(memcpy): Optimize better for smaller than 256 byte copies. Also, use only one unrolled loop instead of two for the large copy case.

1 files changed, 231 insertions, 336 deletions

diff --git a/sysdeps/sparc/sparc64/sparcv9b/memcpy.S b/sysdeps/sparc/sparc64/sparcv9b/memcpy.S
index ccbc3fd..8b70b0a 100644
--- a/sysdeps/sparc/sparc64/sparcv9b/memcpy.S
+++ b/sysdeps/sparc/sparc64/sparcv9b/memcpy.S
@@ -20,32 +20,36 @@
    02111-1307 USA.  */
 
 #include <sysdep.h>
+
 #define ASI_BLK_P 0xf0
 #define FPRS_FEF  0x04
 #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
 
 #ifndef XCC
 #define USE_BPR
 #define XCC xcc
 #endif
 
+	.register	%g2,#scratch
+	.register	%g3,#scratch
+	.register	%g6,#scratch
+
 	.text
 	.align	32
 
 ENTRY(bcopy)
-	sub		%o1, %o0, %o4			/* IEU0		Group		*/
-	mov		%o0, %g3			/* IEU1				*/
-	cmp		%o4, %o2			/* IEU1		Group		*/
-	mov		%o1, %o0			/* IEU0				*/
-	bgeu,pt		%XCC, 100f			/* CTI				*/
-	 mov		%g3, %o1			/* IEU0		Group		*/
+	sub		%o1, %o0, %o4
+	mov		%o0, %g4
+	cmp		%o4, %o2
+	mov		%o1, %o0
+	bgeu,pt		%XCC, 100f
+	 mov		%g4, %o1
 #ifndef USE_BPR
-	srl		%o2, 0, %o2			/* IEU1				*/
+	srl		%o2, 0, %o2
 #endif
-	brnz,pn		%o2, 220f			/* CTI		Group		*/
-	 add		%o0, %o2, %o0			/* IEU0				*/
+	brnz,pn		%o2, 220f
+	 add		%o0, %o2, %o0
 	retl
 	 nop
 END(bcopy)
@@ -68,373 +72,264 @@ END(bcopy)
 ENTRY(memcpy)
 
 100: /* %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		%XCC, 102f			! BR
-
-	/* Extremely small copy? */
-218:	 cmp		%o2, 31				! A0	Group
-	ble,pn		%XCC, 101f			! BR
-
-	/* Large enough to use unrolled prefetch loops? */
-	 cmp		%o2, 0x100			! A1
-	bge,a,pt	%XCC, 103f			! BR	Group
-	 andcc		%o0, 0x3f, %g2			! A0
+	mov		%o0, %g5
+	cmp		%o2, 0
+	be,pn		%XCC, out
+218:	 or		%o0, %o1, %o3
+	cmp		%o2, 16
+	bleu,a,pn	%XCC, small_copy
+	 or		%o3, %o2, %o3
 
-	ba,pt		%XCC, 108f			! BR	Group
-	 andcc		%o0, 0x7, %g2			! A0
+	cmp		%o2, 256
+	blu,pt		%XCC, medium_copy
+	 andcc		%o3, 0x7, %g0
 
-	.align		32
-101:
-	/* 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		%XCC, 101b			! BR
-	 stb		%o3, [%o0 + -1]			! MS	Group (1-cycle stall)
-
-102:
-#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
+	ba,pt		%xcc, enter
+	 andcc		%o0, 0x3f, %g2
 
-	/* Here len >= (6 * 64) and condition codes reflect execution
+	/* Here len >= 256 and condition codes reflect execution
 	 * of "andcc %o0, 0x7, %g2", done by caller.
 	 */
 	.align		64
-103:
+enter:
 	/* Is 'dst' already aligned on an 64-byte boundary? */
-	be,pt		%XCC, 2f			! BR
+	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		%g2, 0x40, %g2
+	sub		%g0, %g2, %g2
+	sub		%o2, %g2, %o2
 
 	/* 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
+1:	ldub		[%o1 + 0x00], %o3
+	add		%o1, 0x1, %o1
+	add		%o0, 0x1, %o0
+	subcc		%g2, 0x1, %g2
 
-	bg,pt		%XCC, 1b			! BR	Group
-	 stb		%o3, [%o0 + -1]			! MS	Group
+	bg,pt		%XCC, 1b
+	 stb		%o3, [%o0 + -1]
 
