2 * Software implementation of AES.
4 * This implementation uses a bit-sliced representation. Instead of
5 * the obvious approach of storing the cipher state so that each byte
6 * (or field element, or entry in the cipher matrix) occupies 8
7 * contiguous bits in a machine integer somewhere, we organise the
8 * cipher state as an array of 8 integers, in such a way that each
9 * logical byte of the cipher state occupies one bit in each integer,
10 * all at the same position. This allows us to do parallel logic on
11 * all bytes of the state by doing bitwise operations between the 8
12 * integers; in particular, the S-box (SubBytes) lookup is done this
13 * way, which takes about 110 operations - but for those 110 bitwise
14 * ops you get 64 S-box lookups, not just one.
19 #include "mpint_i.h" /* we reuse the BignumInt system */
21 static bool aes_sw_available(void)
23 /* Software AES is always available */
27 #define SLICE_PARALLELISM (BIGNUM_INT_BYTES / 2)
29 #ifdef BITSLICED_DEBUG
30 /* Dump function that undoes the bitslicing transform, so you can see
31 * the logical data represented by a set of slice words. */
32 static inline void dumpslices_uint16_t(
33 const char *prefix
, const uint16_t slices
[8])
35 printf("%-30s", prefix
);
36 for (unsigned byte
= 0; byte
< 16; byte
++) {
38 for (unsigned bit
= 0; bit
< 8; bit
++)
39 byteval
|= (1 & (slices
[bit
] >> byte
)) << bit
;
40 printf("%02x", byteval
);
45 static inline void dumpslices_BignumInt(
46 const char *prefix
, const BignumInt slices
[8])
48 printf("%-30s", prefix
);
49 for (unsigned iter
= 0; iter
< SLICE_PARALLELISM
; iter
++) {
50 for (unsigned byte
= 0; byte
< 16; byte
++) {
52 for (unsigned bit
= 0; bit
< 8; bit
++)
53 byteval
|= (1 & (slices
[bit
] >> (iter
*16+byte
))) << bit
;
54 printf("%02x", byteval
);
56 if (iter
+1 < SLICE_PARALLELISM
)
62 #define dumpslices_uintN_t(prefix, slices) ((void)0)
63 #define dumpslices_BignumInt(prefix, slices) ((void)0)
67 * Bit-slicing transformation: convert between an array of 16 uint8_t
68 * and an array of 8 uint16_t, so as to interchange the bit index
69 * within each element and the element index within the array. (That
70 * is, bit j of input[i] == bit i of output[j].
73 #define SWAPWORDS(shift) do \
75 uint64_t mask = ~(uint64_t)0 / ((1ULL << shift) + 1); \
76 uint64_t diff = ((i0 >> shift) ^ i1) & mask; \
77 i0 ^= diff << shift; \
81 #define SWAPINWORD(i, bigshift, smallshift) do \
83 uint64_t mask = ~(uint64_t)0; \
84 mask /= ((1ULL << bigshift) + 1); \
85 mask /= ((1ULL << smallshift) + 1); \
86 mask <<= smallshift; \
87 unsigned shift = bigshift - smallshift; \
88 uint64_t diff = ((i >> shift) ^ i) & mask; \
89 i ^= diff ^ (diff << shift); \
92 #define TO_BITSLICES(slices, bytes, uintN_t, assign_op, shift) do \
94 uint64_t i0 = GET_64BIT_LSB_FIRST(bytes); \
95 uint64_t i1 = GET_64BIT_LSB_FIRST(bytes + 8); \
96 SWAPINWORD(i0, 8, 1); \
97 SWAPINWORD(i1, 8, 1); \
98 SWAPINWORD(i0, 16, 2); \
99 SWAPINWORD(i1, 16, 2); \
100 SWAPINWORD(i0, 32, 4); \
101 SWAPINWORD(i1, 32, 4); \
103 slices[0] assign_op (uintN_t)((i0 >> 0) & 0xFFFF) << (shift); \
104 slices[2] assign_op (uintN_t)((i0 >> 16) & 0xFFFF) << (shift); \
105 slices[4] assign_op (uintN_t)((i0 >> 32) & 0xFFFF) << (shift); \
106 slices[6] assign_op (uintN_t)((i0 >> 48) & 0xFFFF) << (shift); \
107 slices[1] assign_op (uintN_t)((i1 >> 0) & 0xFFFF) << (shift); \
108 slices[3] assign_op (uintN_t)((i1 >> 16) & 0xFFFF) << (shift); \
109 slices[5] assign_op (uintN_t)((i1 >> 32) & 0xFFFF) << (shift); \
110 slices[7] assign_op (uintN_t)((i1 >> 48) & 0xFFFF) << (shift); \
113 #define FROM_BITSLICES(bytes, slices, shift) do \
115 uint64_t i1 = ((slices[7] >> (shift)) & 0xFFFF); \
116 i1 = (i1 << 16) | ((slices[5] >> (shift)) & 0xFFFF); \
117 i1 = (i1 << 16) | ((slices[3] >> (shift)) & 0xFFFF); \
118 i1 = (i1 << 16) | ((slices[1] >> (shift)) & 0xFFFF); \
119 uint64_t i0 = ((slices[6] >> (shift)) & 0xFFFF); \
120 i0 = (i0 << 16) | ((slices[4] >> (shift)) & 0xFFFF); \
121 i0 = (i0 << 16) | ((slices[2] >> (shift)) & 0xFFFF); \
122 i0 = (i0 << 16) | ((slices[0] >> (shift)) & 0xFFFF); \
124 SWAPINWORD(i0, 32, 4); \
125 SWAPINWORD(i1, 32, 4); \
126 SWAPINWORD(i0, 16, 2); \
127 SWAPINWORD(i1, 16, 2); \
128 SWAPINWORD(i0, 8, 1); \
129 SWAPINWORD(i1, 8, 1); \
130 PUT_64BIT_LSB_FIRST(bytes, i0); \
131 PUT_64BIT_LSB_FIRST((bytes) + 8, i1); \
135 * Some macros that will be useful repeatedly.
