Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/power/gemm_ncopy_hummer_8.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	r5
kusano 2b45e8
#define	LDA	r6
kusano 2b45e8
#define B	r7
kusano 2b45e8
kusano 2b45e8
#define AO1	r8
kusano 2b45e8
#define AO2	r9
kusano 2b45e8
#define AO3	r10
kusano 2b45e8
#define AO4	r11
kusano 2b45e8
kusano 2b45e8
#define J	r12
kusano 2b45e8
kusano 2b45e8
#define AO5	r26
kusano 2b45e8
#define AO6	r27
kusano 2b45e8
#define AO7	r28
kusano 2b45e8
#define AO8	r29
kusano 2b45e8
#define INC	r30
kusano 2b45e8
#define INC2	r31
kusano 2b45e8
kusano 2b45e8
#define c01	f0
kusano 2b45e8
#define c02	f1
kusano 2b45e8
#define c03	f2
kusano 2b45e8
#define c04	f3
kusano 2b45e8
#define c05	f4
kusano 2b45e8
#define c06	f5
kusano 2b45e8
#define c07	f6
kusano 2b45e8
#define c08	f7
kusano 2b45e8
#define c09	f8
kusano 2b45e8
#define c10	f9
kusano 2b45e8
#define c11	f10
kusano 2b45e8
#define c12	f11
kusano 2b45e8
#define c13	f12
kusano 2b45e8
#define c14	f13
kusano 2b45e8
#define c15	f14
kusano 2b45e8
#define c16	f15
kusano 2b45e8
kusano 2b45e8
#define c17	f16
kusano 2b45e8
#define c18	f17
kusano 2b45e8
#define c19	f18
kusano 2b45e8
#define c20	f19
kusano 2b45e8
#define c21	f20
kusano 2b45e8
#define c22	f21
kusano 2b45e8
#define c23	f22
kusano 2b45e8
#define c24	f23
kusano 2b45e8
#define c25	f24
kusano 2b45e8
#define c26	f25
kusano 2b45e8
#define c27	f26
kusano 2b45e8
#define c28	f27
kusano 2b45e8
#define c29	f28
kusano 2b45e8
#define c30	f29
kusano 2b45e8
#define c31	f30
kusano 2b45e8
#define c32	f31
kusano 2b45e8
kusano 2b45e8
#define	sel_p	f30
kusano 2b45e8
#define	sel_s	f31
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
	li	r0, -16
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
kusano 2b45e8
	lis	r9,   0x3f80
kusano 2b45e8
	lis	r10,  0xbf80
kusano 2b45e8
kusano 2b45e8
	stwu	r9,    -4(SP)
kusano 2b45e8
	stwu	r10,   -4(SP)
kusano 2b45e8
	stwu	r10,   -4(SP)
kusano 2b45e8
	stwu	r9,    -4(SP)
kusano 2b45e8
kusano 2b45e8
	slwi	LDA, LDA, BASE_SHIFT
kusano 2b45e8
kusano 2b45e8
	li	r0, 0
kusano 2b45e8
	lfpsux	sel_p, SP, r0
kusano 2b45e8
	li	r0, 8
kusano 2b45e8
	lfpsux	sel_s, SP, r0
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
	li	INC,  1 * SIZE
kusano 2b45e8
	li	INC2, 2 * SIZE
kusano 2b45e8
kusano 2b45e8
	subi	B, B, 2 * SIZE
kusano 2b45e8
kusano 2b45e8
	andi.	r0, A,   2 * SIZE - 1
kusano 2b45e8
	bne	.L100
kusano 2b45e8
	andi.	r0, LDA, 2 * SIZE - 1
kusano 2b45e8
	bne	.L100
kusano 2b45e8
kusano 2b45e8
	subi	A, A, 2 * SIZE
kusano 2b45e8
	srawi.	J,  N,  3
kusano 2b45e8
	ble	.L20
kusano 2b45e8
	.align 4
kusano 2b45e8
.L11:
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	AO2, A,   LDA
kusano 2b45e8
	add	AO3, AO2, LDA
kusano 2b45e8
	add	AO4, AO3, LDA
kusano 2b45e8
	add	AO5, AO4, LDA
kusano 2b45e8
	add	AO6, AO5, LDA
kusano 2b45e8
	add	AO7, AO6, LDA
kusano 2b45e8
	add	AO8, AO7, LDA
kusano 2b45e8
	add	A,   AO8, LDA
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M,  2
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L15
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L12:
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFXDUX	c02,   AO2, INC2
kusano 2b45e8
	LFPDUX	c03,   AO3, INC2
kusano 2b45e8
	LFXDUX	c04,   AO4, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c05,   AO5, INC2
kusano 2b45e8
	LFXDUX	c06,   AO6, INC2
kusano 2b45e8
	LFPDUX	c07,   AO7, INC2
kusano 2b45e8
	LFXDUX	c08,   AO8, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c09,   AO1, INC2
kusano 2b45e8
	LFXDUX	c10,   AO2, INC2
kusano 2b45e8
	LFPDUX	c11,   AO3, INC2
kusano 2b45e8
	LFXDUX	c12,   AO4, INC2
kusano 2b45e8
	fpsel	c17, sel_p, c01, c02
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c13,   AO5, INC2
kusano 2b45e8
	fpsel	c18, sel_p, c03, c04
kusano 2b45e8
	LFXDUX	c14,   AO6, INC2
kusano 2b45e8
	fpsel	c19, sel_p, c05, c06
kusano 2b45e8
	LFPDUX	c15,   AO7, INC2
kusano 2b45e8
	fpsel	c20, sel_p, c07, c08
