Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/power/zcopy_hummer.S

kusano 2b45e8
/*********************************************************************/
kusano 2b45e8
/* Copyright 2009, 2010 The University of Texas at Austin.           */
kusano 2b45e8
/* All rights reserved.                                              */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/* Redistribution and use in source and binary forms, with or        */
kusano 2b45e8
/* without modification, are permitted provided that the following   */
kusano 2b45e8
/* conditions are met:                                               */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*   1. Redistributions of source code must retain the above         */
kusano 2b45e8
/*      copyright notice, this list of conditions and the following  */
kusano 2b45e8
/*      disclaimer.                                                  */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*   2. Redistributions in binary form must reproduce the above      */
kusano 2b45e8
/*      copyright notice, this list of conditions and the following  */
kusano 2b45e8
/*      disclaimer in the documentation and/or other materials       */
kusano 2b45e8
/*      provided with the distribution.                              */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
kusano 2b45e8
/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
kusano 2b45e8
/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
kusano 2b45e8
/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
kusano 2b45e8
/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
kusano 2b45e8
/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
kusano 2b45e8
/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
kusano 2b45e8
/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
kusano 2b45e8
/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
kusano 2b45e8
/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
kusano 2b45e8
/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
kusano 2b45e8
/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
kusano 2b45e8
/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
kusano 2b45e8
/*    POSSIBILITY OF SUCH DAMAGE.                                    */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/* The views and conclusions contained in the software and           */
kusano 2b45e8
/* documentation are those of the authors and should not be          */
kusano 2b45e8
/* interpreted as representing official policies, either expressed   */
kusano 2b45e8
/* or implied, of The University of Texas at Austin.                 */
kusano 2b45e8
/*********************************************************************/
kusano 2b45e8
kusano 2b45e8
#define ASSEMBLER
kusano 2b45e8
#include "common.h"
kusano 2b45e8
		
kusano 2b45e8
#define N	r3
kusano 2b45e8
#define X	r4
kusano 2b45e8
#define INCX	r5	
kusano 2b45e8
#define Y	r6
kusano 2b45e8
#define INCY	r7	
