Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/power/gemv_hummer_n.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 M	r3
kusano 2b45e8
#define	N	r4
kusano 2b45e8
#define A	r6
kusano 2b45e8
#define LDA	r7
kusano 2b45e8
#define X	r8
kusano 2b45e8
#define	INCX	r9
kusano 2b45e8
#define	Y	r10
kusano 2b45e8
#define	INCY	r5
kusano 2b45e8
kusano 2b45e8
#define I	r11
kusano 2b45e8
#define	J	r12
kusano 2b45e8
kusano 2b45e8
#define INCY2	r24
kusano 2b45e8
#define A1	r25
kusano 2b45e8
#define A2	r26
kusano 2b45e8
#define A3	r27
kusano 2b45e8
#define A4	r28
kusano 2b45e8
kusano 2b45e8
#define YL	r29
kusano 2b45e8
#define	YS	r30
kusano 2b45e8
#define INC2	r31
kusano 2b45e8
kusano 2b45e8
#define yl1 f0
kusano 2b45e8
#define yl2 f2
kusano 2b45e8
#define yl3 f3
kusano 2b45e8
#define yl4 f4
kusano 2b45e8
#define ys1 f5
kusano 2b45e8
#define ys2 f6
kusano 2b45e8
#define ys3 f7
kusano 2b45e8
#define ys4 f8
kusano 2b45e8
#define	yl5 f27
kusano 2b45e8
#define ys5 f28
kusano 2b45e8
kusano 2b45e8
#define alpha1 f9
kusano 2b45e8
#define alpha2 f10
kusano 2b45e8
kusano 2b45e8
#define a1     f11
kusano 2b45e8
#define a2     f12
kusano 2b45e8
#define a3     f13
kusano 2b45e8
#define a4     f14
kusano 2b45e8
#define a5     f15
kusano 2b45e8
#define a6     f16
kusano 2b45e8
#define a7     f17
kusano 2b45e8
#define a8     f18
kusano 2b45e8
kusano 2b45e8
#define a9     f19
kusano 2b45e8
#define a10    f20
kusano 2b45e8
#define a11    f21
kusano 2b45e8
#define a12    f22
kusano 2b45e8
#define a13    f23
kusano 2b45e8
#define a14    f24
kusano 2b45e8
#define a15    f25
kusano 2b45e8
#define a16    f26
kusano 2b45e8
kusano 2b45e8
#define alpha  f1
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
	li	r0, -16
kusano 2b45e8
	lwz	INCY,      8(SP)
kusano 2b45e8
kusano 2b45e8
	stfpdux	f14, SP, r0
kusano 2b45e8
	stfpdux	f15, SP, r0
kusano 2b45e8
	stfpdux	f16, SP, r0
kusano 2b45e8
	stfpdux	f17, SP, r0
kusano 2b45e8
	stfpdux	f18, SP, r0
kusano 2b45e8
	stfpdux	f19, SP, r0
kusano 2b45e8
	stfpdux	f20, SP, r0
kusano 2b45e8
	stfpdux	f21, SP, r0
kusano 2b45e8
	stfpdux	f22, SP, r0
kusano 2b45e8
	stfpdux	f23, SP, r0
kusano 2b45e8
	stfpdux	f24, SP, r0
kusano 2b45e8
	stfpdux	f25, SP, r0
kusano 2b45e8
	stfpdux	f26, SP, r0
kusano 2b45e8
	stfpdux	f27, SP, r0
kusano 2b45e8
	stfpdux	f28, SP, r0
kusano 2b45e8
	stfpdux	f29, SP, r0
kusano 2b45e8
	stfpdux	f30, SP, r0
kusano 2b45e8
	stfpdux	f31, SP, r0
kusano 2b45e8
kusano 2b45e8
	stwu	r31,  -4(SP)
kusano 2b45e8
	stwu	r30,  -4(SP)
kusano 2b45e8
	stwu	r29,  -4(SP)
kusano 2b45e8
	stwu	r28,  -4(SP)
kusano 2b45e8
kusano 2b45e8
	stwu	r27,  -4(SP)
kusano 2b45e8
	stwu	r26,  -4(SP)
kusano 2b45e8
	stwu	r25,  -4(SP)
kusano 2b45e8
	stwu	r24,  -4(SP)
kusano 2b45e8
kusano 2b45e8
	stwu	r23,  -4(SP)
kusano 2b45e8
	stwu	r22,  -4(SP)
kusano 2b45e8
	stwu	r21,  -4(SP)
kusano 2b45e8
	stwu	r20,  -4(SP)
kusano 2b45e8
kusano 2b45e8
	stwu	r19,  -4(SP)
kusano 2b45e8
	stwu	r18,  -4(SP)
kusano 2b45e8
	stwu	r17,  -4(SP)
kusano 2b45e8
	stwu	r16,  -4(SP)
kusano 2b45e8
kusano 2b45e8
	slwi	LDA,  LDA,  BASE_SHIFT
kusano 2b45e8
	slwi	INCX, INCX, BASE_SHIFT
kusano 2b45e8
	slwi	INCY, INCY, BASE_SHIFT
kusano 2b45e8
kusano 2b45e8
	fsmfp	alpha, alpha
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, M, 0
kusano 2b45e8
	ble-	.L999
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	ble-	.L999
kusano 2b45e8
kusano 2b45e8
	add	INCY2, INCY, INCY
kusano 2b45e8
	li	INC2, 2 * SIZE
kusano 2b45e8
	sub	X, X, INCX
kusano 2b45e8
kusano 2b45e8
	andi.	r0, A,  2 * SIZE - 1
kusano 2b45e8
#	bne	.L100
kusano 2b45e8
kusano 2b45e8
# All cases for aligned A, even LDA
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, INCY,  SIZE
kusano 2b45e8
	bne	.L70
kusano 2b45e8
kusano 2b45e8
	andi.	r0, Y,  2 * SIZE - 1
kusano 2b45e8
	bne	.L40
kusano 2b45e8
kusano 2b45e8
# A : aligned  LDA : even  Y : Unit Aligned
kusano 2b45e8
kusano 2b45e8
	sub	A, A, INC2
kusano 2b45e8
	sub	Y, Y, INCY2
kusano 2b45e8
kusano 2b45e8
	srawi.	J, N, 2
kusano 2b45e8
	ble	.L20
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L11:
kusano 2b45e8
	LFDUX	alpha1, X, INCX
kusano 2b45e8
	mr	A1, A
kusano 2b45e8
	add	A2, A,  LDA
kusano 2b45e8
	add	A3, A2, LDA
kusano 2b45e8
	LFSDUX	alpha1, X, INCX
kusano 2b45e8
	LFDUX	alpha2, X, INCX
kusano 2b45e8
	add	A4, A3, LDA
kusano 2b45e8
	add	A,  A4, LDA
kusano 2b45e8
	mr	YL, Y
kusano 2b45e8
	LFSDUX	alpha2, X, INCX
kusano 2b45e8
	fpmul	alpha1, alpha, alpha1
kusano 2b45e8
	mr	YS, Y
kusano 2b45e8
	srawi.	r0,  M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	fpmul	alpha2, alpha, alpha2
kusano 2b45e8
	ble	.L15
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
	LFPDUX	yl2, YL, INCY2
kusano 2b45e8
	LFPDUX	yl3, YL, INCY2
kusano 2b45e8
	LFPDUX	yl4, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	LFPDUX	a10, A2, INC2
kusano 2b45e8
	LFPDUX	a14, A2, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	LFPDUX	a7,  A3, INC2
kusano 2b45e8
	LFPDUX	a11, A3, INC2
kusano 2b45e8
	LFPDUX	a15, A3, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  yl1
kusano 2b45e8
	LFPDUX	a8,  A4, INC2
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  yl2
kusano 2b45e8
	LFPDUX	a12, A4, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a9,  yl3
kusano 2b45e8
	LFPDUX	a16, A4, INC2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a13, yl4
kusano 2b45e8
	bdz	.L13
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L12:
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2,  ys1
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6,  ys2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a10, ys3
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a14, ys4
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl2, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha2, a3,  ys1
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a7,  ys2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha2, a11, ys3
kusano 2b45e8
	LFPDUX	a10, A2, INC2
kusano 2b45e8
	fxcpmadd  ys4, alpha2, a15, ys4
kusano 2b45e8
