X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=arch%2Fsparc64%2Flib%2FU3memcpy.S;h=8fe195a10bbadd6b437e508bf77928ed9d093a37;hb=6a77f38946aaee1cd85eeec6cf4229b204c15071;hp=45994010399ecb0a475159b3747c3a7ff998cc6b;hpb=5273a3df6485dc2ad6aa7ddd441b9a21970f003b;p=linux-2.6.git

diff --git a/arch/sparc64/lib/U3memcpy.S b/arch/sparc64/lib/U3memcpy.S
index 459940103..8fe195a10 100644
--- a/arch/sparc64/lib/U3memcpy.S
+++ b/arch/sparc64/lib/U3memcpy.S
@@ -1,23 +1,63 @@
-/* $Id: U3memcpy.S,v 1.2 2000/11/01 09:29:19 davem Exp $
- * U3memcpy.S: UltraSparc-III optimized memcpy.
+/* U3memcpy.S: UltraSparc-III optimized memcpy.
  *
- * Copyright (C) 1999, 2000 David S. Miller (davem@redhat.com)
+ * Copyright (C) 1999, 2000, 2004 David S. Miller (davem@redhat.com)
  */
 
 #ifdef __KERNEL__
 #include <asm/visasm.h>
 #include <asm/asi.h>
-#include <asm/dcu.h>
-#include <asm/spitfire.h>
-#undef SMALL_COPY_USES_FPU
 #else
 #define ASI_BLK_P 0xf0
 #define FPRS_FEF  0x04
+#ifdef MEMCPY_DEBUG
+#define VISEntryHalf rd %fprs, %o5; wr %g0, FPRS_FEF, %fprs; \
+		     clr %g1; clr %g2; clr %g3; subcc %g0, %g0, %g0;
+#define VISExitHalf and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
+#else
 #define VISEntryHalf rd %fprs, %o5; wr %g0, FPRS_FEF, %fprs
 #define VISExitHalf and %o5, FPRS_FEF, %o5; wr %o5, 0x0, %fprs
-#define SMALL_COPY_USES_FPU
+#endif
+#endif
+
+#ifndef EX_LD
+#define EX_LD(x)	x
+#endif
+
+#ifndef EX_ST
+#define EX_ST(x)	x
+#endif
+
+#ifndef EX_RETVAL
+#define EX_RETVAL(x)	x
+#endif
+
+#ifndef LOAD
+#define LOAD(type,addr,dest)	type [addr], dest
+#endif
+
+#ifndef STORE
+#define STORE(type,src,addr)	type src, [addr]
+#endif
+
+#ifndef STORE_BLK
+#define STORE_BLK(src,addr)	stda src, [addr] ASI_BLK_P
+#endif
+
+#ifndef FUNC_NAME
+#define FUNC_NAME	U3memcpy
+#endif
+
+#ifndef PREAMBLE
+#define PREAMBLE
+#endif
+
+#ifndef XCC
+#define XCC xcc
 #endif
 
+	.register	%g2,#scratch
+	.register	%g3,#scratch
+
 	/* Special/non-trivial issues of this code:
 	 *
 	 * 1) %o5 is preserved from VISEntryHalf to VISExitHalf
@@ -29,7 +69,7 @@
 	 */
 
 	.text
-	.align	32
+	.align		64
 
 	/* The cheetah's flexible spine, oversized liver, enlarged heart,
 	 * slender muscular body, and claws make it the swiftest hunter
@@ -37,392 +77,344 @@
 	 * of up to 2.4GB per second.
 	 */
 