kusano 2b45e8
kusano 2b45e8
#define INCX2	r8
kusano 2b45e8
#define INCY2	r9
kusano 2b45e8
#define X2	r10
kusano 2b45e8
#define Y2	r11
kusano 2b45e8
kusano 2b45e8
#define A1	f0
kusano 2b45e8
#define A2	f1
kusano 2b45e8
#define A3	f2
kusano 2b45e8
#define A4	f3
kusano 2b45e8
#define A5	f4
kusano 2b45e8
#define A6	f5
kusano 2b45e8
#define A7	f6
kusano 2b45e8
#define A8	f7
kusano 2b45e8
#define A9	f8
kusano 2b45e8
kusano 2b45e8
#define T1	f9
kusano 2b45e8
#define T2	f10
kusano 2b45e8
#define T3	f11
kusano 2b45e8
#define T4	f12
kusano 2b45e8
#define T5	f13
kusano 2b45e8
#define T6	f14
kusano 2b45e8
#define T7	f15
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
	li	r10, -16
kusano 2b45e8
kusano 2b45e8
	stfpdux	f14, SP, r10
kusano 2b45e8
	stfpdux	f15, SP, r10
kusano 2b45e8
	
kusano 2b45e8
	slwi	INCX,  INCX, BASE_SHIFT
kusano 2b45e8
	slwi	INCY,  INCY, BASE_SHIFT
kusano 2b45e8
	add	INCX2, INCX, INCX
kusano 2b45e8
	add	INCY2, INCY, INCY
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	ble	LL(999)
kusano 2b45e8
kusano 2b45e8
	sub	X, X, INCX2
kusano 2b45e8
	sub	Y, Y, INCY2
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, INCX, SIZE
kusano 2b45e8
	bne	LL(100)
kusano 2b45e8
	cmpwi	cr0, INCY, SIZE
kusano 2b45e8
	bne	LL(100)
kusano 2b45e8
kusano 2b45e8
	andi.	r0, X, 2 * SIZE - 1
kusano 2b45e8
	bne	LL(30)
kusano 2b45e8
	andi.	r0, Y, 2 * SIZE - 1
kusano 2b45e8
	bne	LL(20)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(10):  /* X ): aligned     Y ): aligned */
kusano 2b45e8
	srawi.	r0, N, 3
kusano 2b45e8
	mtspr	CTR,  r0
kusano 2b45e8
	beq-	LL(15)
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A1,   X, INCX2
kusano 2b45e8
	LFPDUX	A2,   X, INCX2
kusano 2b45e8
	LFPDUX	A3,   X, INCX2
kusano 2b45e8
	LFPDUX	A4,   X, INCX2
kusano 2b45e8
	LFPDUX	A5,   X, INCX2
kusano 2b45e8
	LFPDUX	A6,   X, INCX2
kusano 2b45e8
	LFPDUX	A7,   X, INCX2
kusano 2b45e8
	LFPDUX	A8,   X, INCX2
kusano 2b45e8
	bdz	LL(13)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(12):
kusano 2b45e8
	STFPDUX	A1,   Y, INCY2
kusano 2b45e8
	LFPDUX	A1,   X, INCX2
kusano 2b45e8
	STFPDUX	A2,   Y, INCY2
kusano 2b45e8
	LFPDUX	A2,   X, INCX2
kusano 2b45e8
	STFPDUX	A3,   Y, INCY2
kusano 2b45e8
	LFPDUX	A3,   X, INCX2
kusano 2b45e8
	STFPDUX	A4,   Y, INCY2
kusano 2b45e8
	LFPDUX	A4,   X, INCX2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	A5,   Y, INCY2
kusano 2b45e8
	LFPDUX	A5,   X, INCX2
kusano 2b45e8
	STFPDUX	A6,   Y, INCY2
kusano 2b45e8
	LFPDUX	A6,   X, INCX2
kusano 2b45e8
	STFPDUX	A7,   Y, INCY2
kusano 2b45e8
	LFPDUX	A7,   X, INCX2
kusano 2b45e8
	STFPDUX	A8,   Y, INCY2
kusano 2b45e8
	LFPDUX	A8,   X, INCX2
kusano 2b45e8
	bdnz	LL(12)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(13):
kusano 2b45e8
	STFPDUX	A1,   Y, INCY2
kusano 2b45e8
	STFPDUX	A2,   Y, INCY2
kusano 2b45e8
	STFPDUX	A3,   Y, INCY2
kusano 2b45e8
	STFPDUX	A4,   Y, INCY2
kusano 2b45e8
	STFPDUX	A5,   Y, INCY2
kusano 2b45e8
	STFPDUX	A6,   Y, INCY2
kusano 2b45e8
	STFPDUX	A7,   Y, INCY2
