memcpy_mck.s

这个linux源代码是很全面的~基本完整了~使用c编译的~由于时间问题我没有亲自测试~但就算用来做参考资料也是非常好的

/*
 * Itanium 2-optimized version of memcpy and copy_user function
 *
 * Inputs:
 * 	in0:	destination address
 *	in1:	source address
 *	in2:	number of bytes to copy
 * Output:
 *	for bcopy:     return nothing
 *	for memcpy:    return dest
 * 	for copy_user: 0 if success,
 *		       or number of bytes NOT copied if error occurred.
 *
 * Copyright (C) 2002 Intel Corp.
 * Copyright (C) 2002 Ken Chen <[e-mail]>
 */
#include <linux/config.h>
#include <asm/asmmacro.h>
#include <asm/page.h>

#if __GNUC__ >= 3
# define EK(y...)	EX(y)
#else
# define EK(y,x...)	x
#endif

/* McKinley specific optimization */

#define retval		r8
#define saved_pfs	r31
#define saved_lc	r10
#define saved_pr	r11
#define saved_in0	r14
#define saved_in1	r15
#define saved_in2	r16

#define src0		r2
#define src1		r3
#define dst0		r17
#define dst1		r18
#define cnt		r9

/* r19-r30 are temp for each code section */
#define PREFETCH_DIST	8
#define src_pre_mem	r19
#define dst_pre_mem	r20
#define src_pre_l2	r21
#define dst_pre_l2	r22
#define t1		r23
#define t2		r24
#define t3		r25
#define t4		r26
#define t5		t1	// alias!
#define t6		t2	// alias!
#define t7		t3	// alias!
#define n8		r27
#define t9		t5	// alias!
#define t10		t4	// alias!
#define t11		t7	// alias!
#define t12		t6	// alias!
#define t14		t10	// alias!
#define t13		r28
#define t15		r29
#define tmp		r30

/* defines for long_copy block */
#define	A	0
#define B	(PREFETCH_DIST)
#define C	(B + PREFETCH_DIST)
#define D	(C + 1)
#define N	(D + 1)
#define Nrot	((N + 7) & ~7)

/* alias */
#define in0		r32
#define in1		r33
#define in2		r34

GLOBAL_ENTRY(bcopy)
	.regstk 3,0,0,0
	mov r8=in0		// swap the src and dest arguments
	mov in0=in1
	;;
	mov in1=r8
	;;
END(bcopy)			// fall through to memcpy
GLOBAL_ENTRY(memcpy)
	and	r28=0x7,in0
	and	r29=0x7,in1
	mov	f6=f0
	mov	retval=in0
	br.cond.sptk .common_code
	;;
END(memcpy)
GLOBAL_ENTRY(__copy_user)
	.prologue
// check dest alignment
	and	r28=0x7,in0
	and	r29=0x7,in1
	mov	f6=f1
	mov	saved_in0=in0	// save dest pointer
	mov	saved_in1=in1	// save src pointer
	mov	retval=r0	// initialize return value
	;;
.common_code:
	cmp.gt	p15,p0=8,in2	// check for small size
	cmp.ne	p13,p0=0,r28	// check dest alignment
	cmp.ne	p14,p0=0,r29	// check src alignment
	add	src0=0,in1
	sub	r30=8,r28	// for .align_dest
	mov	saved_in2=in2	// save len
	;;
	add	dst0=0,in0
	add	dst1=1,in0	// dest odd index
	cmp.le	p6,p0 = 1,r30	// for .align_dest
(p15)	br.cond.dpnt .memcpy_short
(p13)	br.cond.dpnt .align_dest
(p14)	br.cond.dpnt .unaligned_src
	;;

// both dest and src are aligned on 8-byte boundary
.aligned_src:
	.save ar.pfs, saved_pfs
	alloc	saved_pfs=ar.pfs,3,Nrot-3,0,Nrot
	.save pr, saved_pr
	mov	saved_pr=pr

	shr.u	cnt=in2,7	// this much cache line
	;;
	cmp.lt	p6,p0=2*PREFETCH_DIST,cnt
	cmp.lt	p7,p8=1,cnt
	.save ar.lc, saved_lc
	mov	saved_lc=ar.lc
	.body
	add	cnt=-1,cnt
	add	src_pre_mem=0,in1	// prefetch src pointer
	add	dst_pre_mem=0,in0	// prefetch dest pointer
	;;
(p7)	mov	ar.lc=cnt	// prefetch count
(p8)	mov	ar.lc=r0
(p6)	br.cond.dpnt .long_copy
	;;

.prefetch:
	lfetch	  [src_pre_mem], 128
	lfetch.excl [dst_pre_mem], 128
	br.cloop.dptk.few .prefetch
	;;