-	.globl		U3memcpy
-U3memcpy: /* %o0=dst, %o1=src, %o2=len */
-#ifndef __KERNEL__
-	/* Save away original 'dst' for memcpy return value. */
-	mov		%o0, %g3			! A0	Group
-#endif
-	/* Anything to copy at all? */
-	cmp		%o2, 0				! A1
-	ble,pn		%icc, U3memcpy_short_ret	! BR
-
-	/* Extremely small copy? */
-	 cmp		%o2, 31				! A0	Group
-	ble,pn		%icc, U3memcpy_short		! BR
-
-	/* Large enough to use unrolled prefetch loops? */
-	 cmp		%o2, 0x100			! A1
-	bge,a,pt	%icc, U3memcpy_enter		! BR	Group
-	 andcc		%o0, 0x3f, %g2			! A0
-
-	ba,pt		%xcc, U3memcpy_toosmall		! BR	Group
-	 andcc		%o0, 0x7, %g2			! A0
-
-	.align		32
-U3memcpy_short:
-	/* Copy %o2 bytes from src to dst, one byte at a time. */
-	ldub		[%o1 + 0x00], %o3		! MS	Group
-	add		%o1, 0x1, %o1			! A0
-	add		%o0, 0x1, %o0			! A1
-	subcc		%o2, 1, %o2			! A0	Group
-
-	bg,pt		%icc, U3memcpy_short		! BR
-	 stb		%o3, [%o0 + -1]			! MS	Group (1-cycle stall)
-
-U3memcpy_short_ret:
-#ifdef __KERNEL__
-	retl						! BR	Group (0-4 cycle stall)
-	 clr		%o0				! A0
-#else
-	retl						! BR	Group (0-4 cycle stall)
-	 mov		%g3, %o0			! A0
-#endif
-
-	/* Here len >= (6 * 64) and condition codes reflect execution
-	 * of "andcc %o0, 0x7, %g2", done by caller.
+	.globl	FUNC_NAME
+	.type	FUNC_NAME,#function
+FUNC_NAME:	/* %o0=dst, %o1=src, %o2=len */
+	srlx		%o2, 31, %g2
+	cmp		%g2, 0
+	tne		%xcc, 5
+	PREAMBLE
+	mov		%o0, %g5
+	cmp		%o2, 0
+	be,pn		%XCC, 85f
+	 or		%o0, %o1, %o3
+	cmp		%o2, 16
+	blu,a,pn	%XCC, 80f
+	 or		%o3, %o2, %o3
+
+	cmp		%o2, (3 * 64)
+	blu,pt		%XCC, 70f
+	 andcc		%o3, 0x7, %g0
+
+	/* Clobbers o5/g1/g2/g3/g7/icc/xcc.  We must preserve
+	 * o5 from here until we hit VISExitHalf.
 	 */
-	.align		64
-U3memcpy_enter:
+	VISEntryHalf
+
 	/* Is 'dst' already aligned on an 64-byte boundary? */
-	be,pt		%xcc, 2f			! BR
+	andcc		%o0, 0x3f, %g2
+	be,pt		%XCC, 2f
 
 	/* Compute abs((dst & 0x3f) - 0x40) into %g2.  This is the number
 	 * of bytes to copy to make 'dst' 64-byte aligned.  We pre-
 	 * subtract this from 'len'.
 	 */
-	 sub		%g2, 0x40, %g2			! A0	Group
-	sub		%g0, %g2, %g2			! A0	Group
-	sub		%o2, %g2, %o2			! A0	Group
+	 sub		%o0, %o1, %o4
+	sub		%g2, 0x40, %g2
+	sub		%g0, %g2, %g2
+	sub		%o2, %g2, %o2
+	andcc		%g2, 0x7, %g1
+	be,pt		%icc, 2f
+	 and		%g2, 0x38, %g2
+
+1:	subcc		%g1, 0x1, %g1
+	EX_LD(LOAD(ldub, %o1 + 0x00, %o3))
+	EX_ST(STORE(stb, %o3, %o1 + %o4))
+	bgu,pt		%XCC, 1b
+	 add		%o1, 0x1, %o1
 
-	/* Copy %g2 bytes from src to dst, one byte at a time. */
-1:	ldub		[%o1 + 0x00], %o3		! MS	(Group)
-	add		%o1, 0x1, %o1			! A1
-	add		%o0, 0x1, %o0			! A0	Group
-	subcc		%g2, 0x1, %g2			! A1
+	add		%o1, %o4, %o0
 
-	bg,pt		%icc, 1b			! BR	Group
-	 stb		%o3, [%o0 + -1]			! MS	Group
+2:	cmp		%g2, 0x0
+	and		%o1, 0x7, %g1
+	be,pt		%icc, 3f
+	 alignaddr	%o1, %g0, %o1
 