kusano 2b45e8
	LFXDUX	c16,   AO8, INC2
kusano 2b45e8
	fpsel	c21, sel_s, c01, c02
kusano 2b45e8
kusano 2b45e8
	fpsel	c22, sel_s, c03, c04
kusano 2b45e8
	STFPDUX	c17,   B, INC2
kusano 2b45e8
	fpsel	c23, sel_s, c05, c06
kusano 2b45e8
	STFPDUX	c18,   B, INC2
kusano 2b45e8
	fpsel	c24, sel_s, c07, c08
kusano 2b45e8
	STFPDUX	c19,   B, INC2
kusano 2b45e8
kusano 2b45e8
	fpsel	c01, sel_p, c09, c10
kusano 2b45e8
	STFPDUX	c20,   B, INC2
kusano 2b45e8
	fpsel	c02, sel_p, c11, c12
kusano 2b45e8
	STFXDUX	c21,   B, INC2
kusano 2b45e8
	fpsel	c03, sel_p, c13, c14
kusano 2b45e8
	STFXDUX	c22,   B, INC2
kusano 2b45e8
	fpsel	c04, sel_p, c15, c16
kusano 2b45e8
	STFXDUX	c23,   B, INC2
kusano 2b45e8
kusano 2b45e8
	fpsel	c05, sel_s, c09, c10
kusano 2b45e8
	STFXDUX	c24,   B, INC2
kusano 2b45e8
	fpsel	c06, sel_s, c11, c12
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	fpsel	c07, sel_s, c13, c14
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	fpsel	c08, sel_s, c15, c16
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
	STFXDUX	c05,   B, INC2
kusano 2b45e8
	STFXDUX	c06,   B, INC2
kusano 2b45e8
	STFXDUX	c07,   B, INC2
kusano 2b45e8
	STFXDUX	c08,   B, INC2
kusano 2b45e8
	bdnz	.L12
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
.L15:
kusano 2b45e8
	andi.	r0,  M,  3
kusano 2b45e8
	ble	.L19
kusano 2b45e8
kusano 2b45e8
	andi.	r0,  M,  2
kusano 2b45e8
	beq	.L17
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFXDUX	c02,   AO2, INC2
kusano 2b45e8
	LFPDUX	c03,   AO3, INC2
kusano 2b45e8
	LFXDUX	c04,   AO4, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c05,   AO5, INC2
kusano 2b45e8
	fpsel	c09, sel_p, c01, c02
kusano 2b45e8
	LFXDUX	c06,   AO6, INC2
kusano 2b45e8
	fpsel	c10, sel_p, c03, c04
kusano 2b45e8
	LFPDUX	c07,   AO7, INC2
kusano 2b45e8
	fpsel	c11, sel_p, c05, c06
kusano 2b45e8
	LFXDUX	c08,   AO8, INC2
kusano 2b45e8
	fpsel	c12, sel_p, c07, c08
kusano 2b45e8
kusano 2b45e8
	fpsel	c13, sel_s, c01, c02
kusano 2b45e8
	fpsel	c14, sel_s, c03, c04
kusano 2b45e8
	STFPDUX	c09,   B, INC2
kusano 2b45e8
	fpsel	c15, sel_s, c05, c06
kusano 2b45e8
	STFPDUX	c10,   B, INC2
kusano 2b45e8
	fpsel	c16, sel_s, c07, c08
kusano 2b45e8
	STFPDUX	c11,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c12,   B, INC2
kusano 2b45e8
	STFXDUX	c13,   B, INC2
kusano 2b45e8
	STFXDUX	c14,   B, INC2
kusano 2b45e8
	STFXDUX	c15,   B, INC2
kusano 2b45e8
	STFXDUX	c16,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L17:
kusano 2b45e8
	andi.	r0,  M,  1
kusano 2b45e8
	beq	.L19
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC2
kusano 2b45e8
	LFDUX	c02,   AO3, INC2
kusano 2b45e8
	LFDUX	c03,   AO5, INC2
kusano 2b45e8
	LFDUX	c04,   AO7, INC2
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC2
kusano 2b45e8
	LFSDUX	c02,   AO4, INC2
kusano 2b45e8
	LFSDUX	c03,   AO6, INC2
kusano 2b45e8
	LFSDUX	c04,   AO8, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L19:
kusano 2b45e8
	addic.	J, J, -1
kusano 2b45e8
	bgt	.L11
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L20:
kusano 2b45e8
	andi.	J,  N,  4
kusano 2b45e8
	ble	.L30
kusano 2b45e8
	.align 4
kusano 2b45e8
.L21:
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	AO2, A,   LDA
kusano 2b45e8
	add	AO3, AO2, LDA
kusano 2b45e8
	add	AO4, AO3, LDA
kusano 2b45e8
	add	A,   AO4, LDA
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M,  3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L25
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L22:
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFXDUX	c02,   AO2, INC2
kusano 2b45e8
	LFPDUX	c03,   AO3, INC2
kusano 2b45e8
	LFXDUX	c04,   AO4, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c05,   AO1, INC2
kusano 2b45e8
	LFXDUX	c06,   AO2, INC2
kusano 2b45e8
	LFPDUX	c07,   AO3, INC2
kusano 2b45e8
	LFXDUX	c08,   AO4, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c09,   AO1, INC2
kusano 2b45e8
	LFXDUX	c10,   AO2, INC2
kusano 2b45e8
	LFPDUX	c11,   AO3, INC2
