED E_BAD_PTR_CAST_ALIGN */
515 		x_12 = LOAD_LITTLE_32_c(block);
516 		/* LINTED E_BAD_PTR_CAST_ALIGN */
517 		x_11 = LOAD_LITTLE_32_b(block);
518 		/* LINTED E_BAD_PTR_CAST_ALIGN */
519 		x_10 = LOAD_LITTLE_32_a(block);
520 		/* LINTED E_BAD_PTR_CAST_ALIGN */
521 		x_9  = LOAD_LITTLE_32_9(block);
522 		/* LINTED E_BAD_PTR_CAST_ALIGN */
523 		x_8  = LOAD_LITTLE_32_8(block);
524 		/* LINTED E_BAD_PTR_CAST_ALIGN */
525 		x_7  = LOAD_LITTLE_32_7(block);
526 		/* LINTED E_BAD_PTR_CAST_ALIGN */
527 		x_6  = LOAD_LITTLE_32_6(block);
528 		/* LINTED E_BAD_PTR_CAST_ALIGN */
529 		x_5  = LOAD_LITTLE_32_5(block);
530 		/* LINTED E_BAD_PTR_CAST_ALIGN */
531 		x_4  = LOAD_LITTLE_32_4(block);
532 		/* LINTED E_BAD_PTR_CAST_ALIGN */
533 		x_3  = LOAD_LITTLE_32_3(block);
534 		/* LINTED E_BAD_PTR_CAST_ALIGN */
535 		x_2  = LOAD_LITTLE_32_2(block);
536 		/* LINTED E_BAD_PTR_CAST_ALIGN */
537 		x_1  = LOAD_LITTLE_32_1(block);
538 		/* LINTED E_BAD_PTR_CAST_ALIGN */
539 		x_0  = LOAD_LITTLE_32_0(block);
540 #else
541 		/* LINTED E_BAD_PTR_CAST_ALIGN */
542 		x_15 = LOAD_LITTLE_32(block + 60);
543 		/* LINTED E_BAD_PTR_CAST_ALIGN */
544 		x_14 = LOAD_LITTLE_32(block + 56);
545 		/* LINTED E_BAD_PTR_CAST_ALIGN */
546 		x_13 = LOAD_LITTLE_32(block + 52);
547 		/* LINTED E_BAD_PTR_CAST_ALIGN */
548 		x_12 = LOAD_LITTLE_32(block + 48);
549 		/* LINTED E_BAD_PTR_CAST_ALIGN */
550 		x_11 = LOAD_LITTLE_32(block + 44);
551 		/* LINTED E_BAD_PTR_CAST_ALIGN */
552 		x_10 = LOAD_LITTLE_32(block + 40);
553 		/* LINTED E_BAD_PTR_CAST_ALIGN */
554 		x_9  = LOAD_LITTLE_32(block + 36);
555 		/* LINTED E_BAD_PTR_CAST_ALIGN */
556 		x_8  = LOAD_LITTLE_32(block + 32);
557 		/* LINTED E_BAD_PTR_CAST_ALIGN */
558 		x_7  = LOAD_LITTLE_32(block + 28);
559 		/* LINTED E_BAD_PTR_CAST_ALIGN */
560 		x_6  = LOAD_LITTLE_32(block + 24);
561 		/* LINTED E_BAD_PTR_CAST_ALIGN */
562 		x_5  = LOAD_LITTLE_32(block + 20);
563 		/* LINTED E_BAD_PTR_CAST_ALIGN */
564 		x_4  = LOAD_LITTLE_32(block + 16);
565 		/* LINTED E_BAD_PTR_CAST_ALIGN */
566 		x_3  = LOAD_LITTLE_32(block + 12);
567 		/* LINTED E_BAD_PTR_CAST_ALIGN */
568 		x_2  = LOAD_LITTLE_32(block +  8);
569 		/* LINTED E_BAD_PTR_CAST_ALIGN */
570 		x_1  = LOAD_LITTLE_32(block +  4);
571 		/* LINTED E_BAD_PTR_CAST_ALIGN */
572 		x_0  = LOAD_LITTLE_32(block +  0);
573 #endif /* sun4v */
574 	}
575 
576 	/* round 1 */
577 	FF(a, b, c, d, 	x_0, MD5_SHIFT_11, MD5_CONST_e(0));  /* 1 */
578 	FF(d, a, b, c, 	x_1, MD5_SHIFT_12, MD5_CONST_o(1));  /* 2 */
579 	FF(c, d, a, b, 	x_2, MD5_SHIFT_13, MD5_CONST_e(2));  /* 3 */
580 	FF(b, c, d, a, 	x_3, MD5_SHIFT_14, MD5_CONST_o(3));  /* 4 */
581 	FF(a, b, c, d, 	x_4, MD5_SHIFT_11, MD5_CONST_e(4));  /* 5 */
582 	FF(d, a, b, c, 	x_5, MD5_SHIFT_12, MD5_CONST_o(5));  /* 6 */
