ia64: Add McKinley-tuned versions of copy_user() and memcpy(). Patch by Ken Chen.

arch/ia64/lib/memcpy_mck.S

+/*
+ * Itanium 2-optimized version of memcpy and copy_user function
+ *
+ * Inputs:
+ * 	in0:	destination address
+ *	in1:	source address
+ *	in2:	number of bytes to copy
+ * Output:
+ * 	0 if success, or number of byte NOT copied if error occurred.
+ *
+ * Copyright (C) 2002 Intel Corp.
+ * Copyright (C) 2002 Ken Chen <kenneth.w.chen@intel.com>
+ */
+#include <linux/config.h>
+#include <asm/asmmacro.h>
+#include <asm/page.h>
+
+#define EK(y,x...)  x
+
+GLOBAL_ENTRY(bcopy)
+	.regstk 3,0,0,0
+	mov r8=in0
+	mov in0=in1
+	;;
+	mov in1=r8
+	;;
+END(bcopy)
+
+/* 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(memcpy)
+	and	r28=0x7,in0
+	and	r29=0x7,in1
+	mov	f6=f0
+	br.cond.sptk .common_code
+	;;
+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	saved_in2=in2	// save len
+	;;
+.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	retval=r0	// initialize return value
+	;;
+	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.fault	  [src_pre_mem], 128
+	lfetch.fault.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_handler, (p7)	ld8	t2=[src1],8)
+	cmp.le	p8,p0=24,tmp
+	and	r21=-8,tmp
+	;;
+EX(.ex_handler, (p8)	ld8	t3=[src1])
+EK(.ex_handler, (p6)	st8	[dst0]=t1)	// store byte 1
+	and	in2=7,tmp	// remaining length
+EX(.ex_handler, (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_handler_lcpy, (p[A])	ld8 v[A] = [src_pre_mem], 128)		// M0
+EK(.ex_handler_lcpy, (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.fault	  [src_pre_mem], 128
+	lfetch.fault.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
+// macro below need to confine its size to power of two, so an entry
+// can be caulated using shl instead of an expensive multiply. The
+// size is then hard coded by the following #define to match the
+// actual size.  This make it somewhat tedious when COPYU macro gets
+// changed and this need to be adjusted to match.
+#define LOOP_SIZE 6
+1:
+	mov	r29=ip		// jmp_table thread
+	mov	ar.lc=cnt
+	;;
+	add	r29=.jump_table - 1b - (.jmp1-.jump_table), r29
+	shl	r28=r30, LOOP_SIZE	// jmp_table thread
+	mov	ar.ec=2		// loop setup
+	;;
+	add	r29=r29,r28		// jmp_table thread
+	cmp.eq	p16,p17=r0,r0
+	;;
+	mov	b6=r29			// jmp_table thread
+	;;
+	br.cond.sptk.few b6
+
+// for 8-15 byte case
+// We will skip the loop, but need to replicate the side effect
+// that the loop produces.
+.noloop:
+EX(.ex_handler, (p6)	ld8	r37=[src1],8)
+	nop.m	0
+(p6)	shl	r25=r30,3
+	;;
+EX(.ex_handler, (p6)	ld8	r27=[src1])
+(p6)	shr.u	r28=r37,r25
+(p6)	sub	r26=64,r25
+	;;
+(p6)	shl	r27=r27,r26
+	;;
+(p6)	or	r21=r28,r27
+
+.unaligned_src_tail:
+/* check if we have more than blocksize to copy, if so go back */
+	cmp.gt	p8,p0=saved_in2,blocksize
+	;;
+(p8)	add	dst0=saved_in0,blocksize
+(p8)	add	src0=saved_in1,blocksize
+(p8)	sub	in2=saved_in2,blocksize
+(p8)	br.dpnt	.4k_block
+	;;
+
+/* we have up to 15 byte to copy in the tail.
+ * part of work is already done in the jump table code
+ * we are at the following state.
