|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
/* Copyright 2009, 2010 The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/* All rights reserved. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* Redistribution and use in source and binary forms, with or */
|
|
kusano |
2b45e8 |
/* without modification, are permitted provided that the following */
|
|
kusano |
2b45e8 |
/* conditions are met: */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 1. Redistributions of source code must retain the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 2. Redistributions in binary form must reproduce the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer in the documentation and/or other materials */
|
|
kusano |
2b45e8 |
/* provided with the distribution. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
|
|
kusano |
2b45e8 |
/* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
|
|
kusano |
2b45e8 |
/* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
|
|
kusano |
2b45e8 |
/* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
|
|
kusano |
2b45e8 |
/* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
|
|
kusano |
2b45e8 |
/* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
|
|
kusano |
2b45e8 |
/* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
|
|
kusano |
2b45e8 |
/* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
|
|
kusano |
2b45e8 |
/* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
|
|
kusano |
2b45e8 |
/* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
|
|
kusano |
2b45e8 |
/* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
|
|
kusano |
2b45e8 |
/* POSSIBILITY OF SUCH DAMAGE. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* The views and conclusions contained in the software and */
|
|
kusano |
2b45e8 |
/* documentation are those of the authors and should not be */
|
|
kusano |
2b45e8 |
/* interpreted as representing official policies, either expressed */
|
|
kusano |
2b45e8 |
/* or implied, of The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ASSEMBLER
|
|
kusano |
2b45e8 |
#include "common.h"
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define N $4
|
|
kusano |
2b45e8 |
#define X $5
|
|
kusano |
2b45e8 |
#define INCX $6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define I $3
|
|
kusano |
2b45e8 |
#define TEMP $7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define a1 $f4
|
|
kusano |
2b45e8 |
#define a2 $f5
|
|
kusano |
2b45e8 |
#define a3 $f6
|
|
kusano |
2b45e8 |
#define a4 $f7
|
|
kusano |
2b45e8 |
#define a5 $f8
|
|
kusano |
2b45e8 |
#define a6 $f9
|
|
kusano |
2b45e8 |
#define a7 $f10
|
|
kusano |
2b45e8 |
#define a8 $f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define t1 $f12
|
|
kusano |
2b45e8 |
#define t2 $f13
|
|
kusano |
2b45e8 |
#define t3 $f14
|
|
kusano |
2b45e8 |
#define t4 $f15
|
|
kusano |
2b45e8 |
#define t5 $f16
|
|
kusano |
2b45e8 |
#define t6 $f17
|
|
kusano |
2b45e8 |
#define t7 $f18
|
|
kusano |
2b45e8 |
#define t8 $f19
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define s1 $f0
|
|
kusano |
2b45e8 |
#define s2 $f1
|
|
kusano |
2b45e8 |
#define s3 $f2
|
|
kusano |
2b45e8 |
#define s4 $f3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define x1 $2
|
|
kusano |
2b45e8 |
#define x2 $8
|
|
kusano |
2b45e8 |
#define x3 $9
|
|
kusano |
2b45e8 |
#define x4 $10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROLOGUE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef F_INTERFACE
|
|
kusano |
2b45e8 |
LDINT N, 0(N)
|
|
kusano |
2b45e8 |
LDINT INCX, 0(INCX)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
blez N, .L999
|
|
kusano |
2b45e8 |
li x1, 0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
blez INCX, .L999
|
|
kusano |
2b45e8 |
dsll INCX, INCX, ZBASE_SHIFT
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LD a1, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
LD a2, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FABS t1, a1
|
|
kusano |
2b45e8 |
FABS t2, a2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD s1, t1, t2
|
|
kusano |
2b45e8 |
ADD s2, t1, t2
|
|
kusano |
2b45e8 |
ADD s3, t1, t2
|
|
kusano |
2b45e8 |
ADD s4, t1, t2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
daddiu N, N, -1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
blez N, .L999
|
|
kusano |
2b45e8 |
li x1, 1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
daddu X, X, INCX
|
|
kusano |
2b45e8 |
li x2, 1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
dsra I, N, 2
|
|
kusano |
2b45e8 |
li x3, 1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
li TEMP, 2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
blez I, .L15
|
|
kusano |
2b45e8 |
li x4, 1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LD a1, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
LD a2, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
daddu X, X, INCX
|
|
kusano |
2b45e8 |
LD a3, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
LD a4, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
daddu X, X, INCX
|
|
kusano |
2b45e8 |
LD a5, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
LD a6, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
daddu X, X, INCX
|
|
kusano |
2b45e8 |
LD a7, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
