1 // SPDX-License-Identifier: GPL-2.0-only
3  * arch/arm64/lib/xor-neon.c
11 #include <asm/neon-intrinsics.h>
20 	long lines = bytes / (sizeof(uint64x2_t) * 4);  in xor_arm64_neon_2()
26 		v2 = veorq_u64(vld1q_u64(dp1 +  4), vld1q_u64(dp2 +  4));  in xor_arm64_neon_2()
27 		v3 = veorq_u64(vld1q_u64(dp1 +  6), vld1q_u64(dp2 +  6));  in xor_arm64_neon_2()
32 		vst1q_u64(dp1 +  4, v2);  in xor_arm64_neon_2()
33 		vst1q_u64(dp1 +  6, v3);  in xor_arm64_neon_2()
37 	} while (--lines > 0);  in xor_arm64_neon_2()
49 	long lines = bytes / (sizeof(uint64x2_t) * 4);  in xor_arm64_neon_3()
55 		v2 = veorq_u64(vld1q_u64(dp1 +  4), vld1q_u64(dp2 +  4));  in xor_arm64_neon_3()
56 		v3 = veorq_u64(vld1q_u64(dp1 +  6), vld1q_u64(dp2 +  6));  in xor_arm64_neon_3()
61 		v2 = veorq_u64(v2, vld1q_u64(dp3 +  4));  in xor_arm64_neon_3()
62 		v3 = veorq_u64(v3, vld1q_u64(dp3 +  6));  in xor_arm64_neon_3()
67 		vst1q_u64(dp1 +  4, v2);  in xor_arm64_neon_3()
68 		vst1q_u64(dp1 +  6, v3);  in xor_arm64_neon_3()
73 	} while (--lines > 0);  in xor_arm64_neon_3()
87 	long lines = bytes / (sizeof(uint64x2_t) * 4);  in xor_arm64_neon_4()
93 		v2 = veorq_u64(vld1q_u64(dp1 +  4), vld1q_u64(dp2 +  4));  in xor_arm64_neon_4()
94 		v3 = veorq_u64(vld1q_u64(dp1 +  6), vld1q_u64(dp2 +  6));  in xor_arm64_neon_4()
99 		v2 = veorq_u64(v2, vld1q_u64(dp3 +  4));  in xor_arm64_neon_4()
100 		v3 = veorq_u64(v3, vld1q_u64(dp3 +  6));  in xor_arm64_neon_4()
105 		v2 = veorq_u64(v2, vld1q_u64(dp4 +  4));  in xor_arm64_neon_4()
106 		v3 = veorq_u64(v3, vld1q_u64(dp4 +  6));  in xor_arm64_neon_4()
111 		vst1q_u64(dp1 +  4, v2);  in xor_arm64_neon_4()
112 		vst1q_u64(dp1 +  6, v3);  in xor_arm64_neon_4()
118 	} while (--lines > 0);  in xor_arm64_neon_4()
134 	long lines = bytes / (sizeof(uint64x2_t) * 4);  in xor_arm64_neon_5()
140 		v2 = veorq_u64(vld1q_u64(dp1 +  4), vld1q_u64(dp2 +  4));  in xor_arm64_neon_5()
141 		v3 = veorq_u64(vld1q_u64(dp1 +  6), vld1q_u64(dp2 +  6));  in xor_arm64_neon_5()
146 		v2 = veorq_u64(v2, vld1q_u64(dp3 +  4));  in xor_arm64_neon_5()
147 		v3 = veorq_u64(v3, vld1q_u64(dp3 +  6));  in xor_arm64_neon_5()
152 		v2 = veorq_u64(v2, vld1q_u64(dp4 +  4));  in xor_arm64_neon_5()
153 		v3 = veorq_u64(v3, vld1q_u64(dp4 +  6));  in xor_arm64_neon_5()
158 		v2 = veorq_u64(v2, vld1q_u64(dp5 +  4));  in xor_arm64_neon_5()
159 		v3 = veorq_u64(v3, vld1q_u64(dp5 +  6));  in xor_arm64_neon_5()
164 		vst1q_u64(dp1 +  4, v2);  in xor_arm64_neon_5()
165 		vst1q_u64(dp1 +  6, v3);  in xor_arm64_neon_5()
172 	} while (--lines > 0);  in xor_arm64_neon_5()
204 	long lines = bytes / (sizeof(uint64x2_t) * 4);  in xor_arm64_eor3_3()
212 		v2 = eor3(vld1q_u64(dp1 + 4), vld1q_u64(dp2 + 4),  in xor_arm64_eor3_3()
213 			  vld1q_u64(dp3 + 4));  in xor_arm64_eor3_3()
214 		v3 = eor3(vld1q_u64(dp1 + 6), vld1q_u64(dp2 + 6),  in xor_arm64_eor3_3()
215 			  vld1q_u64(dp3 + 6));  in xor_arm64_eor3_3()
220 		vst1q_u64(dp1 + 4, v2);  in xor_arm64_eor3_3()
221 		vst1q_u64(dp1 + 6, v3);  in xor_arm64_eor3_3()
226 	} while (--lines > 0);  in xor_arm64_eor3_3()
241 	long lines = bytes / (sizeof(uint64x2_t) * 4);  in xor_arm64_eor3_4()
249 		v2 = eor3(vld1q_u64(dp1 + 4), vld1q_u64(dp2 + 4),  in xor_arm64_eor3_4()
250 			  vld1q_u64(dp3 + 4));  in xor_arm64_eor3_4()
251 		v3 = eor3(vld1q_u64(dp1 + 6), vld1q_u64(dp2 + 6),  in xor_arm64_eor3_4()
252 			  vld1q_u64(dp3 + 6));  in xor_arm64_eor3_4()
257 		v2 = veorq_u64(v2, vld1q_u64(dp4 + 4));  in xor_arm64_eor3_4()
258 		v3 = veorq_u64(v3, vld1q_u64(dp4 + 6));  in xor_arm64_eor3_4()
263 		vst1q_u64(dp1 + 4, v2);  in xor_arm64_eor3_4()
264 		vst1q_u64(dp1 + 6, v3);  in xor_arm64_eor3_4()
270 	} while (--lines > 0);  in xor_arm64_eor3_4()
287 	long lines = bytes / (sizeof(uint64x2_t) * 4);  in xor_arm64_eor3_5()
295 		v2 = eor3(vld1q_u64(dp1 + 4), vld1q_u64(dp2 + 4),  in xor_arm64_eor3_5()
296 			  vld1q_u64(dp3 + 4));  in xor_arm64_eor3_5()
297 		v3 = eor3(vld1q_u64(dp1 + 6), vld1q_u64(dp2 + 6),  in xor_arm64_eor3_5()
298 			  vld1q_u64(dp3 + 6));  in xor_arm64_eor3_5()
303 		v2 = eor3(v2, vld1q_u64(dp4 + 4), vld1q_u64(dp5 + 4));  in xor_arm64_eor3_5()
304 		v3 = eor3(v3, vld1q_u64(dp4 + 6), vld1q_u64(dp5 + 6));  in xor_arm64_eor3_5()
309 		vst1q_u64(dp1 + 4, v2);  in xor_arm64_eor3_5()
310 		vst1q_u64(dp1 + 6, v3);  in xor_arm64_eor3_5()
317 	} while (--lines > 0);  in xor_arm64_eor3_5()