	LFPDUX	a14, A2, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl3, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha2, a4,  ys1
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a8,  ys2
kusano 2b45e8
	LFPDUX	a7,  A3, INC2
kusano 2b45e8
	fxcsmadd  ys3, alpha2, a12, ys3
kusano 2b45e8
	LFPDUX	a11, A3, INC2
kusano 2b45e8
	fxcsmadd  ys4, alpha2, a16, ys4
kusano 2b45e8
	LFPDUX	a15, A3, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl4, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	STFPDUX	ys2, YS, INCY2
kusano 2b45e8
	STFPDUX	ys3, YS, INCY2
kusano 2b45e8
	STFPDUX	ys4, YS, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  yl1
kusano 2b45e8
	LFPDUX	a8,  A4, INC2
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  yl2
kusano 2b45e8
	LFPDUX	a12, A4, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a9,  yl3
kusano 2b45e8
	LFPDUX	a16, A4, INC2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a13, yl4
kusano 2b45e8
	bdnz	.L12
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L13:
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2,  ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6,  ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a10, ys3
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a14, ys4
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha2, a3,  ys1
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a7,  ys2
kusano 2b45e8
	fxcpmadd  ys3, alpha2, a11, ys3
kusano 2b45e8
	fxcpmadd  ys4, alpha2, a15, ys4
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha2, a4,  ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a8,  ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha2, a12, ys3
kusano 2b45e8
	fxcsmadd  ys4, alpha2, a16, ys4
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	STFPDUX	ys2, YS, INCY2
kusano 2b45e8
	STFPDUX	ys3, YS, INCY2
kusano 2b45e8
	STFPDUX	ys4, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L15:
kusano 2b45e8
	andi.	r0, M, 7
kusano 2b45e8
	ble	.L19
kusano 2b45e8
kusano 2b45e8
	andi.	r0, M, 4
kusano 2b45e8
	ble	.L17
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	yl2, YL, INCY2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	LFPDUX	a7,  A3, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
	LFPDUX	a8,  A4, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5, yl2
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6, ys2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha2, a3, ys1
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a7, ys2
kusano 2b45e8
	fxcsmadd  ys1, alpha2, a4, ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a8, ys2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	STFPDUX	ys2, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L17:
kusano 2b45e8
	andi.	r0, M, 2
kusano 2b45e8
	ble	.L18
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
	fxcpmadd  ys1, alpha2, a3, ys1
kusano 2b45e8
	fxcsmadd  ys1, alpha2, a4, ys1
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L18:
kusano 2b45e8
	andi.	r0, M, 1
kusano 2b45e8
	ble	.L19
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	LFDUX	a1,  A1, INC2
kusano 2b45e8
	LFDUX	a2,  A2, INC2
kusano 2b45e8
	LFDUX	a3,  A3, INC2
kusano 2b45e8
	LFDUX	a4,  A4, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
	fxcpmadd  ys1, alpha2, a3, ys1
kusano 2b45e8
	fxcsmadd  ys1, alpha2, a4, ys1
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L19:
kusano 2b45e8
	addi	J, J, -1
kusano 2b45e8
	cmpi	cr0, 0, J, 0
kusano 2b45e8
	bgt	.L11
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
.L20:
kusano 2b45e8
	andi.	J, N, 2
kusano 2b45e8
	ble	.L30
kusano 2b45e8
kusano 2b45e8
	LFDUX	alpha1, X, INCX
kusano 2b45e8
kusano 2b45e8
	mr	A1, A
kusano 2b45e8
	add	A2, A,  LDA
kusano 2b45e8
	add	A,  A2, LDA
kusano 2b45e8
	LFSDUX	alpha1, X, INCX
kusano 2b45e8
kusano 2b45e8
	mr	YL, Y
kusano 2b45e8
	mr	YS, Y
kusano 2b45e8
	fpmul	alpha1, alpha, alpha1
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L25
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	yl2, YL, INCY2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl3, YL, INCY2
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	LFPDUX	yl4, YL, INCY2
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	LFPDUX	a10, A2, INC2
kusano 2b45e8
	LFPDUX	a14, A2, INC2
kusano 2b45e8
	bdz	.L23
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L22:
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  yl1
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  yl2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	yl2, YL, INCY2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a9,  yl3
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	LFPDUX	yl3, YL, INCY2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a13, yl4
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
	LFPDUX	yl4, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2,  ys1
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6,  ys2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a10, ys3
kusano 2b45e8
	LFPDUX	a10, A2, INC2
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a14, ys4
kusano 2b45e8
	LFPDUX	a14, A2, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	STFPDUX	ys2, YS, INCY2
kusano 2b45e8
	STFPDUX	ys3, YS, INCY2
kusano 2b45e8
	STFPDUX	ys4, YS, INCY2
kusano 2b45e8
	bdnz	.L22
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L23:
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  yl1
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  yl2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a9,  yl3
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a13, yl4
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2,  ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6,  ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a10, ys3
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a14, ys4
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	STFPDUX	ys2, YS, INCY2
kusano 2b45e8
	STFPDUX	ys3, YS, INCY2
kusano 2b45e8