-2:	VISEntryHalf					! MS+MS
-	and		%o1, 0x7, %g1			! A1
-	ba,pt		%xcc, U3memcpy_begin		! BR
-	 alignaddr	%o1, %g0, %o1			! MS	      (Break-after)
+	EX_LD(LOAD(ldd, %o1, %f4))
+1:	EX_LD(LOAD(ldd, %o1 + 0x8, %f6))
+	add		%o1, 0x8, %o1
+	subcc		%g2, 0x8, %g2
+	faligndata	%f4, %f6, %f0
+	EX_ST(STORE(std, %f0, %o0))
+	be,pn		%icc, 3f
+	 add		%o0, 0x8, %o0
+
+	EX_LD(LOAD(ldd, %o1 + 0x8, %f4))
+	add		%o1, 0x8, %o1
+	subcc		%g2, 0x8, %g2
+	faligndata	%f6, %f4, %f2
+	EX_ST(STORE(std, %f2, %o0))
+	bne,pt		%icc, 1b
+	 add		%o0, 0x8, %o0
+
+3:	LOAD(prefetch, %o1 + 0x000, #one_read)
+	LOAD(prefetch, %o1 + 0x040, #one_read)
+	andn		%o2, (0x40 - 1), %o4
+	LOAD(prefetch, %o1 + 0x080, #one_read)
+	LOAD(prefetch, %o1 + 0x0c0, #one_read)
+	LOAD(prefetch, %o1 + 0x100, #one_read)
+	EX_LD(LOAD(ldd, %o1 + 0x000, %f0))
+	LOAD(prefetch, %o1 + 0x140, #one_read)
+	EX_LD(LOAD(ldd, %o1 + 0x008, %f2))
+	LOAD(prefetch, %o1 + 0x180, #one_read)
+	EX_LD(LOAD(ldd, %o1 + 0x010, %f4))
+	LOAD(prefetch, %o1 + 0x1c0, #one_read)
+	faligndata	%f0, %f2, %f16
+	EX_LD(LOAD(ldd, %o1 + 0x018, %f6))
+	faligndata	%f2, %f4, %f18
+	EX_LD(LOAD(ldd, %o1 + 0x020, %f8))
+	faligndata	%f4, %f6, %f20
+	EX_LD(LOAD(ldd, %o1 + 0x028, %f10))
+	faligndata	%f6, %f8, %f22
+
+	EX_LD(LOAD(ldd, %o1 + 0x030, %f12))
+	faligndata	%f8, %f10, %f24
+	EX_LD(LOAD(ldd, %o1 + 0x038, %f14))
+	faligndata	%f10, %f12, %f26
+	EX_LD(LOAD(ldd, %o1 + 0x040, %f0))
+
+	subcc		%o4, 0x80, %o4
+	add		%o1, 0x40, %o1
+	bgu,pt		%XCC, 1f
+	 srl		%o4, 6, %o3
+	ba,pt		%xcc, 2f
+	 nop
 