138 /* Iterate a unary transformation over all 8 slices. */
139 #define ITERATE(MACRO, output, input, uintN_t) do \
141 MACRO(output[0], input[0], uintN_t); \
142 MACRO(output[1], input[1], uintN_t); \
143 MACRO(output[2], input[2], uintN_t); \
144 MACRO(output[3], input[3], uintN_t); \
145 MACRO(output[4], input[4], uintN_t); \
146 MACRO(output[5], input[5], uintN_t); \
147 MACRO(output[6], input[6], uintN_t); \
148 MACRO(output[7], input[7], uintN_t); \
151 /* Simply add (i.e. XOR) two whole sets of slices together. */
152 #define BITSLICED_ADD(output, lhs, rhs) do \
154 output[0] = lhs[0] ^ rhs[0]; \
155 output[1] = lhs[1] ^ rhs[1]; \
156 output[2] = lhs[2] ^ rhs[2]; \
157 output[3] = lhs[3] ^ rhs[3]; \
158 output[4] = lhs[4] ^ rhs[4]; \
159 output[5] = lhs[5] ^ rhs[5]; \
160 output[6] = lhs[6] ^ rhs[6]; \
161 output[7] = lhs[7] ^ rhs[7]; \
165 * The AES S-box, in pure bitwise logic so that it can be run in
166 * parallel on whole words full of bit-sliced field elements.
168 * Source: 'A new combinational logic minimization technique with
169 * applications to cryptology', https://eprint.iacr.org/2009/191
171 * As a minor speed optimisation, I use a modified version of the
172 * S-box which omits the additive constant 0x63, i.e. this S-box
173 * consists of only the field inversion and linear map components.
174 * Instead, the addition of the constant is deferred until after the
175 * subsequent ShiftRows and MixColumns stages, so that it happens at
176 * the same time as adding the next round key - and then we just make
177 * it _part_ of the round key, so it doesn't cost any extra
178 * instructions to add.
180 * (Obviously adding a constant to each byte commutes with ShiftRows,
181 * which only permutes the bytes. It also commutes with MixColumns:
182 * that's not quite so obvious, but since the effect of MixColumns is
183 * to multiply a constant polynomial M into each column, it is obvious
184 * that adding some polynomial K and then multiplying by M is
185 * equivalent to multiplying by M and then adding the product KM. And
186 * in fact, since the coefficients of M happen to sum to 1, it turns
187 * out that KM = K, so we don't even have to change the constant when
188 * we move it to the far side of MixColumns.)
190 * Of course, one knock-on effect of this is that the use of the S-box
191 * *during* key setup has to be corrected by manually adding on the
192 * constant afterwards!
195 /* Initial linear transformation for the forward S-box, from Fig 2 of
197 #define SBOX_FORWARD_TOP_TRANSFORM(input, uintN_t) \
198 uintN_t y14 = input[4] ^ input[2]; \
199 uintN_t y13 = input[7] ^ input[1]; \
200 uintN_t y9 = input[7] ^ input[4]; \
201 uintN_t y8 = input[7] ^ input[2]; \
202 uintN_t t0 = input[6] ^ input[5]; \
203 uintN_t y1 = t0 ^ input[0]; \
204 uintN_t y4 = y1 ^ input[4]; \
205 uintN_t y12 = y13 ^ y14; \
206 uintN_t y2 = y1 ^ input[7]; \
207 uintN_t y5 = y1 ^ input[1]; \
208 uintN_t y3 = y5 ^ y8; \
209 uintN_t t1 = input[3] ^ y12; \
210 uintN_t y15 = t1 ^ input[2]; \
211 uintN_t y20 = t1 ^ input[6]; \
212 uintN_t y6 = y15 ^ input[0]; \
213 uintN_t y10 = y15 ^ t0; \
214 uintN_t y11 = y20 ^ y9; \
215 uintN_t y7 = input[0] ^ y11; \
216 uintN_t y17 = y10 ^ y11; \
217 uintN_t y19 = y10 ^ y8; \
218 uintN_t y16 = t0 ^ y11; \
219 uintN_t y21 = y13 ^ y16; \
220 uintN_t y18 = input[7] ^ y16; \
221 /* Make a copy of input[0] under a new name, because the core
222 * will refer to it, and in the inverse version of the S-box
223 * the corresponding value will be one of the calculated ones
224 * and not in input[0] itself. */ \
225 uintN_t i0 = input[0]; \
228 /* Core nonlinear component, from Fig 3 of the paper. */
229 #define SBOX_CORE(uintN_t) \
230 uintN_t t2 = y12 & y15; \
231 uintN_t t3 = y3 & y6; \
232 uintN_t t4 = t3 ^ t2; \
233 uintN_t t5 = y4 & i0; \
234 uintN_t t6 = t5 ^ t2; \
235 uintN_t t7 = y13 & y16; \
236 uintN_t t8 = y5 & y1; \
237 uintN_t t9 = t8 ^ t7; \
238 uintN_t t10 = y2 & y7; \
239 uintN_t t11 = t10 ^ t7; \
240 uintN_t t12 = y9 & y11; \
241 uintN_t t13 = y14 & y17; \
242 uintN_t t14 = t13 ^ t12; \
243 uintN_t t15 = y8 & y10; \
244 uintN_t t16 = t15 ^ t12; \
245 uintN_t t17 = t4 ^ t14; \
246 uintN_t t18 = t6 ^ t16; \
247 uintN_t t19 = t9 ^ t14; \
248 uintN_t t20 = t11 ^ t16; \
249 uintN_t t21 = t17 ^ y20; \
250 uintN_t t22 = t18 ^ y19; \