	STFPDUX	ys4, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L25:
kusano 2b45e8
	andi.	r0, M, 7
kusano 2b45e8
	ble	.L30
kusano 2b45e8
kusano 2b45e8
	andi.	r0, M, 4
kusano 2b45e8
	ble	.L27
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl2, YL, INCY2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5, yl2
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6, ys2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	STFPDUX	ys2, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L27:
kusano 2b45e8
	andi.	r0, M, 2
kusano 2b45e8
	ble	.L28
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L28:
kusano 2b45e8
	andi.	r0, M, 1
kusano 2b45e8
	ble	.L30
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY2
kusano 2b45e8
	LFDUX	a1,  A1, INC2
kusano 2b45e8
	LFDUX	a2,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L30:
kusano 2b45e8
	andi.	J, N, 1
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	LFDUX	alpha1, X, INCX
kusano 2b45e8
kusano 2b45e8
	mr	A1, A
kusano 2b45e8
	mr	YL, Y
kusano 2b45e8
	mr	YS, Y
kusano 2b45e8
	fmul	alpha1, alpha, alpha1
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L35
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	yl2, YL, INCY2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl3, YL, INCY2
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	LFPDUX	yl4, YL, INCY2
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
	bdz	.L33
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L32:
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  yl1
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  yl2
kusano 2b45e8
	LFPDUX	yl2, YL, INCY2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a9,  yl3
kusano 2b45e8
	LFPDUX	yl3, YL, INCY2
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a13, yl4
kusano 2b45e8
	LFPDUX	yl4, YL, INCY2
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	STFPDUX	ys2, YS, INCY2
kusano 2b45e8
	STFPDUX	ys3, YS, INCY2
kusano 2b45e8
	STFPDUX	ys4, YS, INCY2
kusano 2b45e8
	bdnz	.L32
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L33:
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  yl1
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  yl2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a9,  yl3
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a13, yl4
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	STFPDUX	ys2, YS, INCY2
kusano 2b45e8
	STFPDUX	ys3, YS, INCY2
kusano 2b45e8
	STFPDUX	ys4, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L35:
kusano 2b45e8
	andi.	r0, M, 7
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	andi.	r0, M, 4
kusano 2b45e8
	ble	.L37
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl2, YL, INCY2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5, yl2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	STFPDUX	ys2, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L37:
kusano 2b45e8
	andi.	r0, M, 2
kusano 2b45e8
	ble	.L38
kusano 2b45e8
kusano 2b45e8
	LFPDUX	yl1, YL, INCY2
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
kusano 2b45e8
	STFPDUX	ys1, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L38:
kusano 2b45e8
	andi.	r0, M, 1
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY2
kusano 2b45e8
	LFDUX	a1,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY2
kusano 2b45e8
	b	.L999
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L40:
kusano 2b45e8
# A : aligned  LDA : even  Y : Unaligned
kusano 2b45e8
kusano 2b45e8
	sub	A, A, INC2
kusano 2b45e8
	sub	Y, Y, INCY
kusano 2b45e8
kusano 2b45e8
	srawi.	J, N, 2
kusano 2b45e8
	ble	.L50
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L41:
kusano 2b45e8
	LFDUX	alpha1, X, INCX
kusano 2b45e8
	LFSDUX	alpha1, X, INCX
kusano 2b45e8
	LFDUX	alpha2, X, INCX
kusano 2b45e8
	LFSDUX	alpha2, X, INCX
kusano 2b45e8
kusano 2b45e8
	fpmul	alpha1, alpha, alpha1
kusano 2b45e8
	fpmul	alpha2, alpha, alpha2
kusano 2b45e8
kusano 2b45e8
	mr	A1, A
kusano 2b45e8
	add	A2, A,  LDA
kusano 2b45e8
	add	A3, A2, LDA
kusano 2b45e8
	add	A4, A3, LDA
kusano 2b45e8
	add	A,  A4, LDA
kusano 2b45e8
kusano 2b45e8
	mr	YL, Y
kusano 2b45e8
	sub	YS, Y, INCY2
kusano 2b45e8
kusano 2b45e8
	LFSDX	ys1, YS, INCY2
kusano 2b45e8
	LFDX	yl1, YL, INCY
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L45
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
	LFXDUX	yl3, YL, INCY2
kusano 2b45e8
	LFXDUX	yl4, YL, INCY2
kusano 2b45e8
	LFXDUX	yl5, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	LFPDUX	a10, A2, INC2
kusano 2b45e8
	LFPDUX	a14, A2, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	LFPDUX	a7,  A3, INC2
kusano 2b45e8
	LFPDUX	a11, A3, INC2
kusano 2b45e8
	LFPDUX	a15, A3, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	LFPDUX	a8,  A4, INC2
kusano 2b45e8
	fsmr	yl2, yl3
kusano 2b45e8
	LFPDUX	a12, A4, INC2
kusano 2b45e8
	fsmr	yl3, yl4
kusano 2b45e8
	LFPDUX	a16, A4, INC2
kusano 2b45e8
	fsmr	yl4, yl5
kusano 2b45e8
	bdz	.L43
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L42:
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1,  yl1
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a5,  yl2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a9,  yl3
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys5, alpha1, a13, yl4
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a2,  ys2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a6,  ys3
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a10, ys4
kusano 2b45e8
	LFPDUX	a10, A2, INC2
kusano 2b45e8
	fxcsmadd  ys5, alpha1, a14, ys5
kusano 2b45e8
	LFPDUX	a14, A2, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a3,  ys2
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha2, a7,  ys3
kusano 2b45e8
	LFPDUX	a7,  A3, INC2
kusano 2b45e8
	fxcpmadd  ys4, alpha2, a11, ys4
kusano 2b45e8
	LFPDUX	a11, A3, INC2
kusano 2b45e8
	fxcpmadd  ys5, alpha2, a15, ys5
kusano 2b45e8
	LFPDUX	a15, A3, INC2
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a4,  ys2
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
	fxcsmadd  ys3, alpha2, a8,  ys3
kusano 2b45e8