.medium_copy:
	and	tmp=31,in2	// copy length after iteration
	shr.u	r29=in2,5	// number of 32-byte iteration
	add	dst1=8,dst0	// 2nd dest pointer
	;;
	add	cnt=-1,r29	// ctop iteration adjustment
	cmp.eq	p10,p0=r29,r0	// do we really need to loop?
	add	src1=8,src0	// 2nd src pointer
	cmp.le	p6,p0=8,tmp
	;;
	cmp.le	p7,p0=16,tmp
	mov	ar.lc=cnt	// loop setup
	cmp.eq	p16,p17 = r0,r0
	mov	ar.ec=2
(p10)	br.dpnt.few .aligned_src_tail
	;;
	.align 32
1:
EX(.ex_handler, (p16)	ld8	r34=[src0],16)
EK(.ex_handler, (p16)	ld8	r38=[src1],16)
EX(.ex_handler, (p17)	st8	[dst0]=r33,16)
EK(.ex_handler, (p17)	st8	[dst1]=r37,16)
	;;
EX(.ex_handler, (p16)	ld8	r32=[src0],16)
EK(.ex_handler, (p16)	ld8	r36=[src1],16)
EX(.ex_handler, (p16)	st8	[dst0]=r34,16)
EK(.ex_handler, (p16)	st8	[dst1]=r38,16)
	br.ctop.dptk.few 1b
	;;

.aligned_src_tail:
EX(.ex_handler, (p6)	ld8	t1=[src0])
	mov	ar.lc=saved_lc
	mov	ar.pfs=saved_pfs
EX(.ex_hndlr_s, (p7)	ld8	t2=[src1],8)
	cmp.le	p8,p0=24,tmp
	and	r21=-8,tmp
	;;
EX(.ex_hndlr_s, (p8)	ld8	t3=[src1])
EX(.ex_handler, (p6)	st8	[dst0]=t1)	// store byte 1
	and	in2=7,tmp	// remaining length
EX(.ex_hndlr_d, (p7)	st8	[dst1]=t2,8)	// store byte 2
	add	src0=src0,r21	// setting up src pointer
	add	dst0=dst0,r21	// setting up dest pointer
	;;
EX(.ex_handler, (p8)	st8	[dst1]=t3)	// store byte 3
	mov	pr=saved_pr,-1
	br.dptk.many .memcpy_short
	;;

/* code taken from copy_page_mck */
.long_copy:
	.rotr v[2*PREFETCH_DIST]
	.rotp p[N]

	mov src_pre_mem = src0
	mov pr.rot = 0x10000
	mov ar.ec = 1				// special unrolled loop

