x86/crypto/sm4-aesni-avx-asm_64.S

27 #define RTMP2        %xmm5  macro
172 	vmovdqa .Lbswap32_mask rRIP, RTMP2;
173 	vpshufb RTMP2, RA0, RA0;
174 	vpshufb RTMP2, RA1, RA1;
175 	vpshufb RTMP2, RA2, RA2;
176 	vpshufb RTMP2, RA3, RA3;
184 	vmovdqa .Linv_shift_row_rol_8 rRIP, RTMP2;
202 	vpshufb RTMP2, RX0, RTMP1;                                  \
227 	vmovdqa .Lbswap128_mask rRIP, RTMP2;
230 	vpshufb RTMP2, RA0, RA0;
231 	vpshufb RTMP2, RA1, RA1;
232 	vpshufb RTMP2, RA2, RA2;
233 	vpshufb RTMP2, RA3, RA3;
262 	vmovdqa .Lbswap32_mask rRIP, RTMP2;
263 	vpshufb RTMP2, RA0, RA0;
264 	vpshufb RTMP2, RA1, RA1;
265 	vpshufb RTMP2, RA2, RA2;
266 	vpshufb RTMP2, RA3, RA3;
267 	vpshufb RTMP2, RB0, RB0;
268 	vpshufb RTMP2, RB1, RB1;
269 	vpshufb RTMP2, RB2, RB2;
270 	vpshufb RTMP2, RB3, RB3;
284 	vmovdqa .Lpost_tf_lo_s rRIP, RTMP2;                         \
296 	transform_post(RX0, RTMP2, RTMP3, MASK_4BIT, RTMP0);        \
297 	transform_post(RX1, RTMP2, RTMP3, MASK_4BIT, RTMP0);        \
302 	vpshufb RTMP4, RX1, RTMP2;                                  \
304 	vpxor RTMP2, r0, r0; /* r0 ^ x */                           \
309 	vpxor RTMP3, RTMP2, RTMP2; /* x ^ rol(x,8) */               \
314 	vpxor RTMP3, RTMP2, RTMP2; /* x ^ rol(x,8) ^ rol(x,16) */   \
325 	vpslld $2, RTMP2, RTMP3;                                    \
326 	vpsrld $30, RTMP2, RTMP2;                                   \
327 	vpxor RTMP2, r0, r0;                                        \
343 	vmovdqa .Lbswap128_mask rRIP, RTMP2;
347 	vpshufb RTMP2, RA0, RA0;
348 	vpshufb RTMP2, RA1, RA1;
349 	vpshufb RTMP2, RA2, RA2;
350 	vpshufb RTMP2, RA3, RA3;
351 	vpshufb RTMP2, RB0, RB0;
352 	vpshufb RTMP2, RB1, RB1;
353 	vpshufb RTMP2, RB2, RB2;
354 	vpshufb RTMP2, RB3, RB3;
444 	inc_le128(RTMP0, RNOT, RTMP2); /* +1 */
446 	inc_le128(RTMP0, RNOT, RTMP2); /* +2 */
448 	inc_le128(RTMP0, RNOT, RTMP2); /* +3 */
450 	inc_le128(RTMP0, RNOT, RTMP2); /* +4 */
452 	inc_le128(RTMP0, RNOT, RTMP2); /* +5 */
454 	inc_le128(RTMP0, RNOT, RTMP2); /* +6 */
456 	inc_le128(RTMP0, RNOT, RTMP2); /* +7 */
458 	inc_le128(RTMP0, RNOT, RTMP2); /* +8 */