+ * src side:
+ * 
+ *   xxxxxx xx                   <----- r21 has xxxxxxxx already
+ * -------- -------- --------
+ * 0        8        16
+ *          ^
+ *          |
+ *          src1
+ * 
+ * dst
+ * -------- -------- --------
+ * ^
+ * |
+ * dst1
+ */
+EX(.ex_handler, (p6)	st8	[dst1]=r21,8)	// more than 8 byte to copy
+(p6)	add	curlen=-8,curlen	// update length
+	mov	ar.pfs=saved_pfs
+	;;
+	mov	ar.lc=saved_lc
+	mov	pr=saved_pr,-1
+	mov	in2=curlen	// remaining length
+	mov	dst0=dst1	// dest pointer
+	add	src0=src1,r30	// forward by src alignment
+	;;
+
+// 7 byte or smaller.
+.memcpy_short:
+	cmp.le	p8,p9   = 1,in2
+	cmp.le	p10,p11 = 2,in2
+	cmp.le	p12,p13 = 3,in2
+	cmp.le	p14,p15 = 4,in2
+	add	src1=1,src0	// second src pointer
+	add	dst1=1,dst0	// second dest pointer
+	;;
+
+EX(.ex_handler_short, (p8)	ld1	t1=[src0],2)
+EK(.ex_handler_short, (p10)	ld1	t2=[src1],2)
+(p9)	br.ret.dpnt rp		// 0 byte copy
+	;;
+
+EX(.ex_handler_short, (p8)	st1	[dst0]=t1,2)
+EK(.ex_handler_short, (p10)	st1	[dst1]=t2,2)
+(p11)	br.ret.dpnt rp		// 1 byte copy
+
+EX(.ex_handler_short, (p12)	ld1	t3=[src0],2)
+EK(.ex_handler_short, (p14)	ld1	t4=[src1],2)
+(p13)	br.ret.dpnt rp		// 2 byte copy
+	;;
+
+	cmp.le	p6,p7   = 5,in2
+	cmp.le	p8,p9   = 6,in2
+	cmp.le	p10,p11 = 7,in2
+
+EX(.ex_handler_short, (p12)	st1	[dst0]=t3,2)
+EK(.ex_handler_short, (p14)	st1	[dst1]=t4,2)
+(p15)	br.ret.dpnt rp		// 3 byte copy
+	;;
+
+EX(.ex_handler_short, (p6)	ld1	t5=[src0],2)
+EK(.ex_handler_short, (p8)	ld1	t6=[src1],2)
+(p7)	br.ret.dpnt rp		// 4 byte copy
+	;;
+
+EX(.ex_handler_short, (p6)	st1	[dst0]=t5,2)
+EK(.ex_handler_short, (p8)	st1	[dst1]=t6,2)
+(p9)	br.ret.dptk rp		// 5 byte copy
+
+EX(.ex_handler_short, (p10)	ld1	t7=[src0],2)
+(p11)	br.ret.dptk rp		// 6 byte copy
+	;;
+
+EX(.ex_handler_short, (p10)	st1	[dst0]=t7,2)
+	br.ret.dptk rp		// done all cases
+
+
+/* Align dest to nearest 8-byte boundary. We know we have at
+ * least 7 bytes to copy, enough to crawl to 8-byte boundary.
+ * Actual number of byte to crawl depend on the dest alignment.