kusano 2b45e8
	LFXDUX	c12,   AO4, INC2
kusano 2b45e8
	fpsel	c17, sel_p, c01, c02
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c13,   AO1, INC2
kusano 2b45e8
	fpsel	c18, sel_p, c03, c04
kusano 2b45e8
	LFXDUX	c14,   AO2, INC2
kusano 2b45e8
	fpsel	c19, sel_s, c01, c02
kusano 2b45e8
	LFPDUX	c15,   AO3, INC2
kusano 2b45e8
	fpsel	c20, sel_s, c03, c04
kusano 2b45e8
	LFXDUX	c16,   AO4, INC2
kusano 2b45e8
	fpsel	c21, sel_p, c05, c06
kusano 2b45e8
kusano 2b45e8
	fpsel	c22, sel_p, c07, c08
kusano 2b45e8
	STFPDUX	c17,   B, INC2
kusano 2b45e8
	fpsel	c23, sel_s, c05, c06
kusano 2b45e8
	STFPDUX	c18,   B, INC2
kusano 2b45e8
	fpsel	c24, sel_s, c07, c08
kusano 2b45e8
	STFXDUX	c19,   B, INC2
kusano 2b45e8
kusano 2b45e8
	fpsel	c01, sel_p, c09, c10
kusano 2b45e8
	STFXDUX	c20,   B, INC2
kusano 2b45e8
	fpsel	c02, sel_p, c11, c12
kusano 2b45e8
	STFPDUX	c21,   B, INC2
kusano 2b45e8
	fpsel	c03, sel_s, c09, c10
kusano 2b45e8
	STFPDUX	c22,   B, INC2
kusano 2b45e8
	fpsel	c04, sel_s, c11, c12
kusano 2b45e8
	STFXDUX	c23,   B, INC2
kusano 2b45e8
kusano 2b45e8
	fpsel	c05, sel_p, c13, c14
kusano 2b45e8
	STFXDUX	c24,   B, INC2
kusano 2b45e8
	fpsel	c06, sel_p, c15, c16
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	fpsel	c07, sel_s, c13, c14
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	fpsel	c08, sel_s, c15, c16
kusano 2b45e8
	STFXDUX	c03,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFXDUX	c04,   B, INC2
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	STFPDUX	c06,   B, INC2
kusano 2b45e8
	STFXDUX	c07,   B, INC2
kusano 2b45e8
	STFXDUX	c08,   B, INC2
kusano 2b45e8
	bdnz	.L22
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
	beq	.L26
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFXDUX	c02,   AO2, INC2
kusano 2b45e8
	LFPDUX	c03,   AO3, INC2
kusano 2b45e8
	LFXDUX	c04,   AO4, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c05,   AO1, INC2
kusano 2b45e8
	fpsel	c09, sel_p, c01, c02
kusano 2b45e8
	LFXDUX	c06,   AO2, INC2
kusano 2b45e8
	fpsel	c10, sel_p, c03, c04
kusano 2b45e8
	LFPDUX	c07,   AO3, INC2
kusano 2b45e8
	fpsel	c11, sel_s, c01, c02
kusano 2b45e8
	LFXDUX	c08,   AO4, INC2
kusano 2b45e8
	fpsel	c12, sel_s, c03, c04
kusano 2b45e8
kusano 2b45e8
	fpsel	c13, sel_p, c05, c06
kusano 2b45e8
	fpsel	c14, sel_p, c07, c08
kusano 2b45e8
	STFPDUX	c09,   B, INC2
kusano 2b45e8
	fpsel	c15, sel_s, c05, c06
kusano 2b45e8
	STFPDUX	c10,   B, INC2
kusano 2b45e8
	fpsel	c16, sel_s, c07, c08
kusano 2b45e8
	STFXDUX	c11,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFXDUX	c12,   B, INC2
kusano 2b45e8
	STFPDUX	c13,   B, INC2
kusano 2b45e8
	STFPDUX	c14,   B, INC2
kusano 2b45e8
	STFXDUX	c15,   B, INC2
kusano 2b45e8
	STFXDUX	c16,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L26:
kusano 2b45e8
	andi.	r0,  M,  2
kusano 2b45e8
	beq	.L27
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFXDUX	c02,   AO2, INC2
kusano 2b45e8
	LFPDUX	c03,   AO3, INC2
kusano 2b45e8
	LFXDUX	c04,   AO4, INC2
kusano 2b45e8
kusano 2b45e8
	fpsel	c05, sel_p, c01, c02
kusano 2b45e8
	fpsel	c06, sel_p, c03, c04
kusano 2b45e8
	fpsel	c07, sel_s, c01, c02
kusano 2b45e8
	fpsel	c08, sel_s, c03, c04
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	STFPDUX	c06,   B, INC2
kusano 2b45e8
	STFXDUX	c07,   B, INC2
kusano 2b45e8
	STFXDUX	c08,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L27:
kusano 2b45e8
	andi.	r0,  M,  1
kusano 2b45e8
	beq	.L30
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC2
kusano 2b45e8
	LFDUX	c02,   AO2, INC2
kusano 2b45e8
	LFDUX	c03,   AO3, INC2
kusano 2b45e8
	LFDUX	c04,   AO4, INC2
kusano 2b45e8
kusano 2b45e8
	fsmfp	c01, c02
kusano 2b45e8
	fsmfp	c03, c04
kusano 2b45e8
	
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
	
kusano 2b45e8
.L30:
kusano 2b45e8
	andi.	J,  N,  2
kusano 2b45e8
	ble	.L40
kusano 2b45e8
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	AO2, A,   LDA
kusano 2b45e8
	add	A,   AO2, LDA