 	.align		64
-U3memcpy_begin:
-#ifdef __KERNEL__
-	.globl		U3memcpy_nop_1_6
-U3memcpy_nop_1_6:
-	ldxa		[%g0] ASI_DCU_CONTROL_REG, %g3
-	sethi		%uhi(DCU_PE), %o3
-	sllx		%o3, 32, %o3
-	or		%g3, %o3, %o3
-	stxa		%o3, [%g0] ASI_DCU_CONTROL_REG	! Enable P-cache
-	membar		#Sync
-#endif
-	prefetch	[%o1 + 0x000], #one_read	! MS	Group1
-	prefetch	[%o1 + 0x040], #one_read	! MS	Group2
-	andn		%o2, (0x40 - 1), %o4		! A0
-	prefetch	[%o1 + 0x080], #one_read	! MS	Group3
-	cmp		%o4, 0x140			! A0
-	prefetch	[%o1 + 0x0c0], #one_read	! MS	Group4
-	ldd		[%o1 + 0x000], %f0		! MS	Group5 (%f0 results at G8)
-	bge,a,pt	%icc, 1f			! BR
-
-	 prefetch	[%o1 + 0x100], #one_read	! MS	Group6
-1:	ldd		[%o1 + 0x008], %f2		! AX	       (%f2 results at G9)
-	cmp		%o4, 0x180			! A1
-	bge,a,pt	%icc, 1f			! BR
-	 prefetch	[%o1 + 0x140], #one_read	! MS	Group7
-1:	ldd		[%o1 + 0x010], %f4		! AX	       (%f4 results at G10)
-	cmp		%o4, 0x1c0			! A1
-	bge,a,pt	%icc, 1f			! BR
-
-	 prefetch	[%o1 + 0x180], #one_read	! MS	Group8
-1:	faligndata	%f0, %f2, %f16			! FGA	Group9 (%f16 at G12)
-	ldd		[%o1 + 0x018], %f6		! AX	       (%f6 results at G12)
-	faligndata	%f2, %f4, %f18			! FGA	Group10 (%f18 results at G13)
-	ldd		[%o1 + 0x020], %f8		! MS	        (%f8 results at G13)
-	faligndata	%f4, %f6, %f20			! FGA	Group12	(1-cycle stall,%f20 at G15)
-	ldd		[%o1 + 0x028], %f10		! MS		(%f10 results at G15)
-	faligndata	%f6, %f8, %f22			! FGA	Group13 (%f22 results at G16)
-
-	ldd		[%o1 + 0x030], %f12		! MS		(%f12 results at G16)
-	faligndata	%f8, %f10, %f24			! FGA	Group15 (1-cycle stall,%f24 at G18)
-	ldd		[%o1 + 0x038], %f14		! MS		(%f14 results at G18)
-	faligndata	%f10, %f12, %f26		! FGA	Group16 (%f26 results at G19)
-	ldd		[%o1 + 0x040], %f0		! MS		(%f0 results at G19)
-
-	/* We only use the first loop if len > (7 * 64). */
-	subcc		%o4, 0x1c0, %o4			! A0	Group17
-	bg,pt		%icc, U3memcpy_loop1		! BR
-	 add		%o1, 0x40, %o1			! A1
-
-	add		%o4, 0x140, %o4			! A0	Group18
-	ba,pt		%xcc, U3memcpy_loop2		! BR
-	 srl		%o4, 6, %o3			! A0	Group19
-	nop
-	nop
-	nop
-	nop
-	nop
-
-	nop
-	nop
-
-	/* This loop performs the copy and queues new prefetches.
-	 * We drop into the second loop when len <= (5 * 64).  Note
-	 * that this (5 * 64) factor has been subtracted from len
-	 * already.
-	 */
-U3memcpy_loop1:
-	ldd		[%o1 + 0x008], %f2		! MS	Group2	(%f2 results at G5)
-	faligndata	%f12, %f14, %f28		! FGA		(%f28 results at G5)
-	ldd		[%o1 + 0x010], %f4		! MS	Group3	(%f4 results at G6)
-	faligndata	%f14, %f0, %f30			! FGA	Group4	(1-cycle stall, %f30 at G7)
-	stda		%f16, [%o0] ASI_BLK_P		! MS
-	ldd		[%o1 + 0x018], %f6		! AX		(%f6 results at G7)
-
-	faligndata	%f0, %f2, %f16			! FGA	Group12 (7-cycle stall)
-	ldd		[%o1 + 0x020], %f8		! MS	        (%f8 results at G15)
-	faligndata	%f2, %f4, %f18			! FGA	Group13	(%f18 results at G16)
-	ldd		[%o1 + 0x028], %f10		! MS		(%f10 results at G16)
-	faligndata	%f4, %f6, %f20			! FGA	Group14	(%f20 results at G17)
-	ldd		[%o1 + 0x030], %f12		! MS		(%f12 results at G17)
-	faligndata	%f6, %f8, %f22			! FGA	Group15	(%f22 results at G18)
-	ldd		[%o1 + 0x038], %f14		! MS		(%f14 results at G18)
-
-	faligndata	%f8, %f10, %f24			! FGA	Group16	(%f24 results at G19)
-	ldd		[%o1 + 0x040], %f0		! AX		(%f0 results at G19)
-	prefetch	[%o1 + 0x180], #one_read	! MS
-	faligndata	%f10, %f12, %f26		! FGA	Group17	(%f26 results at G20)
-	subcc		%o4, 0x40, %o4			! A0
-	add		%o1, 0x40, %o1			! A1
-	bg,pt		%xcc, U3memcpy_loop1		! BR
-	 add		%o0, 0x40, %o0			! A0	Group18
-
-U3memcpy_loop2_enter:
-	mov		5, %o3				! A1
-
-	/* This loop performs on the copy, no new prefetches are
-	 * queued.  We do things this way so that we do not perform
-	 * any spurious prefetches past the end of the src buffer.
-	 */
-U3memcpy_loop2:
-	ldd		[%o1 + 0x008], %f2		! MS
-	faligndata	%f12, %f14, %f28		! FGA	Group2
-	ldd		[%o1 + 0x010], %f4		! MS
-	faligndata	%f14, %f0, %f30			! FGA	Group4	(1-cycle stall)
-	stda		%f16, [%o0] ASI_BLK_P		! MS
-	ldd		[%o1 + 0x018], %f6		! AX
-	faligndata	%f0, %f2, %f16			! FGA	Group12 (7-cycle stall)
-
-	ldd		[%o1 + 0x020], %f8		! MS
-	faligndata	%f2, %f4, %f18			! FGA	Group13
-	ldd		[%o1 + 0x028], %f10		! MS
-	faligndata	%f4, %f6, %f20			! FGA	Group14
-	ldd		[%o1 + 0x030], %f12		! MS
-	faligndata	%f6, %f8, %f22			! FGA	Group15
-	ldd		[%o1 + 0x038], %f14		! MS
-	faligndata	%f8, %f10, %f24			! FGA	Group16
-
-	ldd		[%o1 + 0x040], %f0		! AX
-	faligndata	%f10, %f12, %f26		! FGA	Group17
-	subcc		%o3, 0x01, %o3			! A0
-	add		%o1, 0x40, %o1			! A1
-	bg,pt		%xcc, U3memcpy_loop2		! BR
-	 add		%o0, 0x40, %o0			! A0	Group18
+1:
+	EX_LD(LOAD(ldd, %o1 + 0x008, %f2))
+	faligndata	%f12, %f14, %f28
+	EX_LD(LOAD(ldd, %o1 + 0x010, %f4))
+	faligndata	%f14, %f0, %f30
+	EX_ST(STORE_BLK(%f16, %o0))
+	EX_LD(LOAD(ldd, %o1 + 0x018, %f6))
+	faligndata	%f0, %f2, %f16
+	add		%o0, 0x40, %o0
+
+	EX_LD(LOAD(ldd, %o1 + 0x020, %f8))
+	faligndata	%f2, %f4, %f18
+	EX_LD(LOAD(ldd, %o1 + 0x028, %f10))
+	faligndata	%f4, %f6, %f20
+	EX_LD(LOAD(ldd, %o1 + 0x030, %f12))
+	subcc		%o3, 0x01, %o3
+	faligndata	%f6, %f8, %f22
+	EX_LD(LOAD(ldd, %o1 + 0x038, %f14))
+
+	faligndata	%f8, %f10, %f24
+	EX_LD(LOAD(ldd, %o1 + 0x040, %f0))
+	LOAD(prefetch, %o1 + 0x1c0, #one_read)
+	faligndata	%f10, %f12, %f26
+	bg,pt		%XCC, 1b
+	 add		%o1, 0x40, %o1
 