kusano 2b45e8
	STFPDUX	A8,   Y, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(15):
kusano 2b45e8
	andi.	r0,  N, 7
kusano 2b45e8
	beq	LL(999)
kusano 2b45e8
kusano 2b45e8
	andi.	r0,  N, 4
kusano 2b45e8
	beq	LL(16)
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A1,    X, INCX2
kusano 2b45e8
	LFPDUX	A2,    X, INCX2
kusano 2b45e8
	LFPDUX	A3,    X, INCX2
kusano 2b45e8
	LFPDUX	A4,    X, INCX2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	STFPDUX	A2,    Y, INCY2
kusano 2b45e8
	STFPDUX	A3,    Y, INCY2
kusano 2b45e8
	STFPDUX	A4,    Y, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(16):
kusano 2b45e8
	andi.	r0,  N, 2
kusano 2b45e8
	beq	LL(17)
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A1,    X, INCX2
kusano 2b45e8
	LFPDUX	A2,    X, INCX2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	STFPDUX	A2,    Y, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(17):
kusano 2b45e8
	andi.	r0,  N, 1
kusano 2b45e8
	beq	LL(999)
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A1,    X, INCX2
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	b  LL(999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(20):  /* X : aligned     Y : unaligned */
kusano 2b45e8
kusano 2b45e8
	LFXDUX	A1, X, INCX2
kusano 2b45e8
	addi	N, N, -1
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	STFSDX	A1, Y, INCY2
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	ble	LL(29)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, N, 3
kusano 2b45e8
	mtspr	CTR,  r0
kusano 2b45e8
	beq-	LL(25)
kusano 2b45e8
kusano 2b45e8
	LFXDUX	T1,   X, INCX2
kusano 2b45e8
	LFXDUX	T2,   X, INCX2
kusano 2b45e8
	LFXDUX	T3,   X, INCX2
kusano 2b45e8
	LFXDUX	T4,   X, INCX2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A6,   X, INCX2
kusano 2b45e8
	fsmr	A1, T1
kusano 2b45e8
	LFPDUX	A7,   X, INCX2
kusano 2b45e8
	fsmr	T1, T2
kusano 2b45e8
	LFPDUX	A8,   X, INCX2
kusano 2b45e8
	fsmr	T2, T3
kusano 2b45e8
	LFPDUX	A9,   X, INCX2
kusano 2b45e8
	fsmr	T3, T4
kusano 2b45e8
	bdz	LL(23)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(22):
kusano 2b45e8
	STFPDUX	A1,   Y, INCY2
kusano 2b45e8
	fxmr	T5, A6
kusano 2b45e8
	STFPDUX	T1,   Y, INCY2
kusano 2b45e8
	fxmr	T6, A7
kusano 2b45e8
	STFPDUX	T2,   Y, INCY2
kusano 2b45e8
	fxmr	T7, A8
kusano 2b45e8
	STFPDUX	T3,   Y, INCY2
kusano 2b45e8
	fxmr	A1, A9
kusano 2b45e8
kusano 2b45e8
	fsmr	T4, T5
kusano 2b45e8
	LFPDUX	A2,   X, INCX2
kusano 2b45e8
	fsmr	T5, T6
kusano 2b45e8
	LFPDUX	A3,   X, INCX2
kusano 2b45e8
	fsmr	T6, T7
kusano 2b45e8
	LFPDUX	A4,   X, INCX2
kusano 2b45e8
	fsmr	T7, A1
kusano 2b45e8
	LFPDUX	A5,   X, INCX2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	T4,   Y, INCY2
kusano 2b45e8
	fxmr	T1, A2
kusano 2b45e8
	STFPDUX	T5,   Y, INCY2
kusano 2b45e8
	fxmr	T2, A3
kusano 2b45e8
	STFPDUX	T6,   Y, INCY2
kusano 2b45e8
	fxmr	T3, A4
kusano 2b45e8
	STFPDUX	T7,   Y, INCY2
kusano 2b45e8
	fxmr	T4, A5
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A6,   X, INCX2
kusano 2b45e8
	fsmr	A1, T1
kusano 2b45e8
