x86/crypto/sm4-aesni-avx2-asm_64.S

26 #define RTMP0        %ymm3  macro
179 	transpose_4x4(RA0, RA1, RA2, RA3, RTMP0, RTMP1);
180 	transpose_4x4(RB0, RB1, RB2, RB3, RTMP0, RTMP1);
197 	transform_pre(RX0, RTMP4, RTMP1, MASK_4BIT, RTMP0);         \
198 	transform_pre(RX1, RTMP4, RTMP1, MASK_4BIT, RTMP0);         \
208 	transform_post(RX0, RTMP2, RTMP3, MASK_4BIT, RTMP0);        \
209 	transform_post(RX1, RTMP2, RTMP3, MASK_4BIT, RTMP0);        \
212 	vpshufb RTMP4, RX0, RTMP0;                                  \
213 	vpxor RTMP0, s0, s0; /* s0 ^ x */                           \
218 	vpxor RTMP1, RTMP0, RTMP0; /* x ^ rol(x,8) */               \
223 	vpxor RTMP1, RTMP0, RTMP0; /* x ^ rol(x,8) ^ rol(x,16) */   \
229 	vpslld $2, RTMP0, RTMP1;                                    \
230 	vpsrld $30, RTMP0, RTMP0;                                   \
231 	vpxor RTMP0, s0, s0;                                        \
257 	transpose_4x4(RA0, RA1, RA2, RA3, RTMP0, RTMP1);
258 	transpose_4x4(RB0, RB1, RB2, RB3, RTMP0, RTMP1);
306 	vinserti128 $1, RTMP4x, RTMP0, RTMP0;
307 	vpshufb RTMP3, RTMP0, RA0; /* +1 ; +0 */
314 	vpsubq RTMP2, RTMP0, RTMP0; /* +3 ; +2 */
315 	vpshufb RTMP3, RTMP0, RA1;
316 	vpsubq RTMP2, RTMP0, RTMP0; /* +5 ; +4 */
317 	vpshufb RTMP3, RTMP0, RA2;
318 	vpsubq RTMP2, RTMP0, RTMP0; /* +7 ; +6 */
319 	vpshufb RTMP3, RTMP0, RA3;
320 	vpsubq RTMP2, RTMP0, RTMP0; /* +9 ; +8 */
321 	vpshufb RTMP3, RTMP0, RB0;
322 	vpsubq RTMP2, RTMP0, RTMP0; /* +11 ; +10 */
323 	vpshufb RTMP3, RTMP0, RB1;
324 	vpsubq RTMP2, RTMP0, RTMP0; /* +13 ; +12 */
325 	vpshufb RTMP3, RTMP0, RB2;
326 	vpsubq RTMP2, RTMP0, RTMP0; /* +15 ; +14 */
327 	vpshufb RTMP3, RTMP0, RB3;
328 	vpsubq RTMP2, RTMP0, RTMP0; /* +16 */
335 	inc_le128(RTMP0, RNOT, RTMP1);
336 	inc_le128(RTMP0, RNOT, RTMP1);
337 	vpshufb RTMP3, RTMP0, RA1; /* +3 ; +2 */
338 	inc_le128(RTMP0, RNOT, RTMP1);
339 	inc_le128(RTMP0, RNOT, RTMP1);
340 	vpshufb RTMP3, RTMP0, RA2; /* +5 ; +4 */
341 	inc_le128(RTMP0, RNOT, RTMP1);
342 	inc_le128(RTMP0, RNOT, RTMP1);
343 	vpshufb RTMP3, RTMP0, RA3; /* +7 ; +6 */
344 	inc_le128(RTMP0, RNOT, RTMP1);
345 	inc_le128(RTMP0, RNOT, RTMP1);
346 	vpshufb RTMP3, RTMP0, RB0; /* +9 ; +8 */
347 	inc_le128(RTMP0, RNOT, RTMP1);
348 	inc_le128(RTMP0, RNOT, RTMP1);
349 	vpshufb RTMP3, RTMP0, RB1; /* +11 ; +10 */
350 	inc_le128(RTMP0, RNOT, RTMP1);
351 	inc_le128(RTMP0, RNOT, RTMP1);
352 	vpshufb RTMP3, RTMP0, RB2; /* +13 ; +12 */
353 	inc_le128(RTMP0, RNOT, RTMP1);
354 	inc_le128(RTMP0, RNOT, RTMP1);
355 	vpshufb RTMP3, RTMP0, RB3; /* +15 ; +14 */
356 	inc_le128(RTMP0, RNOT, RTMP1);
357 	vextracti128 $1, RTMP0, RTMP0x;