251 uintN_t t23 = t19 ^ y21; \
252 uintN_t t24 = t20 ^ y18; \
253 uintN_t t25 = t21 ^ t22; \
254 uintN_t t26 = t21 & t23; \
255 uintN_t t27 = t24 ^ t26; \
256 uintN_t t28 = t25 & t27; \
257 uintN_t t29 = t28 ^ t22; \
258 uintN_t t30 = t23 ^ t24; \
259 uintN_t t31 = t22 ^ t26; \
260 uintN_t t32 = t31 & t30; \
261 uintN_t t33 = t32 ^ t24; \
262 uintN_t t34 = t23 ^ t33; \
263 uintN_t t35 = t27 ^ t33; \
264 uintN_t t36 = t24 & t35; \
265 uintN_t t37 = t36 ^ t34; \
266 uintN_t t38 = t27 ^ t36; \
267 uintN_t t39 = t29 & t38; \
268 uintN_t t40 = t25 ^ t39; \
269 uintN_t t41 = t40 ^ t37; \
270 uintN_t t42 = t29 ^ t33; \
271 uintN_t t43 = t29 ^ t40; \
272 uintN_t t44 = t33 ^ t37; \
273 uintN_t t45 = t42 ^ t41; \
274 uintN_t z0 = t44 & y15; \
275 uintN_t z1 = t37 & y6; \
276 uintN_t z2 = t33 & i0; \
277 uintN_t z3 = t43 & y16; \
278 uintN_t z4 = t40 & y1; \
279 uintN_t z5 = t29 & y7; \
280 uintN_t z6 = t42 & y11; \
281 uintN_t z7 = t45 & y17; \
282 uintN_t z8 = t41 & y10; \
283 uintN_t z9 = t44 & y12; \
284 uintN_t z10 = t37 & y3; \
285 uintN_t z11 = t33 & y4; \
286 uintN_t z12 = t43 & y13; \
287 uintN_t z13 = t40 & y5; \
288 uintN_t z14 = t29 & y2; \
289 uintN_t z15 = t42 & y9; \
290 uintN_t z16 = t45 & y14; \
291 uintN_t z17 = t41 & y8; \
294 /* Final linear transformation for the forward S-box, from Fig 4 of
296 #define SBOX_FORWARD_BOTTOM_TRANSFORM(output, uintN_t) \
297 uintN_t t46 = z15 ^ z16; \
298 uintN_t t47 = z10 ^ z11; \
299 uintN_t t48 = z5 ^ z13; \
300 uintN_t t49 = z9 ^ z10; \
301 uintN_t t50 = z2 ^ z12; \
302 uintN_t t51 = z2 ^ z5; \
303 uintN_t t52 = z7 ^ z8; \
304 uintN_t t53 = z0 ^ z3; \
305 uintN_t t54 = z6 ^ z7; \
306 uintN_t t55 = z16 ^ z17; \
307 uintN_t t56 = z12 ^ t48; \
308 uintN_t t57 = t50 ^ t53; \
309 uintN_t t58 = z4 ^ t46; \
310 uintN_t t59 = z3 ^ t54; \
311 uintN_t t60 = t46 ^ t57; \
312 uintN_t t61 = z14 ^ t57; \
313 uintN_t t62 = t52 ^ t58; \
314 uintN_t t63 = t49 ^ t58; \
315 uintN_t t64 = z4 ^ t59; \
316 uintN_t t65 = t61 ^ t62; \
317 uintN_t t66 = z1 ^ t63; \
318 output[7] = t59 ^ t63; \
319 output[1] = t56 ^ t62; \
320 output[0] = t48 ^ t60; \
321 uintN_t t67 = t64 ^ t65; \
322 output[4] = t53 ^ t66; \
323 output[3] = t51 ^ t66; \
324 output[2] = t47 ^ t65; \
325 output[6] = t64 ^ output[4]; \
326 output[5] = t55 ^ t67; \
329 #define BITSLICED_SUBBYTES(output, input, uintN_t) do { \
330 SBOX_FORWARD_TOP_TRANSFORM(input, uintN_t); \
331 SBOX_CORE(uintN_t); \
332 SBOX_FORWARD_BOTTOM_TRANSFORM(output, uintN_t); \
336 * Initial and final linear transformations for the backward S-box. I
337 * generated these myself, by implementing the linear-transform
338 * optimisation algorithm in the paper, and applying it to the
339 * matrices calculated by _their_ top and bottom transformations, pre-
340 * and post-multiplied as appropriate by the linear map in the inverse
343 #define SBOX_BACKWARD_TOP_TRANSFORM(input, uintN_t) \
344 uintN_t y5 = input[4] ^ input[6]; \
345 uintN_t y19 = input[3] ^ input[0]; \
346 uintN_t itmp8 = y5 ^ input[0]; \
347 uintN_t y4 = itmp8 ^ input[1]; \
348 uintN_t y9 = input[4] ^ input[3]; \
349 uintN_t y2 = y9 ^ y4; \
350 uintN_t itmp9 = y2 ^ input[7]; \
351 uintN_t y1 = y9 ^ input[0]; \
352 uintN_t y6 = y5 ^ input[7]; \
353 uintN_t y18 = y9 ^ input[5]; \
354 uintN_t y7 = y18 ^ y2; \
355 uintN_t y16 = y7 ^ y1; \
356 uintN_t y21 = y7 ^ input[1]; \
357 uintN_t y3 = input[4] ^ input[7]; \
358 uintN_t y13 = y16 ^ y21; \
359 uintN_t y8 = input[4] ^ y6; \
360 uintN_t y10 = y8 ^ y19; \
361 uintN_t y14 = y8 ^ y9; \
362 uintN_t y20 = itmp9 ^ input[2]; \
363 uintN_t y11 = y9 ^ y20; \
364 uintN_t i0 = y11 ^ y7; \
365 uintN_t y15 = i0 ^ y6; \
366 uintN_t y17 = y16 ^ y15; \
367 uintN_t y12 = itmp9 ^ input[3]; \
369 #define SBOX_BACKWARD_BOTTOM_TRANSFORM(output, uintN_t) \
370 uintN_t otmp18 = z15 ^ z6; \
371 uintN_t otmp19 = z13 ^ otmp18; \
372 uintN_t otmp20 = z12 ^ otmp19; \
373 uintN_t otmp21 = z16 ^ otmp20; \
374 uintN_t otmp22 = z8 ^ otmp21; \
375 uintN_t otmp23 = z0 ^ otmp22; \
376 uintN_t otmp24 = otmp22 ^ z3; \
377 uintN_t otmp25 = otmp24 ^ z4; \
378 uintN_t otmp26 = otmp25 ^ z2; \
379 uintN_t otmp27 = z1 ^ otmp26; \
380 uintN_t otmp28 = z14 ^ otmp27; \
381 uintN_t otmp29 = otmp28 ^ z10; \
382 output[4] = z2 ^ otmp23; \
383 output[7] = z5 ^ otmp24; \
384 uintN_t otmp30 = z11 ^ otmp29; \
385 output[5] = z13 ^ otmp30; \