	LFPDUX	A7,   X, INCX2
kusano 2b45e8
	fsmr	T1, T2
kusano 2b45e8
	LFPDUX	A8,   X, INCX2
kusano 2b45e8
	fsmr	T2, T3
kusano 2b45e8
	LFPDUX	A9,   X, INCX2
kusano 2b45e8
	fsmr	T3, T4
kusano 2b45e8
	bdnz	LL(22)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(23):
kusano 2b45e8
	STFPDUX	A1,   Y, INCY2
kusano 2b45e8
	fxmr	T5, A6
kusano 2b45e8
	STFPDUX	T1,   Y, INCY2
kusano 2b45e8
	fxmr	T6, A7
kusano 2b45e8
	STFPDUX	T2,   Y, INCY2
kusano 2b45e8
	fxmr	T7, A8
kusano 2b45e8
	STFPDUX	T3,   Y, INCY2
kusano 2b45e8
	fxmr	A1, A9
kusano 2b45e8
kusano 2b45e8
	fsmr	T4, T5
kusano 2b45e8
	fsmr	T5, T6
kusano 2b45e8
	fsmr	T6, T7
kusano 2b45e8
	fsmr	T7, A1
kusano 2b45e8
kusano 2b45e8
	STFPDUX	T4,   Y, INCY2
kusano 2b45e8
	STFPDUX	T5,   Y, INCY2
kusano 2b45e8
	STFPDUX	T6,   Y, INCY2
kusano 2b45e8
	STFPDUX	T7,   Y, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(25):
kusano 2b45e8
	andi.	r0,  N, 7
kusano 2b45e8
	beq	LL(29)
kusano 2b45e8
	
kusano 2b45e8
	andi.	r0,  N, 4
kusano 2b45e8
	beq	LL(26)
kusano 2b45e8
kusano 2b45e8
	LFXDUX	A2,    X, INCX2
kusano 2b45e8
	LFXDUX	A3,    X, INCX2
kusano 2b45e8
	LFXDUX	A4,    X, INCX2
kusano 2b45e8
	LFXDUX	A5,    X, INCX2
kusano 2b45e8
kusano 2b45e8
	fsmr	A1, A2
kusano 2b45e8
	fsmr	A2, A3
kusano 2b45e8
	fsmr	A3, A4
kusano 2b45e8
	fsmr	A4, A5
kusano 2b45e8
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	STFPDUX	A2,    Y, INCY2
kusano 2b45e8
	STFPDUX	A3,    Y, INCY2
kusano 2b45e8
	STFPDUX	A4,    Y, INCY2
kusano 2b45e8
	fpmr	A1, A5
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(26):
kusano 2b45e8
	andi.	r0,  N, 2
kusano 2b45e8
	beq	LL(27)
kusano 2b45e8
kusano 2b45e8
	LFXDUX	A2,    X, INCX2
kusano 2b45e8
	LFXDUX	A3,    X, INCX2
kusano 2b45e8
	fsmr	A1, A2
kusano 2b45e8
	fsmr	A2, A3
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	STFPDUX	A2,    Y, INCY2
kusano 2b45e8
	fpmr	A1, A3
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(27):
kusano 2b45e8
	andi.	r0,  N, 1
kusano 2b45e8
	beq	LL(29)
kusano 2b45e8
kusano 2b45e8
	LFXDUX	A2,    X, INCX2
kusano 2b45e8
	fsmr	A1, A2
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	fpmr	A1, A2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(29):
kusano 2b45e8
	STFDUX	A1,    Y, INCY2
kusano 2b45e8
	b  LL(999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(30):  /* X ): unaligned   Y ): aligned */
kusano 2b45e8
	andi.	r0, Y, 2 * SIZE - 1
kusano 2b45e8
	bne	LL(40)
kusano 2b45e8
kusano 2b45e8
	LFDX	A1, X, INCX2
kusano 2b45e8
	add	X, X, INCX
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, N, 3
kusano 2b45e8
	mtspr	CTR,  r0
kusano 2b45e8
	beq-	LL(35)
kusano 2b45e8
kusano 2b45e8
	LFXDUX	T1,   X, INCX2
kusano 2b45e8
	LFXDUX	T2,   X, INCX2
kusano 2b45e8
	LFXDUX	T3,   X, INCX2
kusano 2b45e8
	LFXDUX	T4,   X, INCX2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A6,   X, INCX2
kusano 2b45e8
	fsmr	A1, T1
kusano 2b45e8
	LFPDUX	A7,   X, INCX2
kusano 2b45e8
	fsmr	T1, T2
kusano 2b45e8
	LFPDUX	A8,   X, INCX2
kusano 2b45e8
	fsmr	T2, T3
kusano 2b45e8