	LFPDUX	a8,  A4, INC2
kusano 2b45e8
	fxcsmadd  ys4, alpha2, a12, ys4
kusano 2b45e8
	LFPDUX	a12, A4, INC2
kusano 2b45e8
	fxcsmadd  ys5, alpha2, a16, ys5
kusano 2b45e8
	LFPDUX	a16, A4, INC2
kusano 2b45e8
kusano 2b45e8
	fmr	yl1, yl5
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	LFXDUX	yl3, YL, INCY2
kusano 2b45e8
	fmr	ys2, ys3
kusano 2b45e8
	LFXDUX	yl4, YL, INCY2
kusano 2b45e8
	fmr	ys3, ys4
kusano 2b45e8
	LFXDUX	yl5, YL, INCY2
kusano 2b45e8
	fmr	ys4, ys5
kusano 2b45e8
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	ys1, ys5
kusano 2b45e8
	STFXDUX	ys2, YS, INCY2
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	STFXDUX	ys3, YS, INCY2
kusano 2b45e8
	fsmr	yl2, yl3
kusano 2b45e8
	STFXDUX	ys4, YS, INCY2
kusano 2b45e8
	fsmr	yl3, yl4
kusano 2b45e8
kusano 2b45e8
	fsmr	yl4, yl5
kusano 2b45e8
	bdnz	.L42
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L43:
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1,  yl1
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a5,  yl2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a9,  yl3
kusano 2b45e8
	fxcpmadd  ys5, alpha1, a13, yl4
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a2,  ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a6,  ys3
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a10, ys4
kusano 2b45e8
	fxcsmadd  ys5, alpha1, a14, ys5
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a3,  ys2
kusano 2b45e8
	fxcpmadd  ys3, alpha2, a7,  ys3
kusano 2b45e8
	fxcpmadd  ys4, alpha2, a11, ys4
kusano 2b45e8
	fxcpmadd  ys5, alpha2, a15, ys5
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a4,  ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha2, a8,  ys3
kusano 2b45e8
	fxcsmadd  ys4, alpha2, a12, ys4
kusano 2b45e8
	fxcsmadd  ys5, alpha2, a16, ys5
kusano 2b45e8
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	fmr	ys2, ys3
kusano 2b45e8
	fmr	ys3, ys4
kusano 2b45e8
	fmr	ys4, ys5
kusano 2b45e8
	fmr	yl1, yl5
kusano 2b45e8
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	ys1, ys5
kusano 2b45e8
	STFXDUX	ys2, YS, INCY2
kusano 2b45e8
	STFXDUX	ys3, YS, INCY2
kusano 2b45e8
	STFXDUX	ys4, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L45:
kusano 2b45e8
	andi.	r0, M, 7
kusano 2b45e8
	ble	.L48
kusano 2b45e8
kusano 2b45e8
	andi.	r0, M, 4
kusano 2b45e8
	ble	.L46
kusano 2b45e8
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
	LFXDUX	yl3, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	LFPDUX	a7,  A3, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	LFPDUX	a8,  A4, INC2
kusano 2b45e8
	fsmr	yl2, yl3
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1, yl1
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a5, yl2
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a2, ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a6, ys3
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a3, ys2
kusano 2b45e8
	fxcpmadd  ys3, alpha2, a7, ys3
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a4, ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha2, a8, ys3
kusano 2b45e8
kusano 2b45e8
	fmr	yl1, yl3
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	fmr	ys2, ys3
kusano 2b45e8
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	ys1, ys3
kusano 2b45e8
	STFXDUX	ys2, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L46:
kusano 2b45e8
	andi.	r0, M, 2
kusano 2b45e8
	ble	.L47
kusano 2b45e8
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a2, ys2
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a3, ys2
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a4, ys2
kusano 2b45e8
	fmr	yl1, yl2
kusano 2b45e8
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	ys1, ys2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L47:
kusano 2b45e8
	andi.	r0, M, 1
kusano 2b45e8
	ble	.L48
kusano 2b45e8
kusano 2b45e8
	LFDUX	a1,  A1, INC2
kusano 2b45e8
	LFDUX	a2,  A2, INC2
kusano 2b45e8
	LFDUX	a3,  A3, INC2
kusano 2b45e8
	LFDUX	a4,  A4, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a2, ys2
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a3, ys2
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a4, ys2
kusano 2b45e8
kusano 2b45e8
	STFSDX	ys1, YS, INCY2
kusano 2b45e8
	add	YS, YS, INCY
kusano 2b45e8
	STFDX	ys2, YS, INCY2
kusano 2b45e8
	b	.L49
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L48:
kusano 2b45e8
	STFSDUX	ys1, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L49:
kusano 2b45e8
	addi	J, J, -1
kusano 2b45e8
	cmpi	cr0, 0, J, 0
kusano 2b45e8
	bgt	.L41
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
.L50:
kusano 2b45e8
	andi.	J, N, 2
kusano 2b45e8
	ble	.L60
kusano 2b45e8
kusano 2b45e8
	LFDUX	alpha1, X, INCX
kusano 2b45e8
kusano 2b45e8
	mr	A1, A
kusano 2b45e8
	add	A2, A,  LDA
kusano 2b45e8
	add	A,  A2, LDA
kusano 2b45e8
	LFSDUX	alpha1, X, INCX
kusano 2b45e8
kusano 2b45e8
	mr	YL, Y
kusano 2b45e8
	sub	YS, Y, INCY2
kusano 2b45e8
	fpmul	alpha1, alpha, alpha1
kusano 2b45e8
kusano 2b45e8
	LFSDX	ys1, YS, INCY2
kusano 2b45e8
	LFDX	yl1, YL, INCY
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L55
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
	LFXDUX	yl3, YL, INCY2
kusano 2b45e8
 	LFXDUX	yl4, YL, INCY2
kusano 2b45e8
	LFXDUX	yl5, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	fsmr	yl2, yl3
kusano 2b45e8
	LFPDUX	a10, A2, INC2
kusano 2b45e8
	fsmr	yl3, yl4
kusano 2b45e8
	LFPDUX	a14, A2, INC2
kusano 2b45e8
	fsmr	yl4, yl5
kusano 2b45e8
	bdz	.L53
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L52:
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1,  yl1
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a5,  yl2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a9,  yl3
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys5, alpha1, a13, yl4
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a2,  ys2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a6,  ys3
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a10, ys4
kusano 2b45e8
	LFPDUX	a10, A2, INC2
kusano 2b45e8
	fxcsmadd  ys5, alpha1, a14, ys5
