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
kusano 2b45e8
#define J	r10
kusano 2b45e8
#define B1	r11
kusano 2b45e8
kusano 2b45e8
#define B2	r28
kusano 2b45e8
#define M4	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
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	PROFCODE
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
	slwi	LDA, LDA, ZBASE_SHIFT
kusano 2b45e8
	slwi	M4, M, 1 + ZBASE_SHIFT
kusano 2b45e8
kusano 2b45e8
	li	r9, -2
kusano 2b45e8
kusano 2b45e8
	and	B2, N, r9
kusano 2b45e8
kusano 2b45e8
	mullw	B2, B2, M
kusano 2b45e8
kusano 2b45e8
	slwi	B2, B2, ZBASE_SHIFT
kusano 2b45e8
kusano 2b45e8
	add	B2, B2, B
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, M, 0
kusano 2b45e8
	ble-	LL(99)
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	ble-	LL(99)
kusano 2b45e8
kusano 2b45e8
	subi	B2, B2, 2 * SIZE
kusano 2b45e8
	subi	M4, M4, 6 * SIZE
kusano 2b45e8
kusano 2b45e8
	li	INC,  1 * SIZE
kusano 2b45e8
	li	INC2, 2 * SIZE
kusano 2b45e8
kusano 2b45e8
	andi.	r0, A,   2 * SIZE - 1
kusano 2b45e8
	bne	LL(100)
kusano 2b45e8
kusano 2b45e8
	subi	A, A, 2 * SIZE
kusano 2b45e8
	srawi.	J,  M,  1
kusano 2b45e8
	ble	LL(20)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(10):
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	AO2, A,   LDA
kusano 2b45e8
	add	A,   AO2, LDA
kusano 2b45e8
kusano 2b45e8
	sub	B1, B, M4
kusano 2b45e8
	addi	B, B, 8 * SIZE
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  N,  1
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(15)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(12):
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFPDUX	c02,   AO1, INC2
kusano 2b45e8
	LFPDUX	c03,   AO2, INC2
kusano 2b45e8
	LFPDUX	c04,   AO2, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B1, M4
kusano 2b45e8
	STFPDUX	c02,   B1, INC2
kusano 2b45e8
	STFPDUX	c03,   B1, INC2
kusano 2b45e8
	STFPDUX	c04,   B1, INC2
kusano 2b45e8
	bdnz	LL(12)
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
LL(15):
kusano 2b45e8
	andi.	r0,  N,  1
kusano 2b45e8
	ble	LL(19)
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFPDUX	c02,   AO2, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B2, INC2
kusano 2b45e8
	STFPDUX	c02,   B2, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(19):
kusano 2b45e8
	addic.	J, J, -1
kusano 2b45e8
	bgt	LL(10)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(20):
kusano 2b45e8
	andi.	J,  M,  1
kusano 2b45e8
	addi	M4, M4, 4 * SIZE
kusano 2b45e8
	ble	LL(99)
kusano 2b45e8
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	sub	B1, B, M4
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  N,  1
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(23)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(22):
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
	LFPDUX	c02,   AO1, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B1, M4
kusano 2b45e8
	STFPDUX	c02,   B1, INC2
kusano 2b45e8
	bdnz	LL(22)
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
LL(23):
kusano 2b45e8
	andi.	r0,  N,  1
kusano 2b45e8
	ble	LL(99)
kusano 2b45e8
kusano 2b45e8
	LFPDUX	c01,   AO1, INC2
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B2, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(99):
kusano 2b45e8
	addi	SP, SP, -4
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
	addi	SP, SP, 4
kusano 2b45e8
	blr
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(100):
kusano 2b45e8
	subi	A, A, SIZE
kusano 2b45e8
	srawi.	J,  M,  1
kusano 2b45e8
	ble	LL(120)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(110):
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	add	AO2, A,   LDA
kusano 2b45e8
	add	A,   AO2, LDA
kusano 2b45e8
kusano 2b45e8
	sub	B1, B, M4
kusano 2b45e8
	addi	B, B, 8 * SIZE
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  N,  1
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(115)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(112):
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,   AO2, INC
kusano 2b45e8
	fsmfp	c01, c02
kusano 2b45e8
	LFDUX	c06,   AO2, INC
kusano 2b45e8
	fsmfp	c03, c04
kusano 2b45e8
	LFDUX	c07,   AO2, INC
kusano 2b45e8
	fsmfp	c05, c06
kusano 2b45e8
	LFDUX	c08,   AO2, INC
kusano 2b45e8
	fsmfp	c07, c08
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B1, M4
kusano 2b45e8
	STFPDUX	c03,   B1, INC2
kusano 2b45e8
	STFPDUX	c05,   B1, INC2
kusano 2b45e8
	STFPDUX	c07,   B1, INC2
kusano 2b45e8
	bdnz	LL(112)
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
LL(115):
kusano 2b45e8
	andi.	r0,  N,  1
kusano 2b45e8
	ble	LL(119)
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
	LFDUX	c03,   AO2, INC
kusano 2b45e8
	LFDUX	c04,   AO2, INC
kusano 2b45e8
kusano 2b45e8
	fsmfp	c01, c02
kusano 2b45e8
	fsmfp	c03, c04
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B2, INC2
kusano 2b45e8
	STFPDUX	c03,   B2, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(119):
kusano 2b45e8
	addic.	J, J, -1
kusano 2b45e8
	bgt	LL(110)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(120):
kusano 2b45e8
	andi.	J,  M,  1
kusano 2b45e8
	addi	M4, M4, 4 * SIZE
kusano 2b45e8
	ble	LL(999)
kusano 2b45e8
kusano 2b45e8
	mr	AO1, A
kusano 2b45e8
	sub	B1, B, M4
kusano 2b45e8
kusano 2b45e8
	srawi.	r0,  N,  1
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	ble	LL(123)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(122):
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,   B1, M4
kusano 2b45e8
	STFPDUX	c03,   B1, INC2
kusano 2b45e8
	bdnz	LL(122)
kusano 2b45e8
	.align 4
kusano 2b45e8
	
kusano 2b45e8
LL(123):
kusano 2b45e8
	andi.	r0,  N,  1
kusano 2b45e8
	ble	LL(999)
kusano 2b45e8
kusano 2b45e8
	LFDUX	c01,   AO1, INC
kusano 2b45e8
	LFDUX	c02,   AO1, INC
kusano 2b45e8
kusano 2b45e8
	fsmfp	c01, c02
kusano 2b45e8
kusano 2b45e8
	STFPDUX	c01,   B2, INC2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(999):
kusano 2b45e8
	addi	SP, SP, -4
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
	addi	SP, SP, 4
kusano 2b45e8
	blr
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
	EPILOGUE