	LFPDUX	A9,   X, INCX2
kusano 2b45e8
	fsmr	T3, T4
kusano 2b45e8
	bdz	LL(33)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(32):
kusano 2b45e8
	fxmr	T5, A6
kusano 2b45e8
	STFPDUX	A1,   Y, INCY2
kusano 2b45e8
	fxmr	T6, A7
kusano 2b45e8
	STFPDUX	T1,   Y, INCY2
kusano 2b45e8
	fxmr	T7, A8
kusano 2b45e8
	STFPDUX	T2,   Y, INCY2
kusano 2b45e8
	fxmr	A1, A9
kusano 2b45e8
	STFPDUX	T3,   Y, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A2,   X, INCX2
kusano 2b45e8
	fsmr	T4, T5
kusano 2b45e8
	LFPDUX	A3,   X, INCX2
kusano 2b45e8
	fsmr	T5, T6
kusano 2b45e8
	LFPDUX	A4,   X, INCX2
kusano 2b45e8
	fsmr	T6, T7
kusano 2b45e8
	LFPDUX	A5,   X, INCX2
kusano 2b45e8
	fsmr	T7, A1
kusano 2b45e8
kusano 2b45e8
	fxmr	T1, A2
kusano 2b45e8
	STFPDUX	T4,   Y, INCY2
kusano 2b45e8
	fxmr	T2, A3
kusano 2b45e8
	STFPDUX	T5,   Y, INCY2
kusano 2b45e8
	fxmr	T3, A4
kusano 2b45e8
	STFPDUX	T6,   Y, INCY2
kusano 2b45e8
	fxmr	T4, A5
kusano 2b45e8
	STFPDUX	T7,   Y, INCY2
kusano 2b45e8
kusano 2b45e8
	fsmr	A1, T1
kusano 2b45e8
	LFPDUX	A6,   X, INCX2
kusano 2b45e8
	fsmr	T1, T2
kusano 2b45e8
	LFPDUX	A7,   X, INCX2
kusano 2b45e8
	fsmr	T2, T3
kusano 2b45e8
	LFPDUX	A8,   X, INCX2
kusano 2b45e8
	fsmr	T3, T4
kusano 2b45e8
	LFPDUX	A9,   X, INCX2
kusano 2b45e8
	bdnz	LL(32)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(33):
kusano 2b45e8
	STFPDUX	A1,   Y, INCY2
kusano 2b45e8
	fxmr	T5, A6
kusano 2b45e8
	STFPDUX	T1,   Y, INCY2
kusano 2b45e8
	fxmr	T6, A7
kusano 2b45e8
	STFPDUX	T2,   Y, INCY2
kusano 2b45e8
	fxmr	T7, A8
kusano 2b45e8
	STFPDUX	T3,   Y, INCY2
kusano 2b45e8
	fxmr	A1, A9
kusano 2b45e8
kusano 2b45e8
	fsmr	T4, T5
kusano 2b45e8
	fsmr	T5, T6
kusano 2b45e8
	fsmr	T6, T7
kusano 2b45e8
	fsmr	T7, A1
kusano 2b45e8
kusano 2b45e8
	STFPDUX	T4,   Y, INCY2
kusano 2b45e8
	STFPDUX	T5,   Y, INCY2
kusano 2b45e8
	STFPDUX	T6,   Y, INCY2
kusano 2b45e8
	STFPDUX	T7,   Y, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(35):
kusano 2b45e8
	andi.	r0,  N, 7
kusano 2b45e8
	beq	LL(999)
kusano 2b45e8
	
kusano 2b45e8
	andi.	r0,  N, 4
kusano 2b45e8
	beq	LL(36)
kusano 2b45e8
kusano 2b45e8
	LFXDUX	A2,    X, INCX2
kusano 2b45e8
	LFXDUX	A3,    X, INCX2
kusano 2b45e8
	LFXDUX	A4,    X, INCX2
kusano 2b45e8
	LFXDUX	A5,    X, INCX2
kusano 2b45e8
kusano 2b45e8
	fsmr	A1, A2
kusano 2b45e8
	fsmr	A2, A3
kusano 2b45e8
	fsmr	A3, A4
kusano 2b45e8
	fsmr	A4, A5
kusano 2b45e8
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	STFPDUX	A2,    Y, INCY2
kusano 2b45e8
	STFPDUX	A3,    Y, INCY2
kusano 2b45e8
	STFPDUX	A4,    Y, INCY2
kusano 2b45e8
	fpmr	A1, A5
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(36):
kusano 2b45e8
	andi.	r0,  N, 2
kusano 2b45e8
	beq	LL(37)
kusano 2b45e8
kusano 2b45e8
	LFXDUX	A2,    X, INCX2
kusano 2b45e8
	LFXDUX	A3,    X, INCX2
kusano 2b45e8
	fsmr	A1, A2
kusano 2b45e8
	fsmr	A2, A3
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	STFPDUX	A2,    Y, INCY2
kusano 2b45e8
	fpmr	A1, A3
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(37):
kusano 2b45e8
	andi.	r0,  N, 1