-2:	VISEntryHalf					! MS+MS
-	and		%o1, 0x7, %g1			! A1
-	ba,pt		%XCC, 104f			! BR
-	 alignaddr	%o1, %g0, %o1			! MS	      (Break-after)
+2:	VISEntryHalf
+	and		%o1, 0x7, %g1
+	ba,pt		%xcc, begin
+	 alignaddr	%o1, %g0, %o1
 
 	.align		64
-104:
-	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	%XCC, 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	%XCC, 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	%XCC, 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		%XCC, 105f			! BR
-	 add		%o1, 0x40, %o1			! A1
-
-	add		%o4, 0x140, %o4			! A0	Group18
-	ba,pt		%XCC, 106f			! 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.
-	 */
-105:
-	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, 105b			! BR
-	 add		%o0, 0x40, %o0			! A0	Group18
-
-	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.
-	 */
-106:
-	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, 106b			! BR
-	 add		%o0, 0x40, %o0			! A0	Group18
+begin:
+	prefetch	[%o1 + 0x000], #one_read
+	prefetch	[%o1 + 0x040], #one_read
+	andn		%o2, (0x40 - 1), %o4
+	prefetch	[%o1 + 0x080], #one_read
+	prefetch	[%o1 + 0x0c0], #one_read
+	ldd		[%o1 + 0x000], %f0
+	prefetch	[%o1 + 0x100], #one_read
+	ldd		[%o1 + 0x008], %f2
+	prefetch	[%o1 + 0x140], #one_read
+	ldd		[%o1 + 0x010], %f4
+	prefetch	[%o1 + 0x180], #one_read
+	faligndata	%f0, %f2, %f16
+	ldd		[%o1 + 0x018], %f6
+	faligndata	%f2, %f4, %f18
+	ldd		[%o1 + 0x020], %f8
+	faligndata	%f4, %f6, %f20
+	ldd		[%o1 + 0x028], %f10
+	faligndata	%f6, %f8, %f22
+
+	ldd		[%o1 + 0x030], %f12
+	faligndata	%f8, %f10, %f24
+	ldd		[%o1 + 0x038], %f14
+	faligndata	%f10, %f12, %f26
+	ldd		[%o1 + 0x040], %f0
+
+	sub		%o4, 0x80, %o4
+	add		%o1, 0x40, %o1
+	ba,pt		%xcc, loop
+	 srl		%o4, 6, %o3
+
+	.align		64
+loop:
+	ldd		[%o1 + 0x008], %f2
+	faligndata	%f12, %f14, %f28
+	ldd		[%o1 + 0x010], %f4
+	faligndata	%f14, %f0, %f30
+	stda		%f16, [%o0] ASI_BLK_P
+	ldd		[%o1 + 0x018], %f6
+	faligndata	%f0, %f2, %f16
+
+	ldd		[%o1 + 0x020], %f8
+	faligndata	%f2, %f4, %f18
+	ldd		[%o1 + 0x028], %f10
+	faligndata	%f4, %f6, %f20
+	ldd		[%o1 + 0x030], %f12
+	faligndata	%f6, %f8, %f22
+	ldd		[%o1 + 0x038], %f14
+	faligndata	%f8, %f10, %f24
+
+	ldd		[%o1 + 0x040], %f0
+	prefetch	[%o1 + 0x180], #one_read
+	faligndata	%f10, %f12, %f26
+	subcc		%o3, 0x01, %o3
+	add		%o1, 0x40, %o1
+	bg,pt		%XCC, loop
+	 add		%o0, 0x40, %o0
 
