93 	__asm("movdqu %0, %%xmm0" :: "m" ((ctx)->sse[0]));	\
94 	__asm("movdqu %0, %%xmm1" :: "m" ((ctx)->sse[1]));	\
95 	__asm("movdqu %0, %%xmm2" :: "m" ((ctx)->sse[2]));	\
96 	__asm("movdqu %0, %%xmm3" :: "m" ((ctx)->sse[3]));	\
101 	__asm("movdqu %%xmm0, %0" : "=m" ((ctx)->sse[0]));	\
102 	__asm("movdqu %%xmm1, %0" : "=m" ((ctx)->sse[1]));	\
103 	__asm("movdqu %%xmm2, %0" : "=m" ((ctx)->sse[2]));	\
104 	__asm("movdqu %%xmm3, %0" : "=m" ((ctx)->sse[3]));	\
115 	__asm("pxor %xmm4, %xmm4");  in fletcher_4_sse2_native()
118 		__asm("movdqu %0, %%xmm5" :: "m"(*ip));  in fletcher_4_sse2_native()
119 		__asm("movdqa %xmm5, %xmm6");  in fletcher_4_sse2_native()
120 		__asm("punpckldq %xmm4, %xmm5");  in fletcher_4_sse2_native()
121 		__asm("punpckhdq %xmm4, %xmm6");  in fletcher_4_sse2_native()
122 		__asm("paddq %xmm5, %xmm0");  in fletcher_4_sse2_native()
123 		__asm("paddq %xmm0, %xmm1");  in fletcher_4_sse2_native()
124 		__asm("paddq %xmm1, %xmm2");  in fletcher_4_sse2_native()
125 		__asm("paddq %xmm2, %xmm3");  in fletcher_4_sse2_native()
126 		__asm("paddq %xmm6, %xmm0");  in fletcher_4_sse2_native()
127 		__asm("paddq %xmm0, %xmm1");  in fletcher_4_sse2_native()
128 		__asm("paddq %xmm1, %xmm2");  in fletcher_4_sse2_native()
129 		__asm("paddq %xmm2, %xmm3");  in fletcher_4_sse2_native()
147 		__asm("movd %0, %%xmm5" :: "r"(scratch1));  in fletcher_4_sse2_byteswap()
148 		__asm("movd %0, %%xmm6" :: "r"(scratch2));  in fletcher_4_sse2_byteswap()
149 		__asm("punpcklqdq %xmm6, %xmm5");  in fletcher_4_sse2_byteswap()
150 		__asm("paddq %xmm5, %xmm0");  in fletcher_4_sse2_byteswap()
151 		__asm("paddq %xmm0, %xmm1");  in fletcher_4_sse2_byteswap()
152 		__asm("paddq %xmm1, %xmm2");  in fletcher_4_sse2_byteswap()
153 		__asm("paddq %xmm2, %xmm3");  in fletcher_4_sse2_byteswap()
190 	__asm("movdqa %0, %%xmm7"::"m" (mask));  in fletcher_4_ssse3_byteswap()
191 	__asm("pxor %xmm4, %xmm4");  in fletcher_4_ssse3_byteswap()
194 		__asm("movdqu %0, %%xmm5"::"m" (*ip));  in fletcher_4_ssse3_byteswap()
195 		__asm("pshufb %xmm7, %xmm5");  in fletcher_4_ssse3_byteswap()
196 		__asm("movdqa %xmm5, %xmm6");  in fletcher_4_ssse3_byteswap()
197 		__asm("punpckldq %xmm4, %xmm5");  in fletcher_4_ssse3_byteswap()
198 		__asm("punpckhdq %xmm4, %xmm6");  in fletcher_4_ssse3_byteswap()
199 		__asm("paddq %xmm5, %xmm0");  in fletcher_4_ssse3_byteswap()
200 		__asm("paddq %xmm0, %xmm1");  in fletcher_4_ssse3_byteswap()
201 		__asm("paddq %xmm1, %xmm2");  in fletcher_4_ssse3_byteswap()
202 		__asm("paddq %xmm2, %xmm3");  in fletcher_4_ssse3_byteswap()
203 		__asm("paddq %xmm6, %xmm0");  in fletcher_4_ssse3_byteswap()
204 		__asm("paddq %xmm0, %xmm1");  in fletcher_4_ssse3_byteswap()
205 		__asm("paddq %xmm1, %xmm2");  in fletcher_4_ssse3_byteswap()
206 		__asm("paddq %xmm2, %xmm3");  in fletcher_4_ssse3_byteswap()