+ * 7 byte or less is taken care at .memcpy_short
+
+ * src0 - source even index
+ * src1 - source  odd index
+ * dst0 - dest even index
+ * dst1 - dest  odd index
+ * r30  - distance to 8-byte boundary
+ */
+
+.align_dest:
+	add	src1=1,in1	// source odd index
+	cmp.le	p7,p0 = 2,r30	// for .align_dest
+	cmp.le	p8,p0 = 3,r30	// for .align_dest
+EX(.ex_handler_short, (p6)	ld1	t1=[src0],2)
+	cmp.le	p9,p0 = 4,r30	// for .align_dest
+	cmp.le	p10,p0 = 5,r30
+	;;
+EX(.ex_handler_short, (p7)	ld1	t2=[src1],2)
+EK(.ex_handler_short, (p8)	ld1	t3=[src0],2)
+	cmp.le	p11,p0 = 6,r30
+EX(.ex_handler_short, (p6)	st1	[dst0] = t1,2)
+	cmp.le	p12,p0 = 7,r30
+	;;
+EX(.ex_handler_short, (p9)	ld1	t4=[src1],2)
+EK(.ex_handler_short, (p10)	ld1	t5=[src0],2)
+EX(.ex_handler_short, (p7)	st1	[dst1] = t2,2)
+EK(.ex_handler_short, (p8)	st1	[dst0] = t3,2)
+	;;
+EX(.ex_handler_short, (p11)	ld1	t6=[src1],2)
+EK(.ex_handler_short, (p12)	ld1	t7=[src0],2)
+	cmp.eq	p6,p7=r28,r29
+EX(.ex_handler_short, (p9)	st1	[dst1] = t4,2)
+EK(.ex_handler_short, (p10)	st1	[dst0] = t5,2)
+	sub	in2=in2,r30
+	;;
+EX(.ex_handler_short, (p11)	st1	[dst1] = t6,2)
+EK(.ex_handler_short, (p12)	st1	[dst0] = t7)
+	add	dst0=in0,r30	// setup arguments
+	add	src0=in1,r30
+(p6)	br.cond.dptk .aligned_src
+(p7)	br.cond.dpnt .unaligned_src
+	;;
+
+/* main loop body in jump table format */
+#define COPYU(shift)									\
+1:											\
+EX(.ex_handler,  (p16)	ld8	r32=[src0],8);		/* 1 */				\
+EK(.ex_handler,  (p16)	ld8	r36=[src1],8);						\
+		 (p17)	shrp	r35=r33,r34,shift;;	/* 1 */				\
+EX(.ex_handler,  (p6)	ld8	r22=[src1]);	/* common, prime for tail section */	\
+		 nop.m	0;								\
+		 (p16)	shrp	r38=r36,r37,shift;					\
+EX(.ex_handler,  (p17)	st8	[dst0]=r35,8);		/* 1 */				\
+EK(.ex_handler,  (p17)	st8	[dst1]=r39,8);						\
+		 br.ctop.dptk.few 1b;;							\
+		 (p7)	add	src1=-8,src1;	/* back out for <8 byte case */		\
+		 shrp	r21=r22,r38,shift;	/* speculative work */			\
+		 br.sptk.few .unaligned_src_tail /* branch out of jump table */		\
+		 ;;
+	.align 32
+.jump_table:
+	COPYU(8)	// unaligned cases
+.jmp1:
+	COPYU(16)
+	COPYU(24)
+	COPYU(32)
+	COPYU(40)
+	COPYU(48)
+	COPYU(56)
+
+#undef A
+#undef B
+#undef C
+#undef D
+END(memcpy)
+
+/*
+ * Due to lack of local tag support in gcc 2.x assembler, it is not clear which
+ * instruction failed in the bundle.  The exception algorithm is that we
+ * first figure out the faulting address, then detect if there is any
+ * progress made on the copy, if so, redo the copy from last known copied
+ * location up to the faulting address (exclusive). In the copy_from_user
+ * case, remaining byte in kernel buffer will be zeroed.
+ *
+ * Take copy_from_user as an example, in the code there are multiple loads
+ * in a bundle and those multiple loads could span over two pages, the
+ * faulting address is calculated as page_round_down(max(src0, src1)).
+ * This is based on knowledge that if we can access one byte in a page, we
+ * can access any byte in that page.
+ *
+ * predicate used in the exception handler:
+ * p6-p7: direction
+ * p10-p11: src faulting addr calculation
+ * p12-p13: dst faulting addr calculation
+ */
+
+#define A	r19
+#define B	r20
+#define C	r21
+#define D	r22
+#define F	r28
+
+#define memset_arg0	r32
+#define memset_arg2	r33
+
+#define saved_retval	loc0
+#define saved_rtlink	loc1
+#define saved_pfs_stack	loc2
+
+.ex_handler_lcpy:
+	// in long copy block, the preload addresses should always ahead
+	// of the other two src/det pointers.  Furthermore, src1/dst1 should
+	// always ahead of src0/dst0.
