590 #define transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \  argument
605 #define read_blocks(in, x0, x1, x2, x3, t0, t1, t2) \  argument
611 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2)
613 #define write_blocks(out, x0, x1, x2, x3, t0, t1, t2) \  argument
614 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
621 #define xor_blocks(out, x0, x1, x2, x3, t0, t1, t2) \  argument
622 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
624 	movdqu (0*4*4)(out),	t0; \
625 	pxor t0,		x0; \
627 	movdqu (1*4*4)(out),	t0; \
628 	pxor t0,		x1; \
630 	movdqu (2*4*4)(out),	t0; \
631 	pxor t0,		x2; \
633 	movdqu (3*4*4)(out),	t0; \
634 	pxor t0,		x3; \