kusano 2b45e8
	LFPDUX	a14, A2, INC2
kusano 2b45e8
kusano 2b45e8
	fmr	yl1, yl5
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	LFXDUX	yl3, YL, INCY2
kusano 2b45e8
	fmr	ys2, ys3
kusano 2b45e8
 	LFXDUX	yl4, YL, INCY2
kusano 2b45e8
	fmr	ys3, ys4
kusano 2b45e8
	LFXDUX	yl5, YL, INCY2
kusano 2b45e8
	fmr	ys4, ys5
kusano 2b45e8
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	ys1, ys5
kusano 2b45e8
	STFXDUX	ys2, YS, INCY2
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	STFXDUX	ys3, YS, INCY2
kusano 2b45e8
	fsmr	yl2, yl3
kusano 2b45e8
	STFXDUX	ys4, YS, INCY2
kusano 2b45e8
	fsmr	yl3, yl4
kusano 2b45e8
kusano 2b45e8
	fsmr	yl4, yl5
kusano 2b45e8
	bdnz	.L52
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L53:
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1,  yl1
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a5,  yl2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a9,  yl3
kusano 2b45e8
	fxcpmadd  ys5, alpha1, a13, yl4
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a2,  ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a6,  ys3
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a10, ys4
kusano 2b45e8
	fxcsmadd  ys5, alpha1, a14, ys5
kusano 2b45e8
kusano 2b45e8
	fmr	yl1, yl5
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	fmr	ys2, ys3
kusano 2b45e8
	fmr	ys3, ys4
kusano 2b45e8
	fmr	ys4, ys5
kusano 2b45e8
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	ys1, ys5
kusano 2b45e8
	STFXDUX	ys2, YS, INCY2
kusano 2b45e8
	STFXDUX	ys3, YS, INCY2
kusano 2b45e8
	STFXDUX	ys4, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L55:
kusano 2b45e8
	andi.	r0, M, 7
kusano 2b45e8
	ble	.L59
kusano 2b45e8
kusano 2b45e8
	andi.	r0, M, 4
kusano 2b45e8
	ble	.L57
kusano 2b45e8
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
	LFXDUX	yl3, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	fsmr	yl2, yl3
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a2, ys2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a5, yl2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a6, ys3
kusano 2b45e8
kusano 2b45e8
	fmr	yl1, yl3
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	fmr	ys2, ys3
kusano 2b45e8
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	STFXDUX	ys2, YS, INCY2
kusano 2b45e8
	fsmr	  ys1, ys3
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L57:
kusano 2b45e8
	andi.	r0, M, 2
kusano 2b45e8
	ble	.L58
kusano 2b45e8
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a2, ys2
kusano 2b45e8
	fmr	yl1, yl2
kusano 2b45e8
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	ys1, ys2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L58:
kusano 2b45e8
	andi.	r0, M, 1
kusano 2b45e8
	ble	.L59
kusano 2b45e8
kusano 2b45e8
	LFDUX	a1,  A1, INC2
kusano 2b45e8
	LFDUX	a2,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	fxmr	alpha2, alpha1
kusano 2b45e8
	fmadd	ys1, alpha1, a1, yl1
kusano 2b45e8
	fmadd	ys1, alpha2, a2, ys1
kusano 2b45e8
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	b	.L60
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L59:
kusano 2b45e8
	STFSDUX	ys1, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L60:
kusano 2b45e8
	andi.	J, N, 1
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	LFDUX	alpha1, X, INCX
kusano 2b45e8
	mr	A1, A
kusano 2b45e8
kusano 2b45e8
	mr	YL, Y
kusano 2b45e8
	sub	YS, Y, INCY2
kusano 2b45e8
kusano 2b45e8
	fmul	alpha1, alpha, alpha1
kusano 2b45e8
kusano 2b45e8
	LFSDX	ys1, YS, INCY2
kusano 2b45e8
	LFDX	yl1, YL, INCY
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L65
kusano 2b45e8
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
	LFXDUX	yl3, YL, INCY2
kusano 2b45e8
	LFXDUX	yl4, YL, INCY2
kusano 2b45e8
	LFXDUX	yl5, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	fsmr	yl2, yl3
kusano 2b45e8
	fsmr	yl3, yl4
kusano 2b45e8
	fsmr	yl4, yl5
kusano 2b45e8
	bdz	.L63
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L62:
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1,  yl1
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a5,  yl2
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a9,  yl3
kusano 2b45e8
	LFXDUX	yl3, YL, INCY2
kusano 2b45e8
	fxcpmadd  ys5, alpha1, a13, yl4
kusano 2b45e8
	LFXDUX	yl4, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	fmr	yl1, yl5
kusano 2b45e8
	LFXDUX	yl5, YL, INCY2
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	fmr	ys2, ys3
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	fmr	ys3, ys4
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
	fmr	ys4, ys5
kusano 2b45e8
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	ys1, ys5
kusano 2b45e8
	STFXDUX	ys2, YS, INCY2
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	STFXDUX	ys3, YS, INCY2
kusano 2b45e8
	fsmr	yl2, yl3
kusano 2b45e8
	STFXDUX	ys4, YS, INCY2
kusano 2b45e8
	fsmr	yl3, yl4
kusano 2b45e8
kusano 2b45e8
	fsmr	yl4, yl5
kusano 2b45e8
	bdnz	.L62
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L63:
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1,  yl1
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a5,  yl2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a9,  yl3
kusano 2b45e8
	fxcpmadd  ys5, alpha1, a13, yl4
kusano 2b45e8
kusano 2b45e8
	fmr	yl1, yl5
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	fmr	ys2, ys3
kusano 2b45e8
	fmr	ys3, ys4
kusano 2b45e8
	fmr	ys4, ys5
kusano 2b45e8
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	ys1, ys5
kusano 2b45e8
	STFXDUX	ys2, YS, INCY2
kusano 2b45e8
	STFXDUX	ys3, YS, INCY2
kusano 2b45e8
	STFXDUX	ys4, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L65:
kusano 2b45e8
	andi.	r0, M, 7
kusano 2b45e8
	ble	.L69
kusano 2b45e8
kusano 2b45e8
	andi.	r0, M, 4
kusano 2b45e8
	ble	.L67
kusano 2b45e8
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
	LFXDUX	yl3, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	fsmr	yl2, yl3
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1, yl1
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a5, yl2
kusano 2b45e8
kusano 2b45e8
	fmr	yl1, yl3