 	/* Finally we copy the last full 64-byte block. */
-U3memcpy_loopfini:
-	ldd		[%o1 + 0x008], %f2		! MS
-	faligndata	%f12, %f14, %f28		! FGA
-	ldd		[%o1 + 0x010], %f4		! MS	Group19
-	faligndata	%f14, %f0, %f30			! FGA
-	stda		%f16, [%o0] ASI_BLK_P		! MS	Group20
-	ldd		[%o1 + 0x018], %f6		! AX
-	faligndata	%f0, %f2, %f16			! FGA	Group11 (7-cycle stall)
-	ldd		[%o1 + 0x020], %f8		! MS
-	faligndata	%f2, %f4, %f18			! FGA	Group12
-	ldd		[%o1 + 0x028], %f10		! MS
-	faligndata	%f4, %f6, %f20			! FGA	Group13
-	ldd		[%o1 + 0x030], %f12		! MS
-	faligndata	%f6, %f8, %f22			! FGA	Group14
-	ldd		[%o1 + 0x038], %f14		! MS
-	faligndata	%f8, %f10, %f24			! FGA	Group15
-	cmp		%g1, 0				! A0
-	be,pt		%icc, 1f			! BR
-	 add		%o0, 0x40, %o0			! A1
-	ldd		[%o1 + 0x040], %f0		! MS
-1:	faligndata	%f10, %f12, %f26		! FGA	Group16
-	faligndata	%f12, %f14, %f28		! FGA	Group17
-	faligndata	%f14, %f0, %f30			! FGA	Group18
-	stda		%f16, [%o0] ASI_BLK_P		! MS
-	add		%o0, 0x40, %o0			! A0
-	add		%o1, 0x40, %o1			! A1
-#ifdef __KERNEL__
-	.globl		U3memcpy_nop_2_3
-U3memcpy_nop_2_3:
-	mov		PRIMARY_CONTEXT, %o3
-	stxa		%g0, [%o3] ASI_DMMU		! Flush P-cache
-	stxa		%g3, [%g0] ASI_DCU_CONTROL_REG	! Disable P-cache
-#endif
-	membar		#Sync				! MS	Group26 (7-cycle stall)
+2:
+	EX_LD(LOAD(ldd, %o1 + 0x008, %f2))
+	faligndata	%f12, %f14, %f28
+	EX_LD(LOAD(ldd, %o1 + 0x010, %f4))
+	faligndata	%f14, %f0, %f30
+	EX_ST(STORE_BLK(%f16, %o0))
+	EX_LD(LOAD(ldd, %o1 + 0x018, %f6))
+	faligndata	%f0, %f2, %f16
+	EX_LD(LOAD(ldd, %o1 + 0x020, %f8))
+	faligndata	%f2, %f4, %f18
+	EX_LD(LOAD(ldd, %o1 + 0x028, %f10))
+	faligndata	%f4, %f6, %f20
+	EX_LD(LOAD(ldd, %o1 + 0x030, %f12))
+	faligndata	%f6, %f8, %f22
+	EX_LD(LOAD(ldd, %o1 + 0x038, %f14))
+	faligndata	%f8, %f10, %f24
+	cmp		%g1, 0
+	be,pt		%XCC, 1f
+	 add		%o0, 0x40, %o0
+	EX_LD(LOAD(ldd, %o1 + 0x040, %f0))
+1:	faligndata	%f10, %f12, %f26
+	faligndata	%f12, %f14, %f28
+	faligndata	%f14, %f0, %f30
+	EX_ST(STORE_BLK(%f16, %o0))
+	add		%o0, 0x40, %o0
+	add		%o1, 0x40, %o1
+	membar		#Sync
 