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M,  3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L35
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L32:
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFXDUX	c05,   AO2, INC2
kusano 2b45e8
	LFPDUX	c02,   AO1, INC2
kusano 2b45e8
	LFXDUX	c06,   AO2, INC2
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c03,   AO1, INC2
kusano 2b45e8
	fpsel	c09, sel_p, c01, c05
kusano 2b45e8
	LFXDUX	c07,   AO2, INC2
kusano 2b45e8
	fpsel	c10, sel_s, c01, c05
kusano 2b45e8
	LFPDUX	c04,   AO1, INC2
kusano 2b45e8
	fpsel	c11, sel_p, c02, c06
kusano 2b45e8
	LFXDUX	c08,   AO2, INC2
kusano 2b45e8
	fpsel	c12, sel_s, c02, c06
kusano 2b45e8
kusano 2b45e8
	fpsel	c13, sel_p, c03, c07
kusano 2b45e8
	fpsel	c14, sel_s, c03, c07
kusano 2b45e8
	STFPDUX	c09,   B, INC2
kusano 2b45e8
	fpsel	c15, sel_p, c04, c08
kusano 2b45e8
	STFXDUX	c10,   B, INC2
kusano 2b45e8
	fpsel	c16, sel_s, c04, c08
kusano 2b45e8
	STFPDUX	c11,   B, INC2
kusano 2b45e8
	STFXDUX	c12,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c13,   B, INC2
kusano 2b45e8
	STFXDUX	c14,   B, INC2
kusano 2b45e8
	STFPDUX	c15,   B, INC2
kusano 2b45e8
	STFXDUX	c16,   B, INC2
kusano 2b45e8
	bdnz	.L32
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
.L35:
kusano 2b45e8
	andi.	r0,  M,  7
kusano 2b45e8
	ble	.L40
kusano 2b45e8
kusano 2b45e8
	andi.	r0,  M,  4
kusano 2b45e8
	beq	.L36
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFXDUX	c03,   AO2, INC2
kusano 2b45e8
	LFPDUX	c02,   AO1, INC2
kusano 2b45e8
	LFXDUX	c04,   AO2, INC2
kusano 2b45e8
kusano 2b45e8
	fpsel	c05, sel_p, c01, c03
kusano 2b45e8
	fpsel	c06, sel_s, c01, c03
kusano 2b45e8
	fpsel	c07, sel_p, c02, c04
kusano 2b45e8
	fpsel	c08, sel_s, c02, c04
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	STFXDUX	c06,   B, INC2
kusano 2b45e8
	STFPDUX	c07,   B, INC2
kusano 2b45e8
	STFXDUX	c08,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L36:
kusano 2b45e8
	andi.	r0,  M,  2
kusano 2b45e8
	beq	.L37
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFXDUX	c02,   AO2, INC2
kusano 2b45e8
kusano 2b45e8
	fpsel	c03, sel_p, c01, c02
kusano 2b45e8
	fpsel	c04, sel_s, c01, c02
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFXDUX	c04,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L37:
kusano 2b45e8
	andi.	r0,  M,  1
kusano 2b45e8
	beq	.L40
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC2
kusano 2b45e8
	LFDUX	c02,   AO2, INC2
kusano 2b45e8
kusano 2b45e8
	fsmfp	c01, c02
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L40:
kusano 2b45e8
	andi.	J,  N,  1
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M,  3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L45
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L42:
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFPDUX	c02,   AO1, INC2
kusano 2b45e8
	LFPDUX	c03,   AO1, INC2
kusano 2b45e8
	LFPDUX	c04,   AO1, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
	bdnz	.L42
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
.L45:
kusano 2b45e8
	andi.	r0,  M,  7
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	andi.	r0,  M,  4
kusano 2b45e8
	beq	.L46
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFPDUX	c02,   AO1, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L46:
kusano 2b45e8
	andi.	r0,  M,  2
kusano 2b45e8
	beq	.L47
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L47:
kusano 2b45e8
	andi.	r0,  M,  1
kusano 2b45e8
	beq	.L999
kusano 2b45e8
kusano 2b45e8
	LFDX	c01,   AO1, INC2
kusano 2b45e8
	STFDX	c01,   B,  INC2
kusano 2b45e8
	b	.L999
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
.L100:
kusano 2b45e8
	subi	A, A, 1 * SIZE
kusano 2b45e8
	srawi.	J,  N,  3
kusano 2b45e8
	ble	.L120
kusano 2b45e8
	.align 4
kusano 2b45e8
.L111:
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	AO2, A,   LDA
kusano 2b45e8
	add	AO3, AO2, LDA
kusano 2b45e8
	add	AO4, AO3, LDA
kusano 2b45e8
	add	AO5, AO4, LDA
kusano 2b45e8