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	fmr	ys2, ys3
kusano 2b45e8
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	  ys1, ys3
kusano 2b45e8
	STFXDUX	ys2, YS, INCY2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L67:
kusano 2b45e8
	andi.	r0, M, 2
kusano 2b45e8
	ble	.L68
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFXDUX	yl2, YL, INCY2
kusano 2b45e8
kusano 2b45e8
	fsmr	yl1, yl2
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a1, yl1
kusano 2b45e8
	fmr	yl1, yl2
kusano 2b45e8
	fmr	ys1, ys2
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	fsmr	ys1, ys2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L68:
kusano 2b45e8
	andi.	r0, M, 1
kusano 2b45e8
	ble	.L69
kusano 2b45e8
kusano 2b45e8
	LFDUX	a1,  A1, INC2
kusano 2b45e8
	fmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	STFXDUX	ys1, YS, INCY2
kusano 2b45e8
	b	.L999
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L69:
kusano 2b45e8
	STFSDUX	ys1, YS, INCY2
kusano 2b45e8
	b	.L999
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L70:
kusano 2b45e8
	sub	A, A, INC2
kusano 2b45e8
	sub	Y, Y, INCY
kusano 2b45e8
	srawi.	J, N, 2
kusano 2b45e8
	ble	.L80
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L71:
kusano 2b45e8
	LFDUX	alpha1, X, INCX
kusano 2b45e8
	mr	A1, A
kusano 2b45e8
	add	A2, A,  LDA
kusano 2b45e8
	add	A3, A2, LDA
kusano 2b45e8
	LFSDUX	alpha1, X, INCX
kusano 2b45e8
	LFDUX	alpha2, X, INCX
kusano 2b45e8
	add	A4, A3, LDA
kusano 2b45e8
	add	A,  A4, LDA
kusano 2b45e8
	mr	YL, Y
kusano 2b45e8
	LFSDUX	alpha2, X, INCX
kusano 2b45e8
	fpmul	alpha1, alpha, alpha1
kusano 2b45e8
	mr	YS, Y
kusano 2b45e8
	srawi.	r0,  M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	fpmul	alpha2, alpha, alpha2
kusano 2b45e8
	ble	.L75
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
	LFSDUX	yl1, YL, INCY
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl2, YL, INCY
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	LFPDUX	a10, A2, INC2
kusano 2b45e8
	LFPDUX	a14, A2, INC2
kusano 2b45e8
	LFSDUX	yl2, YL, INCY
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl3, YL, INCY
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	LFPDUX	a7,  A3, INC2
kusano 2b45e8
	LFPDUX	a11, A3, INC2
kusano 2b45e8
	LFPDUX	a15, A3, INC2
kusano 2b45e8
	LFSDUX	yl3, YL, INCY
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl4, YL, INCY
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
	LFPDUX	a8,  A4, INC2
kusano 2b45e8
	LFPDUX	a12, A4, INC2
kusano 2b45e8
	LFPDUX	a16, A4, INC2
kusano 2b45e8
	LFSDUX	yl4, YL, INCY
kusano 2b45e8
	bdz	.L73
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L72:
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  yl1
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  yl2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a9,  yl3
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a13, yl4
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
	LFSDUX	yl1, YL, INCY
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2,  ys1
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFDUX	yl2, YL, INCY
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6,  ys2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a10, ys3
kusano 2b45e8
	LFPDUX	a10, A2, INC2
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a14, ys4
kusano 2b45e8
	LFPDUX	a14, A2, INC2
kusano 2b45e8
	LFSDUX	yl2, YL, INCY
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha2, a3,  ys1
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	LFDUX	yl3, YL, INCY
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a7,  ys2
kusano 2b45e8
	LFPDUX	a7,  A3, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha2, a11, ys3
kusano 2b45e8
	LFPDUX	a11, A3, INC2
kusano 2b45e8
	fxcpmadd  ys4, alpha2, a15, ys4
kusano 2b45e8
	LFPDUX	a15, A3, INC2
kusano 2b45e8
	LFSDUX	yl3, YL, INCY
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha2, a4,  ys1
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
	LFDUX	yl4, YL, INCY
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a8,  ys2
kusano 2b45e8
	LFPDUX	a8,  A4, INC2
kusano 2b45e8
	fxcsmadd  ys3, alpha2, a12, ys3
kusano 2b45e8
	LFPDUX	a12, A4, INC2
kusano 2b45e8
	fxcsmadd  ys4, alpha2, a16, ys4
kusano 2b45e8
	LFPDUX	a16, A4, INC2
kusano 2b45e8
	LFSDUX	yl4, YL, INCY
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFSDUX	ys1, YS, INCY
kusano 2b45e8
	STFDUX	ys2, YS, INCY
kusano 2b45e8
	STFSDUX	ys2, YS, INCY
kusano 2b45e8
	STFDUX	ys3, YS, INCY
kusano 2b45e8
	STFSDUX	ys3, YS, INCY
kusano 2b45e8
	STFDUX	ys4, YS, INCY
kusano 2b45e8
	STFSDUX	ys4, YS, INCY
kusano 2b45e8
	bdnz	.L72
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L73:
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  yl1
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  yl2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a9,  yl3
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a13, yl4
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2,  ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6,  ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a10, ys3
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a14, ys4
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha2, a3,  ys1
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a7,  ys2
kusano 2b45e8
	fxcpmadd  ys3, alpha2, a11, ys3
kusano 2b45e8
	fxcpmadd  ys4, alpha2, a15, ys4
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha2, a4,  ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a8,  ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha2, a12, ys3
kusano 2b45e8
	fxcsmadd  ys4, alpha2, a16, ys4
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFSDUX	ys1, YS, INCY
kusano 2b45e8
	STFDUX	ys2, YS, INCY
kusano 2b45e8
	STFSDUX	ys2, YS, INCY
kusano 2b45e8
	STFDUX	ys3, YS, INCY
kusano 2b45e8
	STFSDUX	ys3, YS, INCY
kusano 2b45e8
	STFDUX	ys4, YS, INCY
kusano 2b45e8
	STFSDUX	ys4, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L75:
kusano 2b45e8
	andi.	r0, M, 7
kusano 2b45e8
	ble	.L79