kusano 2b45e8
	beq	LL(999)
kusano 2b45e8
kusano 2b45e8
	LFXDUX	A2,    X, INCX2
kusano 2b45e8
	fsmr	A1, A2
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	b  LL(999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(40):  /* X : unaligned   Y : unaligned */
kusano 2b45e8
kusano 2b45e8
	LFDX	A1, X, INCX2
kusano 2b45e8
	add	X, X, INCX
kusano 2b45e8
kusano 2b45e8
	addi	N, N, -1
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	STFDX	A1, Y, INCY2
kusano 2b45e8
	add	Y, Y, INCY
kusano 2b45e8
	ble	LL(49)
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, N, 3
kusano 2b45e8
	mtspr	CTR,  r0
kusano 2b45e8
	beq-	LL(45)
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A1,   X, INCX2
kusano 2b45e8
	LFPDUX	A2,   X, INCX2
kusano 2b45e8
	LFPDUX	A3,   X, INCX2
kusano 2b45e8
	LFPDUX	A4,   X, INCX2
kusano 2b45e8
	LFPDUX	A5,   X, INCX2
kusano 2b45e8
	LFPDUX	A6,   X, INCX2
kusano 2b45e8
	LFPDUX	A7,   X, INCX2
kusano 2b45e8
	LFPDUX	A8,   X, INCX2
kusano 2b45e8
	bdz	LL(43)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(42):
kusano 2b45e8
	STFPDUX	A1,   Y, INCY2
kusano 2b45e8
	LFPDUX	A1,   X, INCX2
kusano 2b45e8
	STFPDUX	A2,   Y, INCY2
kusano 2b45e8
	LFPDUX	A2,   X, INCX2
kusano 2b45e8
	STFPDUX	A3,   Y, INCY2
kusano 2b45e8
	LFPDUX	A3,   X, INCX2
kusano 2b45e8
	STFPDUX	A4,   Y, INCY2
kusano 2b45e8
	LFPDUX	A4,   X, INCX2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	A5,   Y, INCY2
kusano 2b45e8
	LFPDUX	A5,   X, INCX2
kusano 2b45e8
	STFPDUX	A6,   Y, INCY2
kusano 2b45e8
	LFPDUX	A6,   X, INCX2
kusano 2b45e8
	STFPDUX	A7,   Y, INCY2
kusano 2b45e8
	LFPDUX	A7,   X, INCX2
kusano 2b45e8
	STFPDUX	A8,   Y, INCY2
kusano 2b45e8
	LFPDUX	A8,   X, INCX2
kusano 2b45e8
	bdnz	LL(42)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(43):
kusano 2b45e8
	STFPDUX	A1,   Y, INCY2
kusano 2b45e8
	STFPDUX	A2,   Y, INCY2
kusano 2b45e8
	STFPDUX	A3,   Y, INCY2
kusano 2b45e8
	STFPDUX	A4,   Y, INCY2
kusano 2b45e8
	STFPDUX	A5,   Y, INCY2
kusano 2b45e8
	STFPDUX	A6,   Y, INCY2
kusano 2b45e8
	STFPDUX	A7,   Y, INCY2
kusano 2b45e8
	STFPDUX	A8,   Y, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(45):
kusano 2b45e8
	andi.	r0,  N, 7
kusano 2b45e8
	beq	LL(49)
kusano 2b45e8
kusano 2b45e8
	andi.	r0,  N, 4
kusano 2b45e8
	beq	LL(46)
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A1,    X, INCX2
kusano 2b45e8
	LFPDUX	A2,    X, INCX2
kusano 2b45e8
	LFPDUX	A3,    X, INCX2
kusano 2b45e8
	LFPDUX	A4,    X, INCX2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	STFPDUX	A2,    Y, INCY2
kusano 2b45e8
	STFPDUX	A3,    Y, INCY2
kusano 2b45e8
	STFPDUX	A4,    Y, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(46):
kusano 2b45e8
	andi.	r0,  N, 2
kusano 2b45e8
	beq	LL(47)
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A1,    X, INCX2
kusano 2b45e8
	LFPDUX	A2,    X, INCX2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
	STFPDUX	A2,    Y, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(47):
kusano 2b45e8
	andi.	r0,  N, 1
kusano 2b45e8
