arm64/crypto/sm4-ce-ccm-core.S

1 /* SPDX-License-Identifier: GPL-2.0-or-later */
3  * SM4-CCM AEAD Algorithm using ARMv8 Crypto Extensions
13 #include "sm4-ce-asm.h"
15 .arch	armv8-a+crypto
17 .irp b, 0, 1, 8, 9, 10, 11, 12, 13, 14, 15, 16, 24, 25, 26, 27, 28, 29, 30, 31
32 		mov		vctr.d[1], x8;		\
34 		adds		x8, x8, #1;		\
35 		rev64		vctr.16b, vctr.16b;	\
49 	ld1		{RMAC.16b}, [x1]
57 	ld1		{v0.16b-v3.16b}, [x2], #64
60 	eor		RMAC.16b, RMAC.16b, v0.16b
62 	eor		RMAC.16b, RMAC.16b, v1.16b
64 	eor		RMAC.16b, RMAC.16b, v2.16b
66 	eor		RMAC.16b, RMAC.16b, v3.16b
72 	sub		w3, w3, #1
74 	ld1		{v0.16b}, [x2], #16
77 	eor		RMAC.16b, RMAC.16b, v0.16b
82 	st1		{RMAC.16b}, [x1]
95 	ld1		{RMAC.16b}, [x2]
96 	ld1		{v0.16b}, [x1]
100 	/* en-/decrypt the mac with ctr0 */
101 	eor		RMAC.16b, RMAC.16b, v0.16b
102 	st1		{RMAC.16b}, [x2]
123 	ld1		{RMAC.16b}, [x5]
126 	cmp		w4, #(4 * 16)
129 	sub		w4, w4, #(4 * 16)
133 	inc_le128(v9)			/* +1 */
137 	ld1		{v0.16b-v3.16b}, [x2], #64
140 	eor		v8.16b, v8.16b, v0.16b
141 	eor		RMAC.16b, RMAC.16b, v0.16b
143 	eor		v9.16b, v9.16b, v1.16b
144 	eor		RMAC.16b, RMAC.16b, v1.16b
146 	eor		v10.16b, v10.16b, v2.16b
147 	eor		RMAC.16b, RMAC.16b, v2.16b
149 	eor		v11.16b, v11.16b, v3.16b
150 	eor		RMAC.16b, RMAC.16b, v3.16b
152 	st1		{v8.16b-v11.16b}, [x1], #64
158 	cmp		w4, #16
161 	sub		w4, w4, #16
166 	ld1		{v0.16b}, [x2], #16
169 	eor		v8.16b, v8.16b, v0.16b
170 	eor		RMAC.16b, RMAC.16b, v0.16b
172 	st1		{v8.16b}, [x1], #16
184 	st1		{RMAC.16b}, [x5]
187 	ldrb		w0, [x2], #1		/* get 1 byte from input */
194 	strb		w9, [x1], #1		/* store out byte */
195 	strb		w6, [x5], #1		/* store MAC byte */
197 	subs		w4, w4, #1
201 	ext		RMAC.16b, RMAC.16b, RMAC.16b, #1
202 	ext		v8.16b, v8.16b, v8.16b, #1
208 	st1		{RMAC.16b}, [x5]
235 	ld1		{RMAC.16b}, [x5]
238 	cmp		w4, #(4 * 16)
241 	sub		w4, w4, #(4 * 16)
245 	inc_le128(v9)			/* +1 */
249 	ld1		{v0.16b-v3.16b}, [x2], #64
252 	eor		v8.16b, v8.16b, v0.16b
253 	eor		RMAC.16b, RMAC.16b, v8.16b
255 	eor		v9.16b, v9.16b, v1.16b
256 	eor		RMAC.16b, RMAC.16b, v9.16b
258 	eor		v10.16b, v10.16b, v2.16b
259 	eor		RMAC.16b, RMAC.16b, v10.16b
261 	eor		v11.16b, v11.16b, v3.16b
262 	eor		RMAC.16b, RMAC.16b, v11.16b
264 	st1		{v8.16b-v11.16b}, [x1], #64
270 	cmp		w4, #16
273 	sub		w4, w4, #16
278 	ld1		{v0.16b}, [x2], #16
281 	eor		v8.16b, v8.16b, v0.16b
282 	eor		RMAC.16b, RMAC.16b, v8.16b
284 	st1		{v8.16b}, [x1], #16
296 	st1		{RMAC.16b}, [x5]
299 	ldrb		w0, [x2], #1		/* get 1 byte from input */
306 	strb		w9, [x1], #1		/* store out byte */
307 	strb		w6, [x5], #1		/* store MAC byte */
309 	subs		w4, w4, #1
313 	ext		RMAC.16b, RMAC.16b, RMAC.16b, #1
314 	ext		v8.16b, v8.16b, v8.16b, #1
320 	st1		{RMAC.16b}, [x5]