386 uintN_t otmp31 = otmp25 ^ z8; \
387 output[1] = z7 ^ otmp31; \
388 uintN_t otmp32 = z11 ^ z9; \
389 uintN_t otmp33 = z17 ^ otmp32; \
390 uintN_t otmp34 = otmp30 ^ otmp33; \
391 output[0] = z15 ^ otmp33; \
392 uintN_t otmp35 = z12 ^ otmp34; \
393 output[6] = otmp35 ^ z16; \
394 uintN_t otmp36 = z1 ^ otmp23; \
395 uintN_t otmp37 = z5 ^ otmp36; \
396 output[2] = z4 ^ otmp37; \
397 uintN_t otmp38 = z11 ^ output[1]; \
398 uintN_t otmp39 = z2 ^ otmp38; \
399 uintN_t otmp40 = z17 ^ otmp39; \
400 uintN_t otmp41 = z0 ^ otmp40; \
401 uintN_t otmp42 = z5 ^ otmp41; \
402 uintN_t otmp43 = otmp42 ^ z10; \
403 uintN_t otmp44 = otmp43 ^ z3; \
404 output[3] = otmp44 ^ z16; \
407 #define BITSLICED_INVSUBBYTES(output, input, uintN_t) do { \
408 SBOX_BACKWARD_TOP_TRANSFORM(input, uintN_t); \
409 SBOX_CORE(uintN_t); \
410 SBOX_BACKWARD_BOTTOM_TRANSFORM(output, uintN_t); \
415 * The ShiftRows transformation. This operates independently on each
419 #define SINGLE_BITSLICE_SHIFTROWS(output, input, uintN_t) do \
421 uintN_t mask, mask2, mask3, diff, x = (input); \
422 /* Rotate rows 2 and 3 by 16 bits */ \
423 mask = 0x00CC * (((uintN_t)~(uintN_t)0) / 0xFFFF); \
424 diff = ((x >> 8) ^ x) & mask; \
425 x ^= diff ^ (diff << 8); \
426 /* Rotate rows 1 and 3 by 8 bits */ \
427 mask = 0x0AAA * (((uintN_t)~(uintN_t)0) / 0xFFFF); \
428 mask2 = 0xA000 * (((uintN_t)~(uintN_t)0) / 0xFFFF); \
429 mask3 = 0x5555 * (((uintN_t)~(uintN_t)0) / 0xFFFF); \
430 x = ((x >> 4) & mask) | ((x << 12) & mask2) | (x & mask3); \
435 #define SINGLE_BITSLICE_INVSHIFTROWS(output, input, uintN_t) do \
437 uintN_t mask, mask2, mask3, diff, x = (input); \
438 /* Rotate rows 2 and 3 by 16 bits */ \
439 mask = 0x00CC * (((uintN_t)~(uintN_t)0) / 0xFFFF); \
440 diff = ((x >> 8) ^ x) & mask; \
441 x ^= diff ^ (diff << 8); \
442 /* Rotate rows 1 and 3 by 8 bits, the opposite way to ShiftRows */ \
443 mask = 0x000A * (((uintN_t)~(uintN_t)0) / 0xFFFF); \
444 mask2 = 0xAAA0 * (((uintN_t)~(uintN_t)0) / 0xFFFF); \
445 mask3 = 0x5555 * (((uintN_t)~(uintN_t)0) / 0xFFFF); \
446 x = ((x >> 12) & mask) | ((x << 4) & mask2) | (x & mask3); \
451 #define BITSLICED_SHIFTROWS(output, input, uintN_t) do \
453 ITERATE(SINGLE_BITSLICE_SHIFTROWS, output, input, uintN_t); \
456 #define BITSLICED_INVSHIFTROWS(output, input, uintN_t) do \
458 ITERATE(SINGLE_BITSLICE_INVSHIFTROWS, output, input, uintN_t); \
462 * The MixColumns transformation. This has to operate on all eight bit
463 * slices at once, and also passes data back and forth between the
464 * bits in an adjacent group of 4 within each slice.
466 * Notation: let F = GF(2)[X]/<X^8+X^4+X^3+X+1> be the finite field
467 * used in AES, and let R = F[Y]/<Y^4+1> be the ring whose elements
468 * represent the possible contents of a column of the matrix. I use X
469 * and Y below in those senses, i.e. X is the value in F that
470 * represents the byte 0x02, and Y is the value in R that cycles the
471 * four bytes around by one if you multiply by it.
474 /* Multiply every column by Y^3, i.e. cycle it round one place to the
475 * right. Operates on one bit slice at a time; you have to wrap it in
476 * ITERATE to affect all the data at once. */
477 #define BITSLICED_MUL_BY_Y3(output, input, uintN_t) do \
479 uintN_t mask, mask2, x; \
480 mask = 0x8 * (((uintN_t)~(uintN_t)0) / 0xF); \
481 mask2 = 0x7 * (((uintN_t)~(uintN_t)0) / 0xF); \
483 output = ((x << 3) & mask) ^ ((x >> 1) & mask2); \
486 /* Multiply every column by Y^2. */
487 #define BITSLICED_MUL_BY_Y2(output, input, uintN_t) do \
489 uintN_t mask, mask2, x; \
490 mask = 0xC * (((uintN_t)~(uintN_t)0) / 0xF); \
491 mask2 = 0x3 * (((uintN_t)~(uintN_t)0) / 0xF); \
493 output = ((x << 2) & mask) ^ ((x >> 2) & mask2); \
496 #define BITSLICED_MUL_BY_1_Y3(output, input, uintN_t) do \
498 uintN_t tmp = input; \
499 BITSLICED_MUL_BY_Y3(tmp, input, uintN_t); \
500 output = input ^ tmp; \
503 /* Multiply every column by 1+Y^2. */
504 #define BITSLICED_MUL_BY_1_Y2(output, input, uintN_t) do \
506 uintN_t tmp = input; \
507 BITSLICED_MUL_BY_Y2(tmp, input, uintN_t); \
508 output = input ^ tmp; \
511 /* Multiply every field element by X. This has to feed data between
512 * slices, so it does the whole job in one go without needing ITERATE. */
513 #define BITSLICED_MUL_BY_X(output, input, uintN_t) do \
515 uintN_t bit7 = input[7]; \
516 output[7] = input[6]; \