583 	FF(c, d, a, b, 	x_6, MD5_SHIFT_13, MD5_CONST_e(6));  /* 7 */
584 	FF(b, c, d, a, 	x_7, MD5_SHIFT_14, MD5_CONST_o(7));  /* 8 */
585 	FF(a, b, c, d, 	x_8, MD5_SHIFT_11, MD5_CONST_e(8));  /* 9 */
586 	FF(d, a, b, c, 	x_9, MD5_SHIFT_12, MD5_CONST_o(9));  /* 10 */
587 	FF(c, d, a, b, x_10, MD5_SHIFT_13, MD5_CONST_e(10)); /* 11 */
588 	FF(b, c, d, a, x_11, MD5_SHIFT_14, MD5_CONST_o(11)); /* 12 */
589 	FF(a, b, c, d, x_12, MD5_SHIFT_11, MD5_CONST_e(12)); /* 13 */
590 	FF(d, a, b, c, x_13, MD5_SHIFT_12, MD5_CONST_o(13)); /* 14 */
591 	FF(c, d, a, b, x_14, MD5_SHIFT_13, MD5_CONST_e(14)); /* 15 */
592 	FF(b, c, d, a, x_15, MD5_SHIFT_14, MD5_CONST_o(15)); /* 16 */
593 
594 	/* round 2 */
595 	GG(a, b, c, d,  x_1, MD5_SHIFT_21, MD5_CONST_e(16)); /* 17 */
596 	GG(d, a, b, c,  x_6, MD5_SHIFT_22, MD5_CONST_o(17)); /* 18 */
597 	GG(c, d, a, b, x_11, MD5_SHIFT_23, MD5_CONST_e(18)); /* 19 */
598 	GG(b, c, d, a,  x_0, MD5_SHIFT_24, MD5_CONST_o(19)); /* 20 */
599 	GG(a, b, c, d,  x_5, MD5_SHIFT_21, MD5_CONST_e(20)); /* 21 */
600 	GG(d, a, b, c, x_10, MD5_SHIFT_22, MD5_CONST_o(21)); /* 22 */
601 	GG(c, d, a, b, x_15, MD5_SHIFT_23, MD5_CONST_e(22)); /* 23 */
602 	GG(b, c, d, a,  x_4, MD5_SHIFT_24, MD5_CONST_o(23)); /* 24 */
603 	GG(a, b, c, d,  x_9, MD5_SHIFT_21, MD5_CONST_e(24)); /* 25 */
604 	GG(d, a, b, c, x_14, MD5_SHIFT_22, MD5_CONST_o(25)); /* 26 */
605 	GG(c, d, a, b,  x_3, MD5_SHIFT_23, MD5_CONST_e(26)); /* 27 */
606 	GG(b, c, d, a,  x_8, MD5_SHIFT_24, MD5_CONST_o(27)); /* 28 */
607 	GG(a, b, c, d, x_13, MD5_SHIFT_21, MD5_CONST_e(28)); /* 29 */
608 	GG(d, a, b, c,  x_2, MD5_SHIFT_22, MD5_CONST_o(29)); /* 30 */
609 	GG(c, d, a, b,  x_7, MD5_SHIFT_23, MD5_CONST_e(30)); /* 31 */
610 	GG(b, c, d, a, x_12, MD5_SHIFT_24, MD5_CONST_o(31)); /* 32 */
611 
612 	/* round 3 */
613 	HH(a, b, c, d,  x_5, MD5_SHIFT_31, MD5_CONST_e(32)); /* 33 */
614 	HH(d, a, b, c,  x_8, MD5_SHIFT_32, MD5_CONST_o(33)); /* 34 */
615 	HH(c, d, a, b, x_11, MD5_SHIFT_33, MD5_CONST_e(34)); /* 35 */
616 	HH(b, c, d, a, x_14, MD5_SHIFT_34, MD5_CONST_o(35)); /* 36 */
617 	HH(a, b, c, d,  x_1, MD5_SHIFT_31, MD5_CONST_e(36)); /* 37 */
618 	HH(d, a, b, c,  x_4, MD5_SHIFT_32, MD5_CONST_o(37)); /* 38 */
619 	HH(c, d, a, b,  x_7, MD5_SHIFT_33, MD5_CONST_e(38)); /* 39 */
620 	HH(b, c, d, a, x_10, MD5_SHIFT_34, MD5_CONST_o(39)); /* 40 */
621 	HH(a, b, c, d, x_13, MD5_SHIFT_31, MD5_CONST_e(40)); /* 41 */
622 	HH(d, a, b, c,  x_0, MD5_SHIFT_32, MD5_CONST_o(41)); /* 42 */
623 	HH(c, d, a, b,  x_3, MD5_SHIFT_33, MD5_CONST_e(42)); /* 43 */
624 	HH(b, c, d, a,  x_6, MD5_SHIFT_34, MD5_CONST_o(43)); /* 44 */
625 	HH(a, b, c, d,  x_9, MD5_SHIFT_31, MD5_CONST_e(44)); /* 45 */
626 	HH(d, a, b, c, x_12, MD5_SHIFT_32, MD5_CONST_o(45)); /* 46 */
627 	HH(c, d, a, b, x_15, MD5_SHIFT_33, MD5_CONST_e(46)); /* 47 */