 	/* Now we copy the (len modulo 64) bytes at the end.
 	 * Note how we borrow the %f0 loaded above.
 	 *
 	 * Also notice how this code is careful not to perform a
-	 * load past the end of the src buffer just like similar
-	 * code found in U3memcpy_toosmall processing.
+	 * load past the end of the src buffer.
 	 */
-U3memcpy_loopend:
-	and		%o2, 0x3f, %o2			! A0	Group
-	andcc		%o2, 0x38, %g2			! A0	Group
-	be,pn		%icc, U3memcpy_endcruft		! BR
-	 subcc		%g2, 0x8, %g2			! A1
-	be,pn		%icc, U3memcpy_endcruft		! BR	Group
-	 cmp		%g1, 0				! A0
-
-	be,a,pt		%icc, 1f			! BR	Group
-	 ldd		[%o1 + 0x00], %f0		! MS
-
-1:	ldd		[%o1 + 0x08], %f2		! MS	Group
-	add		%o1, 0x8, %o1			! A0
-	sub		%o2, 0x8, %o2			! A1
-	subcc		%g2, 0x8, %g2			! A0	Group
-	faligndata	%f0, %f2, %f8			! FGA	Group
-	std		%f8, [%o0 + 0x00]		! MS	(XXX does it stall here? XXX)
-	be,pn		%icc, U3memcpy_endcruft		! BR
-	 add		%o0, 0x8, %o0			! A0
-	ldd		[%o1 + 0x08], %f0		! MS	Group
-	add		%o1, 0x8, %o1			! A0
-	sub		%o2, 0x8, %o2			! A1
-	subcc		%g2, 0x8, %g2			! A0	Group
-	faligndata	%f2, %f0, %f8			! FGA
-	std		%f8, [%o0 + 0x00]		! MS	(XXX does it stall here? XXX)
-	bne,pn		%icc, 1b			! BR
-	 add		%o0, 0x8, %o0			! A0	Group
+	and		%o2, 0x3f, %o2
+	andcc		%o2, 0x38, %g2
+	be,pn		%XCC, 2f
+	 subcc		%g2, 0x8, %g2
+	be,pn		%XCC, 2f
+	 cmp		%g1, 0
+
+	sub		%o2, %g2, %o2
+	be,a,pt		%XCC, 1f
+	 EX_LD(LOAD(ldd, %o1 + 0x00, %f0))
+
+1:	EX_LD(LOAD(ldd, %o1 + 0x08, %f2))
+	add		%o1, 0x8, %o1
+	subcc		%g2, 0x8, %g2
+	faligndata	%f0, %f2, %f8
+	EX_ST(STORE(std, %f8, %o0))
+	be,pn		%XCC, 2f
+	 add		%o0, 0x8, %o0
+	EX_LD(LOAD(ldd, %o1 + 0x08, %f0))
+	add		%o1, 0x8, %o1
+	subcc		%g2, 0x8, %g2
+	faligndata	%f2, %f0, %f8
+	EX_ST(STORE(std, %f8, %o0))
+	bne,pn		%XCC, 1b
+	 add		%o0, 0x8, %o0
 
 	/* If anything is left, we copy it one byte at a time.
 	 * Note that %g1 is (src & 0x3) saved above before the
 	 * alignaddr was performed.
 	 */
-U3memcpy_endcruft:
+2:
 	cmp		%o2, 0
 	add		%o1, %g1, %o1
 	VISExitHalf
-	be,pn		%icc, U3memcpy_short_ret
-	 nop
-	ba,a,pt		%xcc, U3memcpy_short
-
-	/* If we get here, then 32 <= len < (6 * 64) */
-U3memcpy_toosmall:
-
-#ifdef SMALL_COPY_USES_FPU
+	be,pn		%XCC, 85f
+	 sub		%o0, %o1, %o3
 
-	/* Is 'dst' already aligned on an 8-byte boundary? */
-	be,pt		%xcc, 2f			! BR	Group
+	andcc		%g1, 0x7, %g0
+	bne,pn		%icc, 90f
+	 andcc		%o2, 0x8, %g0
+	be,pt		%icc, 1f
+	 nop
+	EX_LD(LOAD(ldx, %o1, %o5))
+	EX_ST(STORE(stx, %o5, %o1 + %o3))
+	add		%o1, 0x8, %o1
 
-	/* Compute abs((dst & 7) - 8) into %g2.  This is the number
-	 * of bytes to copy to make 'dst' 8-byte aligned.  We pre-
-	 * subtract this from 'len'.
-	 */
-	 sub		%g2, 0x8, %g2			! A0
-	sub		%g0, %g2, %g2			! A0	Group (reg-dep)
-	sub		%o2, %g2, %o2			! A0	Group (reg-dep)
+1:	andcc		%o2, 0x4, %g0
+	be,pt		%icc, 1f
+	 nop
+	EX_LD(LOAD(lduw, %o1, %o5))
+	EX_ST(STORE(stw, %o5, %o1 + %o3))
+	add		%o1, 0x4, %o1
 