517 output[6] = input[5]; \
518 output[5] = input[4]; \
519 output[4] = input[3] ^ bit7; \
520 output[3] = input[2] ^ bit7; \
521 output[2] = input[1]; \
522 output[1] = input[0] ^ bit7; \
527 * The MixColumns constant is
528 * M = X + Y + Y^2 + (X+1)Y^3
529 * which we construct by rearranging it into
530 * M = 1 + (1+Y^3) [ X + (1+Y^2) ]
532 #define BITSLICED_MIXCOLUMNS(output, input, uintN_t) do \
534 uintN_t a[8], aX[8], b[8]; \
535 /* a = input * (1+Y^3) */ \
536 ITERATE(BITSLICED_MUL_BY_1_Y3, a, input, uintN_t); \
538 BITSLICED_MUL_BY_X(aX, a, uintN_t); \
539 /* b = a * (1+Y^2) = input * (1+Y+Y^2+Y^3) */ \
540 ITERATE(BITSLICED_MUL_BY_1_Y2, b, a, uintN_t); \
541 /* output = input + aX + b (reusing a as a temp */ \
542 BITSLICED_ADD(a, aX, b); \
543 BITSLICED_ADD(output, input, a); \
547 * The InvMixColumns constant, written out longhand, is
548 * I = (X^3+X^2+X) + (X^3+1)Y + (X^3+X^2+1)Y^2 + (X^3+X+1)Y^3
549 * We represent this as
550 * I = (X^3+X^2+X+1)(Y^3+Y^2+Y+1) + 1 + X(Y+Y^2) + X^2(Y+Y^3)
552 #define BITSLICED_INVMIXCOLUMNS(output, input, uintN_t) do \
554 /* We need input * X^i for i=1,...,3 */ \
555 uintN_t X[8], X2[8], X3[8]; \
556 BITSLICED_MUL_BY_X(X, input, uintN_t); \
557 BITSLICED_MUL_BY_X(X2, X, uintN_t); \
558 BITSLICED_MUL_BY_X(X3, X2, uintN_t); \
559 /* Sum them all and multiply by 1+Y+Y^2+Y^3. */ \
561 BITSLICED_ADD(S, input, X); \
562 BITSLICED_ADD(S, S, X2); \
563 BITSLICED_ADD(S, S, X3); \
564 ITERATE(BITSLICED_MUL_BY_1_Y3, S, S, uintN_t); \
565 ITERATE(BITSLICED_MUL_BY_1_Y2, S, S, uintN_t); \
566 /* Compute the X(Y+Y^2) term. */ \
568 ITERATE(BITSLICED_MUL_BY_1_Y3, A, X, uintN_t); \
569 ITERATE(BITSLICED_MUL_BY_Y2, A, A, uintN_t); \
570 /* Compute the X^2(Y+Y^3) term. */ \
572 ITERATE(BITSLICED_MUL_BY_1_Y2, B, X2, uintN_t); \
573 ITERATE(BITSLICED_MUL_BY_Y3, B, B, uintN_t); \
574 /* And add all the pieces together. */ \
575 BITSLICED_ADD(S, S, input); \
576 BITSLICED_ADD(S, S, A); \
577 BITSLICED_ADD(output, S, B); \
581 * Put it all together into a cipher round.
584 /* Dummy macro to get rid of the MixColumns in the final round. */
585 #define NO_MIXCOLUMNS(out, in, uintN_t) do {} while (0)
587 #define ENCRYPT_ROUND_FN(suffix, uintN_t, mixcol_macro) \
588 static void aes_sliced_round_e_##suffix( \
589 uintN_t output[8], const uintN_t input[8], const uintN_t roundkey[8]) \
591 BITSLICED_SUBBYTES(output, input, uintN_t); \
592 BITSLICED_SHIFTROWS(output, output, uintN_t); \
593 mixcol_macro(output, output, uintN_t); \
594 BITSLICED_ADD(output, output, roundkey); \
597 ENCRYPT_ROUND_FN(serial
, uint16_t, BITSLICED_MIXCOLUMNS
)
598 ENCRYPT_ROUND_FN(serial_last
, uint16_t, NO_MIXCOLUMNS
)
599 ENCRYPT_ROUND_FN(parallel
, BignumInt
, BITSLICED_MIXCOLUMNS
)
600 ENCRYPT_ROUND_FN(parallel_last
, BignumInt
, NO_MIXCOLUMNS
)
602 #define DECRYPT_ROUND_FN(suffix, uintN_t, mixcol_macro) \
603 static void aes_sliced_round_d_##suffix( \
604 uintN_t output[8], const uintN_t input[8], const uintN_t roundkey[8]) \
606 BITSLICED_ADD(output, input, roundkey); \
607 mixcol_macro(output, output, uintN_t); \
608 BITSLICED_INVSUBBYTES(output, output, uintN_t); \
609 BITSLICED_INVSHIFTROWS(output, output, uintN_t); \
612 #if 0 /* no cipher mode we support requires serial decryption */
613 DECRYPT_ROUND_FN(serial
, uint16_t, BITSLICED_INVMIXCOLUMNS
)
614 DECRYPT_ROUND_FN(serial_first
, uint16_t, NO_MIXCOLUMNS
)
616 DECRYPT_ROUND_FN(parallel
, BignumInt
, BITSLICED_INVMIXCOLUMNS
)
617 DECRYPT_ROUND_FN(parallel_first
, BignumInt
, NO_MIXCOLUMNS
)
620 * Key setup function.
623 typedef struct aes_sliced_key aes_sliced_key
;
624 struct aes_sliced_key
{
625 BignumInt roundkeys_parallel
[MAXROUNDKEYS
* 8];
626 uint16_t roundkeys_serial
[MAXROUNDKEYS
* 8];
630 static void aes_sliced_key_setup(
631 aes_sliced_key
*sk
, const void *vkey
, size_t keybits
)
633 const unsigned char *key
= (const unsigned char *)vkey
;
635 size_t key_words
= keybits
/ 32;
636 sk
->rounds
= key_words
+ 6;
637 size_t sched_words
= (sk
->rounds
+ 1) * 4;
639 unsigned rconpos
= 0;
641 uint16_t *outslices
= sk
->roundkeys_serial
;
642 unsigned outshift
= 0;
644 memset(sk
->roundkeys_serial
, 0, sizeof(sk
->roundkeys_serial
));
647 memset(inblk
, 0, 16);
650 for (size_t i
= 0; i
< sched_words
; i
++) {
652 * Prepare a word of round key in the low 4 bits of each
653 * integer in slices[].