LD a8, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
daddiu I, I, -1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
blez I, .L13
|
|
kusano |
2b45e8 |
daddu X, X, INCX
|
|
kusano |
2b45e8 |
.align 3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L12:
|
|
kusano |
2b45e8 |
FABS t1, a1
|
|
kusano |
2b45e8 |
LD a1, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
FABS t2, a2
|
|
kusano |
2b45e8 |
LD a2, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FABS t3, a3
|
|
kusano |
2b45e8 |
daddu X, X, INCX
|
|
kusano |
2b45e8 |
FABS t4, a4
|
|
kusano |
2b45e8 |
NOP
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FABS t5, a5
|
|
kusano |
2b45e8 |
LD a3, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
FABS t6, a6
|
|
kusano |
2b45e8 |
LD a4, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
FABS t7, a7
|
|
kusano |
2b45e8 |
daddu X, X, INCX
|
|
kusano |
2b45e8 |
FABS t8, a8
|
|
kusano |
2b45e8 |
NOP
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD t1, t1, t2
|
|
kusano |
2b45e8 |
LD a5, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
ADD t3, t3, t4
|
|
kusano |
2b45e8 |
LD a6, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
ADD t5, t5, t6
|
|
kusano |
2b45e8 |
daddu X, X, INCX
|
|
kusano |
2b45e8 |
ADD t7, t7, t8
|
|
kusano |
2b45e8 |
NOP
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
CMPLT $fcc0, t1, s1
|
|
kusano |
2b45e8 |
LD a7, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
CMPLT $fcc1, t3, s2
|
|
kusano |
2b45e8 |
LD a8, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
CMPLT $fcc2, t5, s3
|
|
kusano |
2b45e8 |
daddu X, X, INCX
|
|
kusano |
2b45e8 |
CMPLT $fcc3, t7, s4
|
|
kusano |
2b45e8 |
daddiu I, I, -1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
CMOVT s1, t1, $fcc0
|
|
kusano |
2b45e8 |
movt x1, TEMP, $fcc0
|
|
kusano |
2b45e8 |
CMOVT s2, t3, $fcc1
|
|
kusano |
2b45e8 |
movt x2, TEMP, $fcc1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
CMOVT s3, t5, $fcc2
|
|
kusano |
2b45e8 |
movt x3, TEMP, $fcc2
|
|
kusano |
2b45e8 |
CMOVT s4, t7, $fcc3
|
|
kusano |
2b45e8 |
movt x4, TEMP, $fcc3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
bgtz I, .L12
|
|
kusano |
2b45e8 |
daddiu TEMP, TEMP, 4
|
|
kusano |
2b45e8 |
.align 3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L13:
|
|
kusano |
2b45e8 |
FABS t1, a1
|
|
kusano |
2b45e8 |
FABS t2, a2
|
|
kusano |
2b45e8 |
FABS t3, a3
|
|
kusano |
2b45e8 |
FABS t4, a4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FABS t5, a5
|
|
kusano |
2b45e8 |
FABS t6, a6
|
|
kusano |
2b45e8 |
FABS t7, a7
|
|
kusano |
2b45e8 |
FABS t8, a8
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD t1, t1, t2
|
|
kusano |
2b45e8 |
ADD t3, t3, t4
|
|
kusano |
2b45e8 |
ADD t5, t5, t6
|
|
kusano |
2b45e8 |
ADD t7, t7, t8
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
CMPLT $fcc0, t1, s1
|
|
kusano |
2b45e8 |
CMPLT $fcc1, t3, s2
|
|
kusano |
2b45e8 |
CMPLT $fcc2, t5, s3
|
|
kusano |
2b45e8 |
CMPLT $fcc3, t7, s4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
CMOVT s1, t1, $fcc0
|
|
kusano |
2b45e8 |
movt x1, TEMP, $fcc0
|
|
kusano |
2b45e8 |
CMOVT s2, t3, $fcc1
|
|
kusano |
2b45e8 |
movt x2, TEMP, $fcc1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
CMOVT s3, t5, $fcc2
|
|
kusano |
2b45e8 |
movt x3, TEMP, $fcc2
|
|
kusano |
2b45e8 |
CMOVT s4, t7, $fcc3
|
|
kusano |
2b45e8 |
movt x4, TEMP, $fcc3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
daddiu TEMP, TEMP, 4
|
|
kusano |
2b45e8 |
daddiu x2, x2, 1
|
|
kusano |
2b45e8 |
daddiu x3, x3, 2
|
|
kusano |
2b45e8 |
daddiu x4, x4, 3
|
|
kusano |
2b45e8 |
.align 3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L15:
|
|
kusano |
2b45e8 |
andi I, N, 3
|
|
kusano |
2b45e8 |
blez I, .L998
|
|
kusano |
2b45e8 |
NOP
|
|
kusano |
2b45e8 |
.align 3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L16:
|
|
kusano |
2b45e8 |
LD a1, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
LD a2, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
daddu X, X, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FABS t1, a1
|
|
kusano |
2b45e8 |
FABS t2, a2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD t1, t1, t2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
daddiu I, I, -1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
CMPLT $fcc0, t1, s1
|
|
kusano |
2b45e8 |
NOP
|
|
kusano |
2b45e8 |
CMOVT s1, t1, $fcc0
|
|
kusano |
2b45e8 |
movt x1, TEMP, $fcc0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
bgtz I, .L16
|
|
kusano |
2b45e8 |
daddiu TEMP, TEMP, 1
|
|
kusano |
2b45e8 |
.align 3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L998:
|
|
kusano |
2b45e8 |
CMPLT $fcc0, s2, s1
|
|
kusano |
2b45e8 |
CMPLT $fcc1, s4, s3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
CMOVT s1, s2, $fcc0
|
|
kusano |
2b45e8 |
movt x1, x2, $fcc0
|
|
kusano |
2b45e8 |
CMOVT s3, s4, $fcc1
|
|
kusano |
2b45e8 |
movt x3, x4, $fcc1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
CMPLT $fcc0, s3, s1
|
|
kusano |
2b45e8 |
CMOVT s1, s3, $fcc0
|
|
kusano |
2b45e8 |
movt x1, x3, $fcc0
|
|
kusano |
2b45e8 |
.align 3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L999:
|
|
kusano |
2b45e8 |
j $31
|
|
kusano |
2b45e8 |
NOP
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
EPILOGUE
|