	add	AO6, AO5, LDA
kusano 2b45e8
	add	AO7, AO6, LDA
kusano 2b45e8
	add	AO8, AO7, LDA
kusano 2b45e8
	add	A,   AO8, LDA
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M,  3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L115
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L112:
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c05,   AO1, INC
kusano 2b45e8
	LFDUX	c09,   AO1, INC
kusano 2b45e8
	LFDUX	c13,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c17,   AO1, INC
kusano 2b45e8
	LFDUX	c21,   AO1, INC
kusano 2b45e8
	LFDUX	c25,   AO1, INC
kusano 2b45e8
	LFDUX	c29,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c05,   AO2, INC
kusano 2b45e8
	LFSDUX	c09,   AO2, INC
kusano 2b45e8
	LFSDUX	c13,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c17,   AO2, INC
kusano 2b45e8
	LFSDUX	c21,   AO2, INC
kusano 2b45e8
	LFSDUX	c25,   AO2, INC
kusano 2b45e8
	LFSDUX	c29,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c02,   AO3, INC
kusano 2b45e8
	LFDUX	c06,   AO3, INC
kusano 2b45e8
	LFDUX	c10,   AO3, INC
kusano 2b45e8
	LFDUX	c14,   AO3, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c18,   AO3, INC
kusano 2b45e8
	LFDUX	c22,   AO3, INC
kusano 2b45e8
	LFDUX	c26,   AO3, INC
kusano 2b45e8
	LFDUX	c30,   AO3, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c02,   AO4, INC
kusano 2b45e8
	LFSDUX	c06,   AO4, INC
kusano 2b45e8
	LFSDUX	c10,   AO4, INC
kusano 2b45e8
	LFSDUX	c14,   AO4, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c18,   AO4, INC
kusano 2b45e8
	LFSDUX	c22,   AO4, INC
kusano 2b45e8
	LFSDUX	c26,   AO4, INC
kusano 2b45e8
	LFSDUX	c30,   AO4, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c03,   AO5, INC
kusano 2b45e8
	LFDUX	c07,   AO5, INC
kusano 2b45e8
	LFDUX	c11,   AO5, INC
kusano 2b45e8
	LFDUX	c15,   AO5, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c19,   AO5, INC
kusano 2b45e8
	LFDUX	c23,   AO5, INC
kusano 2b45e8
	LFDUX	c27,   AO5, INC
kusano 2b45e8
	LFDUX	c31,   AO5, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c03,   AO6, INC
kusano 2b45e8
	LFSDUX	c07,   AO6, INC
kusano 2b45e8
	LFSDUX	c11,   AO6, INC
kusano 2b45e8
	LFSDUX	c15,   AO6, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c19,   AO6, INC
kusano 2b45e8
	LFSDUX	c23,   AO6, INC
kusano 2b45e8
	LFSDUX	c27,   AO6, INC
kusano 2b45e8
	LFSDUX	c31,   AO6, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c04,   AO7, INC
kusano 2b45e8
	LFDUX	c08,   AO7, INC
kusano 2b45e8
	LFDUX	c12,   AO7, INC
kusano 2b45e8
	LFDUX	c16,   AO7, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c20,   AO7, INC
kusano 2b45e8
	LFDUX	c24,   AO7, INC
kusano 2b45e8
	LFDUX	c28,   AO7, INC
kusano 2b45e8
	LFDUX	c32,   AO7, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c04,   AO8, INC
kusano 2b45e8
	LFSDUX	c08,   AO8, INC
kusano 2b45e8
	LFSDUX	c12,   AO8, INC
kusano 2b45e8
	LFSDUX	c16,   AO8, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c20,   AO8, INC
kusano 2b45e8
	LFSDUX	c24,   AO8, INC
kusano 2b45e8
	LFSDUX	c28,   AO8, INC
kusano 2b45e8
	LFSDUX	c32,   AO8, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	STFPDUX	c06,   B, INC2
kusano 2b45e8
	STFPDUX	c07,   B, INC2
kusano 2b45e8
	STFPDUX	c08,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c09,   B, INC2
kusano 2b45e8
	STFPDUX	c10,   B, INC2
kusano 2b45e8
	STFPDUX	c11,   B, INC2
kusano 2b45e8
	STFPDUX	c12,   B, INC2
kusano 2b45e8
	STFPDUX	c13,   B, INC2
kusano 2b45e8
	STFPDUX	c14,   B, INC2
kusano 2b45e8
	STFPDUX	c15,   B, INC2
kusano 2b45e8
	STFPDUX	c16,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c17,   B, INC2
kusano 2b45e8
	STFPDUX	c18,   B, INC2
kusano 2b45e8
	STFPDUX	c19,   B, INC2
kusano 2b45e8
	STFPDUX	c20,   B, INC2
kusano 2b45e8
	STFPDUX	c21,   B, INC2
kusano 2b45e8
	STFPDUX	c22,   B, INC2
kusano 2b45e8
	STFPDUX	c23,   B, INC2
kusano 2b45e8
	STFPDUX	c24,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c25,   B, INC2
kusano 2b45e8
	STFPDUX	c26,   B, INC2