 	/* Finally we copy the last full 64-byte block. */
-	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		%XCC, 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
-	membar		#Sync				! MS	Group26 (7-cycle stall)
+loopfini:
+	ldd		[%o1 + 0x008], %f2
+	faligndata	%f12, %f14, %f28
+	ldd		[%o1 + 0x010], %f4
+	faligndata	%f14, %f0, %f30
+	stda		%f16, [%o0] ASI_BLK_P
+	ldd		[%o1 + 0x018], %f6
+	faligndata	%f0, %f2, %f16
+	ldd		[%o1 + 0x020], %f8
+	faligndata	%f2, %f4, %f18
+	ldd		[%o1 + 0x028], %f10
+	faligndata	%f4, %f6, %f20
+	ldd		[%o1 + 0x030], %f12
+	faligndata	%f6, %f8, %f22
+	ldd		[%o1 + 0x038], %f14
+	faligndata	%f8, %f10, %f24
+	cmp		%g1, 0
+	be,pt		%XCC, 1f
+	 add		%o0, 0x40, %o0
+	ldd		[%o1 + 0x040], %f0
+1:	faligndata	%f10, %f12, %f26
+	faligndata	%f12, %f14, %f28
+	faligndata	%f14, %f0, %f30
+	stda		%f16, [%o0] ASI_BLK_P
+	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 'toosmall' processing.
+	 * load past the end of the src buffer.
 	 */
-	and		%o2, 0x3f, %o2			! A0	Group
-	andcc		%o2, 0x38, %g2			! A0	Group
-	be,pn		%XCC, 107f			! BR
-	 subcc		%g2, 0x8, %g2			! A1
-	be,pn		%XCC, 107f			! BR	Group
-	 cmp		%g1, 0				! A0
-
-	be,a,pt		%XCC, 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		%XCC, 107f			! 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		%XCC, 1b			! BR
-	 add		%o0, 0x8, %o0			! A0	Group
+loopend:
+	and		%o2, 0x3f, %o2
+	andcc		%o2, 0x38, %g2
+	be,pn		%XCC, endcruft
+	 subcc		%g2, 0x8, %g2
+	be,pn		%XCC, endcruft
+	 cmp		%g1, 0
+
+	be,a,pt		%XCC, 1f
+	 ldd		[%o1 + 0x00], %f0
+
+1:	ldd		[%o1 + 0x08], %f2
+	add		%o1, 0x8, %o1
+	sub		%o2, 0x8, %o2
+	subcc		%g2, 0x8, %g2
+	faligndata	%f0, %f2, %f8
+	std		%f8, [%o0 + 0x00]
+	be,pn		%XCC, endcruft
+	 add		%o0, 0x8, %o0
+	ldd		[%o1 + 0x08], %f0
+	add		%o1, 0x8, %o1
+	sub		%o2, 0x8, %o2
+	subcc		%g2, 0x8, %g2
+	faligndata	%f2, %f0, %f8
+	std		%f8, [%o0 + 0x00]
+	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.
 	 */
-107:
+endcruft:
 	cmp		%o2, 0
 	add		%o1, %g1, %o1
 	VISExitHalf
-	be,pn		%XCC, 102b
-	 nop
-	ba,a,pt		%XCC, 101b
-
-	/* If we get here, then 32 <= len < (6 * 64) */
-108:
+	be,pn		%XCC, out
+	 sub		%o0, %o1, %o3
 
-#ifdef SMALL_COPY_USES_FPU
-
-	/* Is 'dst' already aligned on an 8-byte boundary? */
-	be,pt		%XCC, 2f			! BR	Group
-
-	/* 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)
-
-	/* 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
-
-	bg,pt		%XCC, 1b			! BR	Group
-	 stb		%o3, [%o0 + -1]			! MS	Group
+	andcc		%g1, 0x7, %g0
+	bne,pn		%icc, small_copy_unaligned
+	 andcc		%o2, 0x8, %g0
+	be,pt		%icc, 1f
+	 nop
+	ldx		[%o1], %o5
+	stx		%o5, [%o1 + %o3]
+	add		%o1, 0x8, %o1
 
-2:	VISEntryHalf					! MS+MS
+1:	andcc		%o2, 0x4, %g0
+	be,pt		%icc, 1f
+	 nop
+	lduw		[%o1], %o5
+	stw		%o5, [%o1 + %o3]
+	add		%o1, 0x4, %o1
 
-	/* Compute (len - (len % 8)) into %g2.  This is guarenteed
-	 * 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 guarenteed 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		%XCC, 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		%XCC, 1b			! BR
-	 add		%o0, 0x8, %o0			! A1
-
-	/* Nothing left to copy? */
-2:	cmp		%o2, 0				! A0	Group
-	VISExitHalf					! A0+MS
-	be,pn		%XCC, 102b			! BR	Group
-	 nop						! A0
-	ba,a,pt		%XCC, 101b			! BR	Group
-
-#else /* !(SMALL_COPY_USES_FPU) */
-
-	xor		%o1, %o0, %g2
-	andcc		%g2, 0x7, %g0
-	bne,pn		%XCC, 101b
-	 andcc		%o1, 0x7, %g2
+1:	andcc		%o2, 0x2, %g0
+	be,pt		%icc, 1f
+	 nop
+	lduh		[%o1], %o5
+	sth		%o5, [%o1 + %o3]
+	add		%o1, 0x2, %o1
 
