prefetch.old.t3e

; prefetch.t3e: put objects into cache
; assuming 8 byte alinment, 32 byte cache line
; need to go from address to (address+size-8)
#include <mpp/asdef.h>
CRI_REGISTER_NAMES

	.ident sub$c
	;.psect	kernel@data,data,cache

	;.endp

	.psect	kernel@code,code,cache
;	 ENTER	prefetch_vector,zero,user
prefetch_vector::	; su3_vector
	ldq	zero,0(a0)
	ldq	zero,16(a0)
	ret	zero,(ra)
	.endp

	.psect	kernel@code,code,cache
;	 ENTER	prefetch_4_vectors,zero,user
prefetch_4_vectors::	; su3_vector
	ldq	zero,0(a0)
	ldq	zero,16(a0)
	ldq	zero,0(a1)
	ldq	zero,16(a1)
	ldq	zero,0(a2)
	ldq	zero,16(a2)
	ldq	zero,0(a3)
	ldq	zero,16(a3)
	ret	zero,(ra)
	.endp

	.psect	kernel@code,code,cache
;	 ENTER	prefetch_wvec,zero,user
prefetch_wvec::	; wilson_vector
	ldq	zero,0(a0)
	ldq	zero,32(a0)
	ldq	zero,64(a0)
	ldq	zero,88(a0)
	ret	zero,(ra)
	.endp

	.psect	kernel@code,code,cache
;	 ENTER	prefetch_hwvec,zero,user
prefetch_hwvec::	; half_wilson_vector
	ldq	zero,0(a0)
	ldq	zero,32(a0)
	ldq	zero,40(a0)
	ret	zero,(ra)
	.endp

	.psect	kernel@code,code,cache
;	 ENTER	prefetch_matrix,zero,user
prefetch_matrix::	; su3_matrix
	ldq	zero,0(a0)
	ldq	zero,32(a0)
	ldq	zero,64(a0)
	ret	zero,(ra)
	.endp
	.end