656 memcpy(inblk
, key
+ 4*i
, 4);
657 TO_BITSLICES(slices
, inblk
, uint16_t, =, 0);
659 unsigned wordindex
, bitshift
;
660 uint16_t *prevslices
;
662 /* Fetch the (i-1)th key word */
664 bitshift
= 4 * (wordindex
& 3);
665 prevslices
= sk
->roundkeys_serial
+ 8 * (wordindex
>> 2);
666 for (size_t i
= 0; i
< 8; i
++)
667 slices
[i
] = prevslices
[i
] >> bitshift
;
669 /* Decide what we're doing in this expansion stage */
670 bool rotate_and_round_constant
= (i
% key_words
== 0);
671 bool sub
= rotate_and_round_constant
||
672 (key_words
== 8 && i
% 8 == 4);
674 if (rotate_and_round_constant
) {
675 for (size_t i
= 0; i
< 8; i
++)
676 slices
[i
] = ((slices
[i
] << 3) | (slices
[i
] >> 1)) & 0xF;
680 /* Apply the SubBytes transform to the key word. But
681 * here we need to apply the _full_ SubBytes from the
682 * spec, including the constant which our S-box leaves
684 BITSLICED_SUBBYTES(slices
, slices
, uint16_t);
691 if (rotate_and_round_constant
) {
692 assert(rconpos
< lenof(aes_key_setup_round_constants
));
693 uint8_t rcon
= aes_key_setup_round_constants
[rconpos
++];
694 for (size_t i
= 0; i
< 8; i
++)
695 slices
[i
] ^= 1 & (rcon
>> i
);
698 /* Combine with the (i-Nk)th key word */
699 wordindex
= i
- key_words
;
700 bitshift
= 4 * (wordindex
& 3);
701 prevslices
= sk
->roundkeys_serial
+ 8 * (wordindex
>> 2);
702 for (size_t i
= 0; i
< 8; i
++)
703 slices
[i
] ^= prevslices
[i
] >> bitshift
;
707 * Now copy it into sk.
709 for (unsigned b
= 0; b
< 8; b
++)
710 outslices
[b
] |= (slices
[b
] & 0xF) << outshift
;
712 if (outshift
== 16) {
718 smemclr(inblk
, sizeof(inblk
));
719 smemclr(slices
, sizeof(slices
));
722 * Add the S-box constant to every round key after the first one,
723 * compensating for it being left out in the main cipher.
725 for (size_t i
= 8; i
< 8 * (sched_words
/4); i
+= 8) {
726 sk
->roundkeys_serial
[i
+0] ^= 0xFFFF;
727 sk
->roundkeys_serial
[i
+1] ^= 0xFFFF;
728 sk
->roundkeys_serial
[i
+5] ^= 0xFFFF;
729 sk
->roundkeys_serial
[i
+6] ^= 0xFFFF;
733 * Replicate that set of round keys into larger integers for the
734 * parallel versions of the cipher.
736 for (size_t i
= 0; i
< 8 * (sched_words
/ 4); i
++) {
737 sk
->roundkeys_parallel
[i
] = sk
->roundkeys_serial
[i
] *
738 ((BignumInt
)~(BignumInt
)0 / 0xFFFF);
743 * The full cipher primitive, including transforming the input and
744 * output to/from bit-sliced form.
747 #define ENCRYPT_FN(suffix, uintN_t, nblocks) \
748 static void aes_sliced_e_##suffix( \
749 uint8_t *output, const uint8_t *input, const aes_sliced_key *sk) \
752 TO_BITSLICES(state, input, uintN_t, =, 0); \
753 for (unsigned i = 1; i < nblocks; i++) { \
755 TO_BITSLICES(state, input, uintN_t, |=, i*16); \
757 const uintN_t *keys = sk->roundkeys_##suffix; \
758 BITSLICED_ADD(state, state, keys); \
760 for (unsigned i = 0; i < sk->rounds-1; i++) { \
761 aes_sliced_round_e_##suffix(state, state, keys); \
764 aes_sliced_round_e_##suffix##_last(state, state, keys); \
765 for (unsigned i = 0; i < nblocks; i++) { \
766 FROM_BITSLICES(output, state, i*16); \
771 #define DECRYPT_FN(suffix, uintN_t, nblocks) \
772 static void aes_sliced_d_##suffix( \
773 uint8_t *output, const uint8_t *input, const aes_sliced_key *sk) \
776 TO_BITSLICES(state, input, uintN_t, =, 0); \
777 for (unsigned i = 1; i < nblocks; i++) { \
779 TO_BITSLICES(state, input, uintN_t, |=, i*16); \
781 const uintN_t *keys = sk->roundkeys_##suffix + 8*sk->rounds; \
782 aes_sliced_round_d_##suffix##_first(state, state, keys); \
784 for (unsigned i = 0; i < sk->rounds-1; i++) { \
785 aes_sliced_round_d_##suffix(state, state, keys); \
788 BITSLICED_ADD(state, state, keys); \
789 for (unsigned i = 0; i < nblocks; i++) { \
790 FROM_BITSLICES(output, state, i*16); \
795 ENCRYPT_FN(serial
, uint16_t, 1)
796 #if 0 /* no cipher mode we support requires serial decryption */
797 DECRYPT_FN(serial
, uint16_t, 1)
799 ENCRYPT_FN(parallel
, BignumInt
, SLICE_PARALLELISM
)
800 DECRYPT_FN(parallel
, BignumInt
, SLICE_PARALLELISM
)
803 * The SSH interface and the cipher modes.