kusano 2b45e8
kusano 2b45e8
	andi.	r0, M, 4
kusano 2b45e8
	ble	.L77
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFSDUX	yl1, YL, INCY
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl2, YL, INCY
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	LFPDUX	a7,  A3, INC2
kusano 2b45e8
	LFSDUX	yl2, YL, INCY
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
	LFPDUX	a8,  A4, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5, yl2
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6, ys2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha2, a3, ys1
kusano 2b45e8
	fxcpmadd  ys2, alpha2, a7, ys2
kusano 2b45e8
	fxcsmadd  ys1, alpha2, a4, ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha2, a8, ys2
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFSDUX	ys1, YS, INCY
kusano 2b45e8
	STFDUX	ys2, YS, INCY
kusano 2b45e8
	STFSDUX	ys2, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L77:
kusano 2b45e8
	andi.	r0, M, 2
kusano 2b45e8
	ble	.L78
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFSDUX	yl1, YL, INCY
kusano 2b45e8
	LFPDUX	a3,  A3, INC2
kusano 2b45e8
	LFPDUX	a4,  A4, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
	fxcpmadd  ys1, alpha2, a3, ys1
kusano 2b45e8
	fxcsmadd  ys1, alpha2, a4, ys1
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFSDUX	ys1, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L78:
kusano 2b45e8
	andi.	r0, M, 1
kusano 2b45e8
	ble	.L79
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
kusano 2b45e8
	LFDUX	a1,  A1, INC2
kusano 2b45e8
	LFDUX	a2,  A2, INC2
kusano 2b45e8
	LFDUX	a3,  A3, INC2
kusano 2b45e8
	LFDUX	a4,  A4, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
	fxcpmadd  ys1, alpha2, a3, ys1
kusano 2b45e8
	fxcsmadd  ys1, alpha2, a4, ys1
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L79:
kusano 2b45e8
	addi	J, J, -1
kusano 2b45e8
	cmpi	cr0, 0, J, 0
kusano 2b45e8
	bgt	.L71
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
.L80:
kusano 2b45e8
	andi.	J, N, 2
kusano 2b45e8
	ble	.L90
kusano 2b45e8
kusano 2b45e8
	LFDUX	alpha1, X, INCX
kusano 2b45e8
kusano 2b45e8
	mr	A1, A
kusano 2b45e8
	add	A2, A,  LDA
kusano 2b45e8
	add	A,  A2, LDA
kusano 2b45e8
	LFSDUX	alpha1, X, INCX
kusano 2b45e8
kusano 2b45e8
	mr	YL, Y
kusano 2b45e8
	mr	YS, Y
kusano 2b45e8
	fpmul	alpha1, alpha, alpha1
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L85
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFDUX	a9,  YL, INCY
kusano 2b45e8
	LFDUX	yl2, YL, INCY
kusano 2b45e8
	LFDUX	a10, YL, INCY
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a3,  A1, INC2
kusano 2b45e8
	LFPDUX	a7,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl3, YL, INCY
kusano 2b45e8
	LFDUX	a11, YL, INCY
kusano 2b45e8
	LFDUX	yl4, YL, INCY
kusano 2b45e8
	LFDUX	a12, YL, INCY
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	LFPDUX	a4,  A2, INC2
kusano 2b45e8
	LFPDUX	a8,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	bdz	.L83
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L82:
kusano 2b45e8
	fsmfp	yl1, a9
kusano 2b45e8
	fsmfp	yl2, a10
kusano 2b45e8
	fsmfp	yl3, a11
kusano 2b45e8
	fsmfp	yl4, a12
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  yl1
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFDUX	a9,  YL, INCY
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  yl2
kusano 2b45e8
	LFDUX	yl2, YL, INCY
kusano 2b45e8
	LFDUX	a10, YL, INCY
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a3,  yl3
kusano 2b45e8
	LFDUX	yl3, YL, INCY
kusano 2b45e8
	LFDUX	a11, YL, INCY
kusano 2b45e8
	LFPDUX	a3,  A1, INC2
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a7,  yl4
kusano 2b45e8
	LFDUX	yl4, YL, INCY
kusano 2b45e8
	LFDUX	a12, YL, INCY
kusano 2b45e8
	LFPDUX	a7,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2,  ys1
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6,  ys2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a4,  ys3
kusano 2b45e8
	LFPDUX	a4,  A2, INC2
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a8,  ys4
kusano 2b45e8
	LFPDUX	a8,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFSDUX	ys1, YS, INCY
kusano 2b45e8
	STFDUX	ys2, YS, INCY
kusano 2b45e8
	STFSDUX	ys2, YS, INCY
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys3, YS, INCY
kusano 2b45e8
	STFSDUX	ys3, YS, INCY
kusano 2b45e8
	STFDUX	ys4, YS, INCY
kusano 2b45e8
	STFSDUX	ys4, YS, INCY
kusano 2b45e8
	bdnz	.L82
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L83:
kusano 2b45e8
	fsmfp	yl1, a9
kusano 2b45e8
	fsmfp	yl2, a10
kusano 2b45e8
	fsmfp	yl3, a11
kusano 2b45e8
	fsmfp	yl4, a12
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  yl1
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  yl2
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a3,  yl3
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a7,  yl4
kusano 2b45e8
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2,  ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6,  ys2
kusano 2b45e8
	fxcsmadd  ys3, alpha1, a4,  ys3
kusano 2b45e8
	fxcsmadd  ys4, alpha1, a8,  ys4
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFSDUX	ys1, YS, INCY
kusano 2b45e8
	STFDUX	ys2, YS, INCY
kusano 2b45e8
	STFSDUX	ys2, YS, INCY
kusano 2b45e8
	STFDUX	ys3, YS, INCY
kusano 2b45e8
	STFSDUX	ys3, YS, INCY
kusano 2b45e8
	STFDUX	ys4, YS, INCY
kusano 2b45e8
	STFSDUX	ys4, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L85:
kusano 2b45e8
	andi.	r0, M, 7
kusano 2b45e8
	ble	.L90
kusano 2b45e8
kusano 2b45e8
	andi.	r0, M, 4
kusano 2b45e8
	ble	.L87
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFSDUX	yl1, YL, INCY
kusano 2b45e8
	LFDUX	yl2, YL, INCY
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a6,  A2, INC2
kusano 2b45e8
	LFSDUX	yl2, YL, INCY
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5, yl2
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
	fxcsmadd  ys2, alpha1, a6, ys2
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFSDUX	ys1, YS, INCY
kusano 2b45e8
	STFDUX	ys2, YS, INCY
kusano 2b45e8
	STFSDUX	ys2, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L87:
kusano 2b45e8
	andi.	r0, M, 2
kusano 2b45e8
	ble	.L88
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a2,  A2, INC2
kusano 2b45e8
	LFSDUX	yl1, YL, INCY
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFSDUX	ys1, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L88:
kusano 2b45e8
	andi.	r0, M, 1
kusano 2b45e8
	ble	.L90
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFDUX	a1,  A1, INC2
kusano 2b45e8
	LFDUX	a2,  A2, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
	fxcsmadd  ys1, alpha1, a2, ys1
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L90:
kusano 2b45e8
	andi.	J, N, 1
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	LFDUX	alpha1, X, INCX
kusano 2b45e8
kusano 2b45e8
	mr	A1, A
kusano 2b45e8
	mr	YL, Y
kusano 2b45e8
	mr	YS, Y
kusano 2b45e8
	fmul	alpha1, alpha, alpha1
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L95
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFSDUX	a2,  YL, INCY
kusano 2b45e8
	LFDUX	yl2, YL, INCY
kusano 2b45e8
	LFSDUX	a4,  YL, INCY
kusano 2b45e8
	LFDUX	yl3, YL, INCY
kusano 2b45e8
	LFSDUX	a6,  YL, INCY
kusano 2b45e8
	LFDUX	yl4, YL, INCY
kusano 2b45e8
	LFSDUX	a8,  YL, INCY
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
	bdz	.L93
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L92:
kusano 2b45e8
	fmr	a2, yl1
kusano 2b45e8
	fmr	a4, yl2
kusano 2b45e8
	fmr	a6, yl3
kusano 2b45e8
	fmr	a8, yl4
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  a2
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFSDUX	a2,  YL, INCY
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  a4
kusano 2b45e8
	LFDUX	yl2, YL, INCY
kusano 2b45e8
	LFSDUX	a4,  YL, INCY
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a9,  a6
kusano 2b45e8
	LFDUX	yl3, YL, INCY
kusano 2b45e8
	LFSDUX	a6,  YL, INCY
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a13, a8
kusano 2b45e8
	LFDUX	yl4, YL, INCY
kusano 2b45e8
	LFSDUX	a8,  YL, INCY
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFPDUX	a5,  A1, INC2
kusano 2b45e8
	LFPDUX	a9,  A1, INC2
kusano 2b45e8
	LFPDUX	a13, A1, INC2
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFSDUX	ys1, YS, INCY
kusano 2b45e8
	STFDUX	ys2, YS, INCY
kusano 2b45e8
	STFSDUX	ys2, YS, INCY
kusano 2b45e8
	STFDUX	ys3, YS, INCY
kusano 2b45e8
	STFSDUX	ys3, YS, INCY
kusano 2b45e8
	STFDUX	ys4, YS, INCY
kusano 2b45e8
	STFSDUX	ys4, YS, INCY
kusano 2b45e8
	bdnz	.L92
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L93:
kusano 2b45e8
	fmr	a2, yl1
kusano 2b45e8
	fmr	a4, yl2
kusano 2b45e8
	fmr	a6, yl3
kusano 2b45e8
	fmr	a8, yl4
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1,  a2
kusano 2b45e8
	fxcpmadd  ys2, alpha1, a5,  a4
kusano 2b45e8
	fxcpmadd  ys3, alpha1, a9,  a6
kusano 2b45e8
	fxcpmadd  ys4, alpha1, a13, a8
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFSDUX	ys1, YS, INCY
kusano 2b45e8
	STFDUX	ys2, YS, INCY
kusano 2b45e8
	STFSDUX	ys2, YS, INCY
kusano 2b45e8
	STFDUX	ys3, YS, INCY
kusano 2b45e8
	STFSDUX	ys3, YS, INCY
kusano 2b45e8
	STFDUX	ys4, YS, INCY
kusano 2b45e8
	STFSDUX	ys4, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L95:
kusano 2b45e8
	andi.	r0, M, 7
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	andi.	r0, M, 4
kusano 2b45e8
	ble	.L97
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFDUX	yl2, YL, INCY
kusano 2b45e8
	LFPDUX	a2,  A1, INC2
kusano 2b45e8
	LFDUX	yl3, YL, INCY
kusano 2b45e8
	LFDUX	yl4, YL, INCY
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, a1, alpha1, yl1
kusano 2b45e8
	fxcsmadd  ys2, a1, alpha1, yl2
kusano 2b45e8
	fxcpmadd  ys3, a2, alpha1, yl3
kusano 2b45e8
	fxcsmadd  ys4, a2, alpha1, yl4
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFDUX	ys2, YS, INCY
kusano 2b45e8
	STFDUX	ys3, YS, INCY
kusano 2b45e8
	STFDUX	ys4, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L97:
kusano 2b45e8
	andi.	r0, M, 2
kusano 2b45e8
	ble	.L98
kusano 2b45e8
kusano 2b45e8
	LFPDUX	a1,  A1, INC2
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFDUX	yl2, YL, INCY
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, a1, alpha1, yl1
kusano 2b45e8
	fxcsmadd  ys2, a1, alpha1, yl2
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	STFDUX	ys2, YS, INCY
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L98:
kusano 2b45e8
	andi.	r0, M, 1
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	LFDUX	yl1, YL, INCY
kusano 2b45e8
	LFDUX	a1,  A1, INC2
kusano 2b45e8
kusano 2b45e8
	fxcpmadd  ys1, alpha1, a1, yl1
kusano 2b45e8
kusano 2b45e8
	STFDUX	ys1, YS, INCY
kusano 2b45e8
	b	.L999
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
.L999:
kusano 2b45e8
	addi	SP, SP, -4
kusano 2b45e8
kusano 2b45e8
	lwzu	r16,   4(SP)
kusano 2b45e8
	lwzu	r17,   4(SP)
kusano 2b45e8
	lwzu	r18,   4(SP)
kusano 2b45e8
	lwzu	r19,   4(SP)
kusano 2b45e8
kusano 2b45e8
	lwzu	r20,   4(SP)
kusano 2b45e8
	lwzu	r21,   4(SP)
kusano 2b45e8
	lwzu	r22,   4(SP)
kusano 2b45e8
	lwzu	r23,   4(SP)
kusano 2b45e8
kusano 2b45e8
	lwzu	r24,   4(SP)
kusano 2b45e8
	lwzu	r25,   4(SP)
kusano 2b45e8
	lwzu	r26,   4(SP)
kusano 2b45e8
	lwzu	r27,   4(SP)
kusano 2b45e8
kusano 2b45e8
	lwzu	r28,   4(SP)
kusano 2b45e8
	lwzu	r29,   4(SP)
kusano 2b45e8
	lwzu	r30,   4(SP)
kusano 2b45e8
	lwzu	r31,   4(SP)
kusano 2b45e8
kusano 2b45e8
	subi	SP, SP, 12
kusano 2b45e8
	li	r0, 16
kusano 2b45e8
kusano 2b45e8
	lfpdux	f31, SP, r0
kusano 2b45e8
	lfpdux	f30, SP, r0
kusano 2b45e8
	lfpdux	f29, SP, r0
kusano 2b45e8
	lfpdux	f28, SP, r0
kusano 2b45e8
	lfpdux	f27, SP, r0
kusano 2b45e8
	lfpdux	f26, SP, r0
kusano 2b45e8
	lfpdux	f25, SP, r0
kusano 2b45e8
	lfpdux	f24, SP, r0
kusano 2b45e8
	lfpdux	f23, SP, r0
kusano 2b45e8
	lfpdux	f22, SP, r0
kusano 2b45e8
	lfpdux	f21, SP, r0
kusano 2b45e8
	lfpdux	f20, SP, r0
kusano 2b45e8
	lfpdux	f19, SP, r0
kusano 2b45e8
	lfpdux	f18, SP, r0
kusano 2b45e8
	lfpdux	f17, SP, r0
kusano 2b45e8
	lfpdux	f16, SP, r0
kusano 2b45e8
	lfpdux	f15, SP, r0
kusano 2b45e8
	lfpdux	f14, SP, r0
kusano 2b45e8
	addi	SP, SP, 16
kusano 2b45e8
	blr
kusano 2b45e8
kusano 2b45e8
	EPILOGUE