kusano 2b45e8
	STFPDUX	c27,   B, INC2
kusano 2b45e8
	STFPDUX	c28,   B, INC2
kusano 2b45e8
	STFPDUX	c29,   B, INC2
kusano 2b45e8
	STFPDUX	c30,   B, INC2
kusano 2b45e8
	STFPDUX	c31,   B, INC2
kusano 2b45e8
	STFPDUX	c32,   B, INC2
kusano 2b45e8
	bdnz	.L112
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
.L115:
kusano 2b45e8
	andi.	r0,  M,  7
kusano 2b45e8
	ble	.L119
kusano 2b45e8
kusano 2b45e8
	andi.	r0,  M,  4
kusano 2b45e8
	beq	.L116
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c05,   AO1, INC
kusano 2b45e8
	LFDUX	c09,   AO1, INC
kusano 2b45e8
	LFDUX	c13,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c05,   AO2, INC
kusano 2b45e8
	LFSDUX	c09,   AO2, INC
kusano 2b45e8
	LFSDUX	c13,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c02,   AO3, INC
kusano 2b45e8
	LFDUX	c06,   AO3, INC
kusano 2b45e8
	LFDUX	c10,   AO3, INC
kusano 2b45e8
	LFDUX	c14,   AO3, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c02,   AO4, INC
kusano 2b45e8
	LFSDUX	c06,   AO4, INC
kusano 2b45e8
	LFSDUX	c10,   AO4, INC
kusano 2b45e8
	LFSDUX	c14,   AO4, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c03,   AO5, INC
kusano 2b45e8
	LFDUX	c07,   AO5, INC
kusano 2b45e8
	LFDUX	c11,   AO5, INC
kusano 2b45e8
	LFDUX	c15,   AO5, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c03,   AO6, INC
kusano 2b45e8
	LFSDUX	c07,   AO6, INC
kusano 2b45e8
	LFSDUX	c11,   AO6, INC
kusano 2b45e8
	LFSDUX	c15,   AO6, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c04,   AO7, INC
kusano 2b45e8
	LFDUX	c08,   AO7, INC
kusano 2b45e8
	LFDUX	c12,   AO7, INC
kusano 2b45e8
	LFDUX	c16,   AO7, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c04,   AO8, INC
kusano 2b45e8
	LFSDUX	c08,   AO8, INC
kusano 2b45e8
	LFSDUX	c12,   AO8, INC
kusano 2b45e8
	LFSDUX	c16,   AO8, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	STFPDUX	c06,   B, INC2
kusano 2b45e8
	STFPDUX	c07,   B, INC2
kusano 2b45e8
	STFPDUX	c08,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c09,   B, INC2
kusano 2b45e8
	STFPDUX	c10,   B, INC2
kusano 2b45e8
	STFPDUX	c11,   B, INC2
kusano 2b45e8
	STFPDUX	c12,   B, INC2
kusano 2b45e8
	STFPDUX	c13,   B, INC2
kusano 2b45e8
	STFPDUX	c14,   B, INC2
kusano 2b45e8
	STFPDUX	c15,   B, INC2
kusano 2b45e8
	STFPDUX	c16,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L116:
kusano 2b45e8
	andi.	r0,  M,  2
kusano 2b45e8
	beq	.L117
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c05,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO3, INC
kusano 2b45e8
	LFDUX	c06,   AO3, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c05,   AO2, INC
kusano 2b45e8
	LFSDUX	c02,   AO4, INC
kusano 2b45e8
	LFSDUX	c06,   AO4, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c03,   AO5, INC
kusano 2b45e8
	LFDUX	c07,   AO5, INC
kusano 2b45e8
	LFDUX	c04,   AO7, INC
kusano 2b45e8
	LFDUX	c08,   AO7, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c03,   AO6, INC
kusano 2b45e8
	LFSDUX	c07,   AO6, INC
kusano 2b45e8
	LFSDUX	c04,   AO8, INC
kusano 2b45e8
	LFSDUX	c08,   AO8, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	STFPDUX	c06,   B, INC2
kusano 2b45e8
	STFPDUX	c07,   B, INC2
kusano 2b45e8
	STFPDUX	c08,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L117:
kusano 2b45e8
	andi.	r0,  M,  1
kusano 2b45e8
	beq	.L119
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO3, INC
kusano 2b45e8
	LFDUX	c03,   AO5, INC
kusano 2b45e8
	LFDUX	c04,   AO7, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c02,   AO4, INC
kusano 2b45e8
	LFSDUX	c03,   AO6, INC
kusano 2b45e8
	LFSDUX	c04,   AO8, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L119:
kusano 2b45e8
	addic.	J, J, -1
kusano 2b45e8
	bgt	.L111
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L120:
kusano 2b45e8
	andi.	J,  N,  4
kusano 2b45e8
	ble	.L130
kusano 2b45e8
	.align 4
kusano 2b45e8
.L121:
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	AO2, A,   LDA
kusano 2b45e8
	add	AO3, AO2, LDA