806 #define SDCTR_WORDS (16 / BIGNUM_INT_BYTES)
808 typedef struct aes_sw_context aes_sw_context
;
809 struct aes_sw_context
{
813 /* In CBC mode, the IV is just a copy of the last seen
818 /* In SDCTR mode, we keep the counter itself in a form
819 * that's easy to increment. We also use the parallel
820 * version of the core AES function, so we'll encrypt
821 * multiple counter values in one go. That won't align
822 * nicely with the sizes of data we're asked to encrypt,
823 * so we must also store a cache of the last set of
824 * keystream blocks we generated, and our current position
825 * within that cache. */
826 BignumInt counter
[SDCTR_WORDS
];
827 uint8_t keystream
[SLICE_PARALLELISM
* 16];
828 uint8_t *keystream_pos
;
831 /* In GCM mode, the cipher preimage consists of three
832 * sections: one fixed, one that increments per message
833 * sent and MACed, and one that increments per cipher
835 uint64_t msg_counter
;
836 uint32_t fixed_iv
, block_counter
;
837 /* But we keep the precomputed keystream chunks just like
839 uint8_t keystream
[SLICE_PARALLELISM
* 16];
840 uint8_t *keystream_pos
;
846 static ssh_cipher
*aes_sw_new(const ssh_cipheralg
*alg
)
848 aes_sw_context
*ctx
= snew(aes_sw_context
);
853 static void aes_sw_free(ssh_cipher
*ciph
)
855 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
856 smemclr(ctx
, sizeof(*ctx
));
860 static void aes_sw_setkey(ssh_cipher
*ciph
, const void *vkey
)
862 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
863 aes_sliced_key_setup(&ctx
->sk
, vkey
, ctx
->ciph
.vt
->real_keybits
);
866 static void aes_sw_setiv_cbc(ssh_cipher
*ciph
, const void *iv
)
868 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
869 memcpy(ctx
->iv
.cbc
.prevblk
, iv
, 16);
872 static void aes_sw_setiv_sdctr(ssh_cipher
*ciph
, const void *viv
)
874 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
875 const uint8_t *iv
= (const uint8_t *)viv
;
877 /* Import the initial counter value into the internal representation */
878 for (unsigned i
= 0; i
< SDCTR_WORDS
; i
++)
879 ctx
->iv
.sdctr
.counter
[i
] =
880 GET_BIGNUMINT_MSB_FIRST(
881 iv
+ 16 - BIGNUM_INT_BYTES
- i
*BIGNUM_INT_BYTES
);
883 /* Set keystream_pos to indicate that the keystream cache is
885 ctx
->iv
.sdctr
.keystream_pos
=
886 ctx
->iv
.sdctr
.keystream
+ sizeof(ctx
->iv
.sdctr
.keystream
);
889 static void aes_sw_setiv_gcm(ssh_cipher
*ciph
, const void *viv
)
891 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
892 const uint8_t *iv
= (const uint8_t *)viv
;
894 ctx
->iv
.gcm
.fixed_iv
= GET_32BIT_MSB_FIRST(iv
);
895 ctx
->iv
.gcm
.msg_counter
= GET_64BIT_MSB_FIRST(iv
+ 4);
896 ctx
->iv
.gcm
.block_counter
= 1;
898 /* Set keystream_pos to indicate that the keystream cache is
900 ctx
->iv
.gcm
.keystream_pos
=
901 ctx
->iv
.gcm
.keystream
+ sizeof(ctx
->iv
.gcm
.keystream
);
904 static void aes_sw_next_message_gcm(ssh_cipher
*ciph
)
906 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
908 ctx
->iv
.gcm
.msg_counter
++;
909 ctx
->iv
.gcm
.block_counter
= 1;
910 ctx
->iv
.gcm
.keystream_pos
=
911 ctx
->iv
.gcm
.keystream
+ sizeof(ctx
->iv
.gcm
.keystream
);
914 typedef void (*aes_sw_fn
)(uint32_t v
[4], const uint32_t *keysched
);
916 static inline void memxor16(void *vout
, const void *vlhs
, const void *vrhs
)
918 uint8_t *out
= (uint8_t *)vout
;
919 const uint8_t *lhs
= (const uint8_t *)vlhs
, *rhs
= (const uint8_t *)vrhs
;
922 w
= GET_64BIT_LSB_FIRST(lhs
);
923 w
^= GET_64BIT_LSB_FIRST(rhs
);
924 PUT_64BIT_LSB_FIRST(out
, w
);
925 w
= GET_64BIT_LSB_FIRST(lhs
+ 8);
926 w
^= GET_64BIT_LSB_FIRST(rhs
+ 8);
927 PUT_64BIT_LSB_FIRST(out
+ 8, w
);
930 static inline void aes_cbc_sw_encrypt(
931 ssh_cipher
*ciph
, void *vblk
, int blklen
)
933 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
936 * CBC encryption has to be done serially, because the input to
937 * each run of the cipher includes the output from the previous
941 for (uint8_t *blk
= (uint8_t *)vblk
, *finish
= blk
+ blklen
;
942 blk
< finish
; blk
+= 16) {
944 * We use the IV array itself as the location for the
945 * encryption, because there's no reason not to.
948 /* XOR the new plaintext block into the previous cipher block */
949 memxor16(ctx
->iv
.cbc
.prevblk
, ctx
->iv
.cbc
.prevblk
, blk
);
951 /* Run the cipher over the result, which leaves it
952 * conveniently already stored in ctx->iv */
954 ctx
->iv
.cbc
.prevblk
, ctx
->iv
.cbc
.prevblk
, &ctx
->sk
);
956 /* Copy it to the output location */
957 memcpy(blk
, ctx
->iv
.cbc
.prevblk
, 16);
961 static inline void aes_cbc_sw_decrypt(
962 ssh_cipher
*ciph
, void *vblk
, int blklen
)
964 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
965 uint8_t *blk
= (uint8_t *)vblk
;
968 * CBC decryption can run in parallel, because all the
969 * _ciphertext_ blocks are already available.
972 size_t blocks_remaining
= blklen
/ 16;
974 uint8_t data
[SLICE_PARALLELISM
* 16];
975 /* Zeroing the data array is probably overcautious, but it avoids
976 * technically undefined behaviour from leaving it uninitialised
977 * if our very first iteration doesn't include enough cipher
978 * blocks to populate it fully */
979 memset(data
, 0, sizeof(data
));
981 while (blocks_remaining
> 0) {
982 /* Number of blocks we'll handle in this iteration. If we're
983 * dealing with fewer than the maximum, it doesn't matter -
984 * it's harmless to run the full parallel cipher function
986 size_t blocks
= (blocks_remaining
< SLICE_PARALLELISM
?