-	/* Copy %g2 bytes from src to dst, one byte at a time. */
-1:	ldub		[%o1 + 0x00], %o3		! MS	(Group) (%o3 in 3 cycles)
-	add		%o1, 0x1, %o1			! A1
-	add		%o0, 0x1, %o0			! A0	Group
-	subcc		%g2, 0x1, %g2			! A1
+1:	andcc		%o2, 0x2, %g0
+	be,pt		%icc, 1f
+	 nop
+	EX_LD(LOAD(lduh, %o1, %o5))
+	EX_ST(STORE(sth, %o5, %o1 + %o3))
+	add		%o1, 0x2, %o1
 
-	bg,pt		%icc, 1b			! BR	Group
-	 stb		%o3, [%o0 + -1]			! MS	Group
+1:	andcc		%o2, 0x1, %g0
+	be,pt		%icc, 85f
+	 nop
+	EX_LD(LOAD(ldub, %o1, %o5))
+	ba,pt		%xcc, 85f
+	 EX_ST(STORE(stb, %o5, %o1 + %o3))
 
-2:	VISEntryHalf					! MS+MS
+	.align		64
+70: /* 16 < len <= 64 */
+	bne,pn		%XCC, 75f
+	 sub		%o0, %o1, %o3
+
+72:
+	andn		%o2, 0xf, %o4
+	and		%o2, 0xf, %o2
+1:	subcc		%o4, 0x10, %o4
+	EX_LD(LOAD(ldx, %o1 + 0x00, %o5))
+	EX_LD(LOAD(ldx, %o1 + 0x08, %g1))
+	EX_ST(STORE(stx, %o5, %o1 + %o3))
+	add		%o1, 0x8, %o1
+	EX_ST(STORE(stx, %g1, %o1 + %o3))
+	bgu,pt		%XCC, 1b
+	 add		%o1, 0x8, %o1
+73:	andcc		%o2, 0x8, %g0
+	be,pt		%XCC, 1f
+	 nop
+	sub		%o2, 0x8, %o2
+	EX_LD(LOAD(ldx, %o1, %o5))
+	EX_ST(STORE(stx, %o5, %o1 + %o3))
+	add		%o1, 0x8, %o1
+1:	andcc		%o2, 0x4, %g0
+	be,pt		%XCC, 1f
+	 nop
+	sub		%o2, 0x4, %o2
+	EX_LD(LOAD(lduw, %o1, %o5))
+	EX_ST(STORE(stw, %o5, %o1 + %o3))
+	add		%o1, 0x4, %o1
+1:	cmp		%o2, 0
+	be,pt		%XCC, 85f
+	 nop
+	ba,pt		%xcc, 90f
+	 nop
 