-	be,pt		%XCC, 2f
-	 sub		%g2, 0x8, %g2
-	sub		%g0, %g2, %g2
-	sub		%o2, %g2, %o2
+1:	andcc		%o2, 0x1, %g0
+	be,pt		%icc, out
+	 nop
+	ldub		[%o1], %o5
+	ba,pt		%xcc, out
+	 stb		%o5, [%o1 + %o3]
+
+medium_copy: /* 16 < len <= 64 */
+	bne,pn		%XCC, small_copy_unaligned
+	 sub		%o0, %o1, %o3
+
+medium_copy_aligned:
+	andn		%o2, 0x7, %o4
+	and		%o2, 0x7, %o2
+1:	subcc		%o4, 0x8, %o4
+	ldx		[%o1], %o5
+	stx		%o5, [%o1 + %o3]
+	bgu,pt		%XCC, 1b
+	 add		%o1, 0x8, %o1
+	andcc		%o2, 0x4, %g0
+	be,pt		%XCC, 1f
+	 nop
+	sub		%o2, 0x4, %o2
+	lduw		[%o1], %o5
+	stw		%o5, [%o1 + %o3]
+	add		%o1, 0x4, %o1
+1:	cmp		%o2, 0
+	be,pt		%XCC, out
+	 nop
+	ba,pt		%xcc, small_copy_unaligned
+	 nop
 
-1:	ldub		[%o1 + 0x00], %o3
-	add		%o1, 0x1, %o1
-	add		%o0, 0x1, %o0
-	subcc		%g2, 0x1, %g2
-	bg,pt		%XCC, 1b
-	 stb		%o3, [%o0 + -1]
+small_copy: /* 0 < len <= 16 */
+	andcc		%o3, 0x3, %g0
+	bne,pn		%XCC, small_copy_unaligned
+	 sub		%o0, %o1, %o3
 
-2:	andn		%o2, 0x7, %g2
-	sub		%o2, %g2, %o2
+small_copy_aligned:
+	subcc		%o2, 4, %o2
+	lduw		[%o1], %g1
+	stw		%g1, [%o1 + %o3]
+	bgu,pt		%XCC, small_copy_aligned
+	 add		%o1, 4, %o1
 
-3:	ldx		[%o1 + 0x00], %o3
-	add		%o1, 0x8, %o1
-	add		%o0, 0x8, %o0
-	subcc		%g2, 0x8, %g2
-	bg,pt		%XCC, 3b
-	 stx		%o3, [%o0 + -8]
+out:	retl
+	 mov		%g5, %o0
 
-	cmp		%o2, 0
-	bne,pn		%XCC, 101b
-	 nop
-	ba,a,pt		%XCC, 102b
+	.align	32
+small_copy_unaligned:
+	subcc		%o2, 1, %o2
+	ldub		[%o1], %g1
+	stb		%g1, [%o1 + %o3]
+	bgu,pt		%XCC, small_copy_unaligned
+	 add		%o1, 1, %o1
+	retl
+	 mov		%g5, %o0
 
-#endif /* !(SMALL_COPY_USES_FPU) */
 END(memcpy)
 
 #define RMOVE_BIGCHUNK(src, dst, offset, t0, t1, t2, t3)	\
@@ -511,11 +406,11 @@ END(memcpy)
 
 	.align		32
 ENTRY(memmove)
-	mov		%o0, %g3
+	mov		%o0, %g5
 #ifndef USE_BPR
 	srl		%o2, 0, %o2			/* IEU1		Group		*/
 #endif
-	brz,pn		%o2, 102b			/* CTI		Group		*/
+	brz,pn		%o2, out			/* CTI		Group		*/
 	 sub		%o0, %o1, %o4			/* IEU0				*/
 	cmp		%o4, %o2			/* IEU1		Group		*/
 	bgeu,pt		%XCC, 218b			/* CTI				*/