kusano 2b45e8
	add	AO4, AO3, LDA
kusano 2b45e8
	add	A,   AO4, LDA
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M,  3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L125
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L122:
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
	LFDUX	c03,   AO1, INC
kusano 2b45e8
	LFDUX	c04,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c09,   AO1, INC
kusano 2b45e8
	LFDUX	c10,   AO1, INC
kusano 2b45e8
	LFDUX	c11,   AO1, INC
kusano 2b45e8
	LFDUX	c12,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c02,   AO2, INC
kusano 2b45e8
	LFSDUX	c03,   AO2, INC
kusano 2b45e8
	LFSDUX	c04,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c09,   AO2, INC
kusano 2b45e8
	LFSDUX	c10,   AO2, INC
kusano 2b45e8
	LFSDUX	c11,   AO2, INC
kusano 2b45e8
	LFSDUX	c12,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c05,   AO3, INC
kusano 2b45e8
	LFDUX	c06,   AO3, INC
kusano 2b45e8
	LFDUX	c07,   AO3, INC
kusano 2b45e8
	LFDUX	c08,   AO3, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c13,   AO3, INC
kusano 2b45e8
	LFDUX	c14,   AO3, INC
kusano 2b45e8
	LFDUX	c15,   AO3, INC
kusano 2b45e8
	LFDUX	c16,   AO3, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c05,   AO4, INC
kusano 2b45e8
	LFSDUX	c06,   AO4, INC
kusano 2b45e8
	LFSDUX	c07,   AO4, INC
kusano 2b45e8
	LFSDUX	c08,   AO4, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c13,   AO4, INC
kusano 2b45e8
	LFSDUX	c14,   AO4, INC
kusano 2b45e8
	LFSDUX	c15,   AO4, INC
kusano 2b45e8
	LFSDUX	c16,   AO4, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c06,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c07,   B, INC2
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
	STFPDUX	c08,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c09,   B, INC2
kusano 2b45e8
	STFPDUX	c13,   B, INC2
kusano 2b45e8
	STFPDUX	c10,   B, INC2
kusano 2b45e8
	STFPDUX	c14,   B, INC2
kusano 2b45e8
	STFPDUX	c11,   B, INC2
kusano 2b45e8
	STFPDUX	c15,   B, INC2
kusano 2b45e8
	STFPDUX	c12,   B, INC2
kusano 2b45e8
	STFPDUX	c16,   B, INC2
kusano 2b45e8
	bdnz	.L122
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
.L125:
kusano 2b45e8
	andi.	r0,  M,  7
kusano 2b45e8
	ble	.L130
kusano 2b45e8
kusano 2b45e8
	andi.	r0,  M,  4
kusano 2b45e8
	beq	.L126
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
	LFDUX	c03,   AO1, INC
kusano 2b45e8
	LFDUX	c04,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c02,   AO2, INC
kusano 2b45e8
	LFSDUX	c03,   AO2, INC
kusano 2b45e8
	LFSDUX	c04,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c05,   AO3, INC
kusano 2b45e8
	LFDUX	c06,   AO3, INC
kusano 2b45e8
	LFDUX	c07,   AO3, INC
kusano 2b45e8
	LFDUX	c08,   AO3, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c05,   AO4, INC
kusano 2b45e8
	LFSDUX	c06,   AO4, INC
kusano 2b45e8
	LFSDUX	c07,   AO4, INC
kusano 2b45e8
	LFSDUX	c08,   AO4, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c06,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c07,   B, INC2
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
	STFPDUX	c08,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L126:
kusano 2b45e8
	andi.	r0,  M,  2
kusano 2b45e8
	beq	.L127
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c02,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c05,   AO3, INC
kusano 2b45e8
	LFDUX	c06,   AO3, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c05,   AO4, INC
kusano 2b45e8
	LFSDUX	c06,   AO4, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c06,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L127:
kusano 2b45e8
	andi.	r0,  M,  1
kusano 2b45e8
	beq	.L130
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c05,   AO3, INC
kusano 2b45e8
kusano 2b45e8
	nop
kusano 2b45e8
	nop
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c05,   AO4, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
	
kusano 2b45e8
.L130:
kusano 2b45e8
	andi.	J,  N,  2
kusano 2b45e8
	ble	.L140