-	/* Compute (len - (len % 8)) into %g2.  This is guaranteed
-	 * to be nonzero.
-	 */
-	andn		%o2, 0x7, %g2			! A0	Group
-
-	/* You may read this and believe that it allows reading
-	 * one 8-byte longword past the end of src.  It actually
-	 * does not, as %g2 is subtracted as loads are done from
-	 * src, so we always stop before running off the end.
-	 * Also, we are guaranteed to have at least 0x10 bytes
-	 * to move here.
-	 */
-	sub		%g2, 0x8, %g2			! A0	Group (reg-dep)
-	alignaddr	%o1, %g0, %g1			! MS	      (Break-after)
-	ldd		[%g1 + 0x00], %f0		! MS	Group (1-cycle stall)
-	add		%g1, 0x8, %g1			! A0
-
-1:	ldd		[%g1 + 0x00], %f2		! MS	Group
-	add		%g1, 0x8, %g1			! A0
-	sub		%o2, 0x8, %o2			! A1
-	subcc		%g2, 0x8, %g2			! A0	Group
-
-	faligndata	%f0, %f2, %f8			! FGA	Group (1-cycle stall)
-	std		%f8, [%o0 + 0x00]		! MS	Group (2-cycle stall)
-	add		%o1, 0x8, %o1			! A0
-	be,pn		%icc, 2f			! BR
-
-	 add		%o0, 0x8, %o0			! A1
-	ldd		[%g1 + 0x00], %f0		! MS	Group
-	add		%g1, 0x8, %g1			! A0
-	sub		%o2, 0x8, %o2			! A1
-
-	subcc		%g2, 0x8, %g2			! A0	Group
-	faligndata	%f2, %f0, %f8			! FGA	Group (1-cycle stall)
-	std		%f8, [%o0 + 0x00]		! MS	Group (2-cycle stall)
-	add		%o1, 0x8, %o1			! A0
-
-	bne,pn		%icc, 1b			! BR
-	 add		%o0, 0x8, %o0			! A1
-
-	/* Nothing left to copy? */
-2:	cmp		%o2, 0				! A0	Group
-	VISExitHalf					! A0+MS
-	be,pn		%icc, U3memcpy_short_ret	! BR	Group
-	 nop						! A0
-	ba,a,pt		%xcc, U3memcpy_short		! BR	Group
-
-#else /* !(SMALL_COPY_USES_FPU) */
-
-	xor		%o1, %o0, %g2
-	andcc		%g2, 0x7, %g0
-	bne,pn		%icc, U3memcpy_short
-	 andcc		%o1, 0x7, %g2
-
-	be,pt		%xcc, 2f
-	 sub		%g2, 0x8, %g2
-	sub		%g0, %g2, %g2
-	sub		%o2, %g2, %o2
+75:
+	andcc		%o0, 0x7, %g1
+	sub		%g1, 0x8, %g1
+	be,pn		%icc, 2f
+	 sub		%g0, %g1, %g1
+	sub		%o2, %g1, %o2
+
+1:	subcc		%g1, 1, %g1
+	EX_LD(LOAD(ldub, %o1, %o5))
+	EX_ST(STORE(stb, %o5, %o1 + %o3))
+	bgu,pt		%icc, 1b
+	 add		%o1, 1, %o1
+
+2:	add		%o1, %o3, %o0
+	andcc		%o1, 0x7, %g1
+	bne,pt		%icc, 8f
+	 sll		%g1, 3, %g1
+
+	cmp		%o2, 16
+	bgeu,pt		%icc, 72b
+	 nop
+	ba,a,pt		%xcc, 73b
+
+8:	mov		64, %o3
+	andn		%o1, 0x7, %o1
+	EX_LD(LOAD(ldx, %o1, %g2))
+	sub		%o3, %g1, %o3
+	andn		%o2, 0x7, %o4
+	sllx		%g2, %g1, %g2
+1:	EX_LD(LOAD(ldx, %o1 + 0x8, %g3))
+	subcc		%o4, 0x8, %o4
+	add		%o1, 0x8, %o1
+	srlx		%g3, %o3, %o5
+	or		%o5, %g2, %o5
+	EX_ST(STORE(stx, %o5, %o0))
+	add		%o0, 0x8, %o0
+	bgu,pt		%icc, 1b
+	 sllx		%g3, %g1, %g2
 
-1:	ldub		[%o1 + 0x00], %o3
-	add		%o1, 0x1, %o1
-	add		%o0, 0x1, %o0
-	subcc		%g2, 0x1, %g2
-	bg,pt		%icc, 1b
-	 stb		%o3, [%o0 + -1]
+	srl		%g1, 3, %g1
+	andcc		%o2, 0x7, %o2
+	be,pn		%icc, 85f
+	 add		%o1, %g1, %o1
+	ba,pt		%xcc, 90f
+	 sub		%o0, %o1, %o3
 
-2:	andn		%o2, 0x7, %g2
-	sub		%o2, %g2, %o2
+	.align		64
+80: /* 0 < len <= 16 */
+	andcc		%o3, 0x3, %g0
+	bne,pn		%XCC, 90f
+	 sub		%o0, %o1, %o3
 
-3:	ldx		[%o1 + 0x00], %o3
-	add		%o1, 0x8, %o1
-	add		%o0, 0x8, %o0
-	subcc		%g2, 0x8, %g2
-	bg,pt		%icc, 3b
-	 stx		%o3, [%o0 + -8]
+1:
+	subcc		%o2, 4, %o2
+	EX_LD(LOAD(lduw, %o1, %g1))
+	EX_ST(STORE(stw, %g1, %o1 + %o3))
+	bgu,pt		%XCC, 1b
+	 add		%o1, 4, %o1
 
-	cmp		%o2, 0
-	bne,pn		%icc, U3memcpy_short
-	 nop
-	ba,a,pt		%xcc, U3memcpy_short_ret
+85:	retl
+	 mov		EX_RETVAL(%g5), %o0
 
-#endif /* !(SMALL_COPY_USES_FPU) */
+	.align		32
+90:
+	subcc		%o2, 1, %o2
+	EX_LD(LOAD(ldub, %o1, %g1))
+	EX_ST(STORE(stb, %g1, %o1 + %o3))
+	bgu,pt		%XCC, 90b
+	 add		%o1, 1, %o1
+	retl
+	 mov		EX_RETVAL(%g5), %o0
+
+	.size		FUNC_NAME, .-FUNC_NAME