987 blocks_remaining
: SLICE_PARALLELISM
);
989 /* Parallel-decrypt the input, in a separate array so we still
990 * have the cipher stream available for XORing. */
991 memcpy(data
, blk
, 16 * blocks
);
992 aes_sliced_d_parallel(data
, data
, &ctx
->sk
);
994 /* Write the output and update the IV */
995 for (size_t i
= 0; i
< blocks
; i
++) {
996 uint8_t *decrypted
= data
+ 16*i
;
997 uint8_t *output
= blk
+ 16*i
;
999 memxor16(decrypted
, decrypted
, ctx
->iv
.cbc
.prevblk
);
1000 memcpy(ctx
->iv
.cbc
.prevblk
, output
, 16);
1001 memcpy(output
, decrypted
, 16);
1004 /* Advance the input pointer. */
1006 blocks_remaining
-= blocks
;
1009 smemclr(data
, sizeof(data
));
1012 static inline void aes_sdctr_sw(
1013 ssh_cipher
*ciph
, void *vblk
, int blklen
)
1015 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
1018 * SDCTR encrypt/decrypt loops round one block at a time XORing
1019 * the keystream into the user's data, and periodically has to run
1020 * a parallel encryption operation to get more keystream.
1023 uint8_t *keystream_end
=
1024 ctx
->iv
.sdctr
.keystream
+ sizeof(ctx
->iv
.sdctr
.keystream
);
1026 for (uint8_t *blk
= (uint8_t *)vblk
, *finish
= blk
+ blklen
;
1027 blk
< finish
; blk
+= 16) {
1029 if (ctx
->iv
.sdctr
.keystream_pos
== keystream_end
) {
1031 * Generate some keystream.
1033 for (uint8_t *block
= ctx
->iv
.sdctr
.keystream
;
1034 block
< keystream_end
; block
+= 16) {
1035 /* Format the counter value into the buffer. */
1036 for (unsigned i
= 0; i
< SDCTR_WORDS
; i
++)
1037 PUT_BIGNUMINT_MSB_FIRST(
1038 block
+ 16 - BIGNUM_INT_BYTES
- i
*BIGNUM_INT_BYTES
,
1039 ctx
->iv
.sdctr
.counter
[i
]);
1041 /* Increment the counter. */
1042 BignumCarry carry
= 1;
1043 for (unsigned i
= 0; i
< SDCTR_WORDS
; i
++)
1044 BignumADC(ctx
->iv
.sdctr
.counter
[i
], carry
,
1045 ctx
->iv
.sdctr
.counter
[i
], 0, carry
);
1048 /* Encrypt all those counter blocks. */
1049 aes_sliced_e_parallel(ctx
->iv
.sdctr
.keystream
,
1050 ctx
->iv
.sdctr
.keystream
, &ctx
->sk
);
1052 /* Reset keystream_pos to the start of the buffer. */
1053 ctx
->iv
.sdctr
.keystream_pos
= ctx
->iv
.sdctr
.keystream
;
1056 memxor16(blk
, blk
, ctx
->iv
.sdctr
.keystream_pos
);
1057 ctx
->iv
.sdctr
.keystream_pos
+= 16;
1061 static inline void aes_encrypt_ecb_block_sw(ssh_cipher
*ciph
, void *blk
)
1063 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
1064 aes_sliced_e_serial(blk
, blk
, &ctx
->sk
);
1067 static inline void aes_gcm_sw(
1068 ssh_cipher
*ciph
, void *vblk
, int blklen
)
1070 aes_sw_context
*ctx
= container_of(ciph
, aes_sw_context
, ciph
);
1073 * GCM encrypt/decrypt looks just like SDCTR, except that the
1074 * method of generating more keystream varies slightly.
1077 uint8_t *keystream_end
=
1078 ctx
->iv
.gcm
.keystream
+ sizeof(ctx
->iv
.gcm
.keystream
);
1080 for (uint8_t *blk
= (uint8_t *)vblk
, *finish
= blk
+ blklen
;
1081 blk
< finish
; blk
+= 16) {
1083 if (ctx
->iv
.gcm
.keystream_pos
== keystream_end
) {
1085 * Generate some keystream.
1087 for (uint8_t *block
= ctx
->iv
.gcm
.keystream
;
1088 block
< keystream_end
; block
+= 16) {
1089 /* Format the counter value into the buffer. */
1090 PUT_32BIT_MSB_FIRST(block
, ctx
->iv
.gcm
.fixed_iv
);
1091 PUT_64BIT_MSB_FIRST(block
+ 4, ctx
->iv
.gcm
.msg_counter
);
1092 PUT_32BIT_MSB_FIRST(block
+ 12, ctx
->iv
.gcm
.block_counter
);
1094 /* Increment the counter. */
1095 ctx
->iv
.gcm
.block_counter
++;
1098 /* Encrypt all those counter blocks. */
1099 aes_sliced_e_parallel(ctx
->iv
.gcm
.keystream
,
1100 ctx
->iv
.gcm
.keystream
, &ctx
->sk
);
1102 /* Reset keystream_pos to the start of the buffer. */
1103 ctx
->iv
.gcm
.keystream_pos
= ctx
->iv
.gcm
.keystream
;
1106 memxor16(blk
, blk
, ctx
->iv
.gcm
.keystream_pos
);
1107 ctx
->iv
.gcm
.keystream_pos
+= 16;
1111 #define SW_ENC_DEC(len) \
1112 static void aes##len##_sw_cbc_encrypt( \
1113 ssh_cipher *ciph, void *vblk, int blklen) \
1114 { aes_cbc_sw_encrypt(ciph, vblk, blklen); } \
1115 static void aes##len##_sw_cbc_decrypt( \
1116 ssh_cipher *ciph, void *vblk, int blklen) \
1117 { aes_cbc_sw_decrypt(ciph, vblk, blklen); } \
1118 static void aes##len##_sw_sdctr( \
1119 ssh_cipher *ciph, void *vblk, int blklen) \
1120 { aes_sdctr_sw(ciph, vblk, blklen); } \
1121 static void aes##len##_sw_gcm( \
1122 ssh_cipher *ciph, void *vblk, int blklen) \
1123 { aes_gcm_sw(ciph, vblk, blklen); } \
1124 static void aes##len##_sw_encrypt_ecb_block( \
1125 ssh_cipher *ciph, void *vblk) \
1126 { aes_encrypt_ecb_block_sw(ciph, vblk); }
1133 AES_ALL_VTABLES(_sw
, "unaccelerated");