kusano 2b45e8
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	AO2, A,   LDA
kusano 2b45e8
	add	A,   AO2, LDA
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M,  3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L135
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L132:
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
	LFDUX	c03,   AO1, INC
kusano 2b45e8
	LFDUX	c04,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c09,   AO1, INC
kusano 2b45e8
	LFDUX	c10,   AO1, INC
kusano 2b45e8
	LFDUX	c11,   AO1, INC
kusano 2b45e8
	LFDUX	c12,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c02,   AO2, INC
kusano 2b45e8
	LFSDUX	c03,   AO2, INC
kusano 2b45e8
	LFSDUX	c04,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c09,   AO2, INC
kusano 2b45e8
	LFSDUX	c10,   AO2, INC
kusano 2b45e8
	LFSDUX	c11,   AO2, INC
kusano 2b45e8
	LFSDUX	c12,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c09,   B, INC2
kusano 2b45e8
	STFPDUX	c10,   B, INC2
kusano 2b45e8
	STFPDUX	c11,   B, INC2
kusano 2b45e8
	STFPDUX	c12,   B, INC2
kusano 2b45e8
	bdnz	.L132
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
.L135:
kusano 2b45e8
	andi.	r0,  M,  7
kusano 2b45e8
	ble	.L140
kusano 2b45e8
kusano 2b45e8
	andi.	r0,  M,  4
kusano 2b45e8
	beq	.L136
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
	LFDUX	c03,   AO1, INC
kusano 2b45e8
	LFDUX	c04,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c02,   AO2, INC
kusano 2b45e8
	LFSDUX	c03,   AO2, INC
kusano 2b45e8
	LFSDUX	c04,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c04,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L136:
kusano 2b45e8
	andi.	r0,  M,  2
kusano 2b45e8
	beq	.L137
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFSDUX	c01,   AO2, INC
kusano 2b45e8
	LFSDUX	c02,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c02,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L137:
kusano 2b45e8
	andi.	r0,  M,  1
kusano 2b45e8
	beq	.L140
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	fsmfp	c01, c02
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L140:
kusano 2b45e8
	andi.	J,  N,  1
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  M,  3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	.L145
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L142:
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
	LFDUX	c03,   AO1, INC
kusano 2b45e8
	LFDUX	c04,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	LFDUX	c05,   AO1, INC
kusano 2b45e8
	LFDUX	c06,   AO1, INC
kusano 2b45e8
	LFDUX	c07,   AO1, INC
kusano 2b45e8
	LFDUX	c08,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	fsmfp	c01, c02
kusano 2b45e8
	fsmfp	c03, c04
kusano 2b45e8
	fsmfp	c05, c06
kusano 2b45e8
	fsmfp	c07, c08
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	STFPDUX	c05,   B, INC2
kusano 2b45e8
	STFPDUX	c07,   B, INC2
kusano 2b45e8
	bdnz	.L142
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
.L145:
kusano 2b45e8
	andi.	r0,  M,  7
kusano 2b45e8
	ble	.L999
kusano 2b45e8
kusano 2b45e8
	andi.	r0,  M,  4
kusano 2b45e8
	beq	.L146
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
	LFDUX	c03,   AO1, INC
kusano 2b45e8
	LFDUX	c04,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	fsmfp	c01, c02
kusano 2b45e8
	fsmfp	c03, c04
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	STFPDUX	c03,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L146:
kusano 2b45e8
	andi.	r0,  M,  2
kusano 2b45e8
	beq	.L147
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	fsmfp	c01, c02
kusano 2b45e8
	STFPDUX	c01,   B, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L147:
kusano 2b45e8
	andi.	r0,  M,  1
kusano 2b45e8
	beq	.L999
kusano 2b45e8
kusano 2b45e8
	LFDX	c01,   AO1, INC
kusano 2b45e8
	STFDX	c01,   B,  INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
.L999:
kusano 2b45e8
	addi	SP, SP, 4
kusano 2b45e8
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
	EPILOGUE