628 	HH(b, c, d, a,  x_2, MD5_SHIFT_34, MD5_CONST_o(47)); /* 48 */
629 
630 	/* round 4 */
631 	II(a, b, c, d,  x_0, MD5_SHIFT_41, MD5_CONST_e(48)); /* 49 */
632 	II(d, a, b, c,  x_7, MD5_SHIFT_42, MD5_CONST_o(49)); /* 50 */
633 	II(c, d, a, b, x_14, MD5_SHIFT_43, MD5_CONST_e(50)); /* 51 */
634 	II(b, c, d, a,  x_5, MD5_SHIFT_44, MD5_CONST_o(51)); /* 52 */
635 	II(a, b, c, d, x_12, MD5_SHIFT_41, MD5_CONST_e(52)); /* 53 */
636 	II(d, a, b, c,  x_3, MD5_SHIFT_42, MD5_CONST_o(53)); /* 54 */
637 	II(c, d, a, b, x_10, MD5_SHIFT_43, MD5_CONST_e(54)); /* 55 */
638 	II(b, c, d, a,  x_1, MD5_SHIFT_44, MD5_CONST_o(55)); /* 56 */
639 	II(a, b, c, d,  x_8, MD5_SHIFT_41, MD5_CONST_e(56)); /* 57 */
640 	II(d, a, b, c, x_15, MD5_SHIFT_42, MD5_CONST_o(57)); /* 58 */
641 	II(c, d, a, b,  x_6, MD5_SHIFT_43, MD5_CONST_e(58)); /* 59 */
642 	II(b, c, d, a, x_13, MD5_SHIFT_44, MD5_CONST_o(59)); /* 60 */
643 	II(a, b, c, d,  x_4, MD5_SHIFT_41, MD5_CONST_e(60)); /* 61 */
644 	II(d, a, b, c, x_11, MD5_SHIFT_42, MD5_CONST_o(61)); /* 62 */
645 	II(c, d, a, b,  x_2, MD5_SHIFT_43, MD5_CONST_e(62)); /* 63 */
646 	II(b, c, d, a,  x_9, MD5_SHIFT_44, MD5_CONST_o(63)); /* 64 */
647 
648 	ctx->state[0] += a;
649 	ctx->state[1] += b;
650 	ctx->state[2] += c;
651 	ctx->state[3] += d;
652 
653 	/*
654 	 * zeroize sensitive information -- compiler will optimize
655 	 * this out if everything is kept in registers
656 	 */
657 
658 	x_0 = x_1  = x_2  = x_3  = x_4  = x_5  = x_6  = x_7 = x_8 = 0;
659 	x_9 = x_10 = x_11 = x_12 = x_13 = x_14 = x_15 = 0;
660 }
661 #endif /* !defined(__amd64) */
662 
663 /*
664  * Encode()
665  *
666  * purpose: to convert a list of numbers from big endian to little endian
667  *   input: uint8_t *	: place to store the converted little endian numbers
668  *	    uint32_t *	: place to get numbers to convert from
669  *          size_t	: the length of the input in bytes
670  *  output: void
671  */
672 
673 static void
674 Encode(uint8_t *_RESTRICT_KYWD output, const uint32_t *_RESTRICT_KYWD input,
675     size_t input_len)
676 {
677 	size_t		i, j;
678 
679 	for (i = 0, j = 0; j < input_len; i++, j += sizeof (uint32_t)) {
680 
681 #ifdef _LITTLE_ENDIAN
682 
683 #ifdef _MD5_CHECK_ALIGNMENT
684 		if ((uintptr_t)output & 0x3)	/* Not 4-byte aligned */
685 			bcopy(input + i, output + j, 4);
686 		else *(uint32_t *)(output + j) = input[i];
687 #else
688 		/*LINTED E_BAD_PTR_CAST_ALIGN*/
689 		*(uint32_t *)(output + j) = input[i];
690 #endif /* _MD5_CHECK_ALIGNMENT */
691 
692 #else	/* big endian -- will work on little endian, but slowly */
693 
694 		output[j] = input[i] & 0xff;
695 		output[j + 1] = (input[i] >> 8)  & 0xff;
696 		output[j + 2] = (input[i] >> 16) & 0xff;
697 		output[j + 3] = (input[i] >> 24) & 0xff;
698 #endif
699 	}
700 }
701