468 #define transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \  argument
483 #define read_blocks(in, x0, x1, x2, x3, t0, t1, t2) \  argument
489 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2)
491 #define write_blocks(out, x0, x1, x2, x3, t0, t1, t2) \  argument
492 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
499 #define xor_blocks(out, x0, x1, x2, x3, t0, t1, t2) \  argument
500 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
502 	movdqu (0*4*4)(out),	t0; \
503 	pxor t0,		x0; \
505 	movdqu (1*4*4)(out),	t0; \
506 	pxor t0,		x1; \
508 	movdqu (2*4*4)(out),	t0; \
509 	pxor t0,		x2; \
511 	movdqu (3*4*4)(out),	t0; \
512 	pxor t0,		x3; \