	beq	LL(49)
kusano 2b45e8
kusano 2b45e8
	LFPDUX	A1,    X, INCX2
kusano 2b45e8
	STFPDUX	A1,    Y, INCY2
kusano 2b45e8
kusano 2b45e8
LL(49):
kusano 2b45e8
	LFDUX	A1,    X, INCX2
kusano 2b45e8
	STFDUX	A1,    Y, INCY2
kusano 2b45e8
	b  LL(999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(100):
kusano 2b45e8
	addi	X2, X, SIZE
kusano 2b45e8
	addi	Y2, Y, SIZE
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, N, 2
kusano 2b45e8
	mtspr	CTR,  r0
kusano 2b45e8
	beq-	LL(115)
kusano 2b45e8
kusano 2b45e8
	LFDUX	A1,   X,  INCX2
kusano 2b45e8
	LFDUX	A2,   X2, INCX2
kusano 2b45e8
	LFDUX	A3,   X,  INCX2
kusano 2b45e8
	LFDUX	A4,   X2, INCX2
kusano 2b45e8
	LFDUX	A5,   X,  INCX2
kusano 2b45e8
	LFDUX	A6,   X2, INCX2
kusano 2b45e8
	LFDUX	A7,   X,  INCX2
kusano 2b45e8
	LFDUX	A8,   X2, INCX2
kusano 2b45e8
	bdz	LL(113)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(112):
kusano 2b45e8
	STFDUX	A1,   Y,  INCY2
kusano 2b45e8
	LFDUX	A1,   X,  INCX2
kusano 2b45e8
	STFDUX	A2,   Y2, INCY2
kusano 2b45e8
	LFDUX	A2,   X2, INCX2
kusano 2b45e8
	STFDUX	A3,   Y,  INCY2
kusano 2b45e8
	LFDUX	A3,   X,  INCX2
kusano 2b45e8
	STFDUX	A4,   Y2, INCY2
kusano 2b45e8
	LFDUX	A4,   X2, INCX2
kusano 2b45e8
kusano 2b45e8
	STFDUX	A5,   Y,  INCY2
kusano 2b45e8
	LFDUX	A5,   X,  INCX2
kusano 2b45e8
	STFDUX	A6,   Y2, INCY2
kusano 2b45e8
	LFDUX	A6,   X2, INCX2
kusano 2b45e8
	STFDUX	A7,   Y,  INCY2
kusano 2b45e8
	LFDUX	A7,   X,  INCX2
kusano 2b45e8
	STFDUX	A8,   Y2, INCY2
kusano 2b45e8
	LFDUX	A8,   X2, INCX2
kusano 2b45e8
	bdnz	LL(112)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(113):
kusano 2b45e8
	STFDUX	A1,   Y,  INCY2
kusano 2b45e8
	STFDUX	A2,   Y2, INCY2
kusano 2b45e8
	STFDUX	A3,   Y,  INCY2
kusano 2b45e8
	STFDUX	A4,   Y2, INCY2
kusano 2b45e8
	STFDUX	A5,   Y,  INCY2
kusano 2b45e8
	STFDUX	A6,   Y2, INCY2
kusano 2b45e8
	STFDUX	A7,   Y,  INCY2
kusano 2b45e8
	STFDUX	A8,   Y2, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(115):
kusano 2b45e8
	andi.	r0,  N, 3
kusano 2b45e8
	beq	LL(999)
kusano 2b45e8
	andi.	r0,  N, 2
kusano 2b45e8
	beq	LL(117)
kusano 2b45e8
kusano 2b45e8
	LFDUX	A1,    X,  INCX2
kusano 2b45e8
	LFDUX	A2,    X2, INCX2
kusano 2b45e8
	LFDUX	A3,    X,  INCX2
kusano 2b45e8
	LFDUX	A4,    X2, INCX2
kusano 2b45e8
kusano 2b45e8
	STFDUX	A1,    Y,  INCY2
kusano 2b45e8
	STFDUX	A2,    Y2, INCY2
kusano 2b45e8
	STFDUX	A3,    Y,  INCY2
kusano 2b45e8
	STFDUX	A4,    Y2, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(117):
kusano 2b45e8
	andi.	r0,  N, 1
kusano 2b45e8
	beq	LL(999)
kusano 2b45e8
kusano 2b45e8
	LFDUX	A1,    X,  INCX2
kusano 2b45e8
	LFDUX	A2,    X2, INCX2
kusano 2b45e8
kusano 2b45e8
	STFDUX	A1,    Y,  INCY2
kusano 2b45e8
	STFDUX	A2,    Y2, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(999):
kusano 2b45e8
	li	r10, 16
kusano 2b45e8
	addi	SP, SP,  -16
kusano 2b45e8
kusano 2b45e8
	lfpdux	f15, SP, r10
kusano 2b45e8
	lfpdux	f14, SP, r10
kusano 2b45e8
	
kusano 2b45e8
	addi	SP, SP,  16
kusano 2b45e8
	blr
kusano 2b45e8
kusano 2b45e8
	EPILOGUE