+	cmp.ltu	p10,p11=src_pre_l2,src_pre_mem
+	cmp.ltu	p12,p13=dst_pre_l2,dst_pre_mem
+	;;
+(p10)	mov	src1=src_pre_mem
+(p11)	mov	src1=src_pre_l2
+(p12)	mov	dst1=dst_pre_mem
+(p13)	mov	dst1=dst_pre_l2
+	;;
+.ex_handler:
+	mov	pr=saved_pr,-1		// first restore pr, lc, and pfs
+	mov	ar.lc=saved_lc
+	mov	ar.pfs=saved_pfs
+	;;
+.ex_handler_short: // fault occurred in these sections didn't change pr, lc, pfs
+	cmp.ltu	p6,p7=saved_in0, saved_in1	// get the copy direction
+	cmp.ltu	p10,p11=src0,src1
+	cmp.ltu	p12,p13=dst0,dst1
+	fcmp.eq	p8,p0=f6,f0		// is it memcpy?
+	mov	tmp = dst0
+	;;
+(p11)	mov	src1 = src0		// pick the larger of the two
+(p13)	mov	dst0 = dst1		// make dst0 the smaller one
+(p13)	mov	dst1 = tmp		// and dst1 the larger one
+	;;
+(p6)	dep	F = r0,dst1,0,PAGE_SHIFT // usr dst round down to page boundary
+(p7)	dep	F = r0,src1,0,PAGE_SHIFT // usr src round down to page boundary
+	;;
+(p6)	cmp.le	p14,p0=F,saved_in0	// bad address to start with
+(p7)	cmp.le	p14,p0=F,saved_in1	// here too
+	mov	retval=saved_in2
+(p8)	ld1	tmp=[src1]		// force an oops for memcpy call
+(p8)	st1	[dst1]=r0		// force an oops for memcpy call
+(p14)	br.ret.sptk.many rp
+
+/*
+ * The remaining byte to copy is calculated as:
+ *
+ * A =	(faulting_addr - orig_src)	-> len to faulting ld address
+ *	or 
+ * 	(faulting_addr - orig_dst)	-> len to faulting st address
+ * B =	(cur_dst - orig_dst)		-> len copied so far
+ * C =	A - B				-> len need to be copied
+ * D =	orig_len - A			-> len need to be zeroed
+ */
+(p6)	sub	A = F, saved_in0
+(p7)	sub	A = F, saved_in1
+	clrrrb
+	;;
+	alloc	saved_pfs_stack=ar.pfs,3,3,3,0
+	sub	B = dst0, saved_in0	// how many byte copied so far
+	;;
+	sub	C = A, B
+	sub	D = saved_in2, A
+	;;
+	cmp.gt	p8,p0=C,r0		// more than 1 byte?
+	add	memset_arg0=saved_in0, A
+(p6)	mov	memset_arg2=0		// copy_to_user should not call memset
+(p7)	mov	memset_arg2=D		// copy_from_user need to have kbuf zeroed
+	mov	r8=0
+	mov	saved_retval = D
+	mov	saved_rtlink = b0
+
+	add	out0=saved_in0, B
+	add	out1=saved_in1, B
+	mov	out2=C
+(p8)	br.call.sptk.few b0=__copy_user	// recursive call
+	;;
+
+	add	saved_retval=saved_retval,r8	// above might return non-zero value
+	cmp.gt	p8,p0=memset_arg2,r0	// more than 1 byte?
+	mov	out0=memset_arg0	// *s
+	mov	out1=r0			// c
+	mov	out2=memset_arg2	// n
+(p8)	br.call.sptk.few b0=memset
+	;;
+
+	mov	retval=saved_retval
+	mov	ar.pfs=saved_pfs_stack
+	mov	b0=saved_rtlink
+	br.ret.sptk.many rp
+
+/* end of McKinley specific optimization */
+END(__copy_user)