	mov dst_pre_mem = dst0

	add src_pre_l2 = 8*8, src0
	add dst_pre_l2 = 8*8, dst0
	;;
	add src0 = 8, src_pre_mem		// first t1 src
	mov ar.lc = 2*PREFETCH_DIST - 1
	shr.u cnt=in2,7				// number of lines
	add src1 = 3*8, src_pre_mem		// first t3 src
	add dst0 = 8, dst_pre_mem		// first t1 dst
	add dst1 = 3*8, dst_pre_mem		// first t3 dst
	;;
	and tmp=127,in2				// remaining bytes after this block
	add cnt = -(2*PREFETCH_DIST) - 1, cnt
	// same as .line_copy loop, but with all predicated-off instructions removed:
.prefetch_loop:
EX(.ex_hndlr_lcpy_1, (p[A])	ld8 v[A] = [src_pre_mem], 128)		// M0
EK(.ex_hndlr_lcpy_1, (p[B])	st8 [dst_pre_mem] = v[B], 128)		// M2
	br.ctop.sptk .prefetch_loop
	;;
	cmp.eq p16, p0 = r0, r0			// reset p16 to 1
	mov ar.lc = cnt
	mov ar.ec = N				// # of stages in pipeline
	;;
.line_copy:
EX(.ex_handler,	(p[D])	ld8 t2 = [src0], 3*8)			// M0
EK(.ex_handler,	(p[D])	ld8 t4 = [src1], 3*8)			// M1
EX(.ex_handler_lcpy,	(p[B])	st8 [dst_pre_mem] = v[B], 128)		// M2 prefetch dst from memory
EK(.ex_handler_lcpy,	(p[D])	st8 [dst_pre_l2] = n8, 128)		// M3 prefetch dst from L2
	;;
EX(.ex_handler_lcpy,	(p[A])	ld8 v[A] = [src_pre_mem], 128)		// M0 prefetch src from memory
EK(.ex_handler_lcpy,	(p[C])	ld8 n8 = [src_pre_l2], 128)		// M1 prefetch src from L2
EX(.ex_handler,	(p[D])	st8 [dst0] =  t1, 8)			// M2
EK(.ex_handler,	(p[D])	st8 [dst1] =  t3, 8)			// M3
	;;
EX(.ex_handler,	(p[D])	ld8  t5 = [src0], 8)
EK(.ex_handler,	(p[D])	ld8  t7 = [src1], 3*8)
EX(.ex_handler,	(p[D])	st8 [dst0] =  t2, 3*8)
EK(.ex_handler,	(p[D])	st8 [dst1] =  t4, 3*8)
	;;
EX(.ex_handler,	(p[D])	ld8  t6 = [src0], 3*8)
EK(.ex_handler,	(p[D])	ld8 t10 = [src1], 8)
EX(.ex_handler,	(p[D])	st8 [dst0] =  t5, 8)
EK(.ex_handler,	(p[D])	st8 [dst1] =  t7, 3*8)
	;;
EX(.ex_handler,	(p[D])	ld8  t9 = [src0], 3*8)
EK(.ex_handler,	(p[D])	ld8 t11 = [src1], 3*8)
EX(.ex_handler,	(p[D])	st8 [dst0] =  t6, 3*8)
EK(.ex_handler,	(p[D])	st8 [dst1] = t10, 8)
	;;
EX(.ex_handler,	(p[D])	ld8 t12 = [src0], 8)
EK(.ex_handler,	(p[D])	ld8 t14 = [src1], 8)
EX(.ex_handler,	(p[D])	st8 [dst0] =  t9, 3*8)
EK(.ex_handler,	(p[D])	st8 [dst1] = t11, 3*8)
	;;
EX(.ex_handler,	(p[D])	ld8 t13 = [src0], 4*8)
EK(.ex_handler,	(p[D])	ld8 t15 = [src1], 4*8)
EX(.ex_handler,	(p[D])	st8 [dst0] = t12, 8)
EK(.ex_handler,	(p[D])	st8 [dst1] = t14, 8)
	;;
EX(.ex_handler,	(p[C])	ld8  t1 = [src0], 8)
EK(.ex_handler,	(p[C])	ld8  t3 = [src1], 8)
EX(.ex_handler,	(p[D])	st8 [dst0] = t13, 4*8)
EK(.ex_handler,	(p[D])	st8 [dst1] = t15, 4*8)
	br.ctop.sptk .line_copy
	;;

	add dst0=-8,dst0
	add src0=-8,src0
	mov in2=tmp
	.restore sp
	br.sptk.many .medium_copy
	;;

#define BLOCK_SIZE	128*32
#define blocksize	r23
#define curlen		r24

// dest is on 8-byte boundary, src is not. We need to do
// ld8-ld8, shrp, then st8.  Max 8 byte copy per cycle.
.unaligned_src:
	.prologue
	.save ar.pfs, saved_pfs
	alloc	saved_pfs=ar.pfs,3,5,0,8
	.save ar.lc, saved_lc
	mov	saved_lc=ar.lc
	.save pr, saved_pr
	mov	saved_pr=pr
	.body
.4k_block:
	mov	saved_in0=dst0	// need to save all input arguments
	mov	saved_in2=in2
	mov	blocksize=BLOCK_SIZE
	;;
	cmp.lt	p6,p7=blocksize,in2
	mov	saved_in1=src0
	;;
(p6)	mov	in2=blocksize
	;;
	shr.u	r21=in2,7	// this much cache line
	shr.u	r22=in2,4	// number of 16-byte iteration
	and	curlen=15,in2	// copy length after iteration
	and	r30=7,src0	// source alignment
	;;
	cmp.lt	p7,p8=1,r21
	add	cnt=-1,r21
	;;

	add	src_pre_mem=0,src0	// prefetch src pointer
	add	dst_pre_mem=0,dst0	// prefetch dest pointer
	and	src0=-8,src0		// 1st src pointer
(p7)	mov	ar.lc = r21
(p8)	mov	ar.lc = r0
	;;
	.align 32
1:	lfetch	  [src_pre_mem], 128
	lfetch.excl [dst_pre_mem], 128
	br.cloop.dptk.few 1b
	;;

	shladd	dst1=r22,3,dst0	// 2nd dest pointer
	shladd	src1=r22,3,src0	// 2nd src pointer
	cmp.eq	p8,p9=r22,r0	// do we really need to loop?
	cmp.le	p6,p7=8,curlen;	// have at least 8 byte remaining?
	add	cnt=-1,r22	// ctop iteration adjustment
	;;
EX(.ex_handler, (p9)	ld8	r33=[src0],8)	// loop primer
EK(.ex_handler, (p9)	ld8	r37=[src1],8)
(p8)	br.dpnt.few .noloop
	;;

// The jump address is calculated based on src alignment. The COPYU