|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
/* Copyright 2009, 2010 The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/* All rights reserved. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* Redistribution and use in source and binary forms, with or */
|
|
kusano |
2b45e8 |
/* without modification, are permitted provided that the following */
|
|
kusano |
2b45e8 |
/* conditions are met: */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 1. Redistributions of source code must retain the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 2. Redistributions in binary form must reproduce the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer in the documentation and/or other materials */
|
|
kusano |
2b45e8 |
/* provided with the distribution. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
|
|
kusano |
2b45e8 |
/* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
|
|
kusano |
2b45e8 |
/* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
|
|
kusano |
2b45e8 |
/* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
|
|
kusano |
2b45e8 |
/* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
|
|
kusano |
2b45e8 |
/* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
|
|
kusano |
2b45e8 |
/* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
|
|
kusano |
2b45e8 |
/* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
|
|
kusano |
2b45e8 |
/* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
|
|
kusano |
2b45e8 |
/* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
|
|
kusano |
2b45e8 |
/* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
|
|
kusano |
2b45e8 |
/* POSSIBILITY OF SUCH DAMAGE. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* The views and conclusions contained in the software and */
|
|
kusano |
2b45e8 |
/* documentation are those of the authors and should not be */
|
|
kusano |
2b45e8 |
/* interpreted as representing official policies, either expressed */
|
|
kusano |
2b45e8 |
/* or implied, of The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ASSEMBLER
|
|
kusano |
2b45e8 |
#include "common.h"
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define N r3
|
|
kusano |
2b45e8 |
#define X r6
|
|
kusano |
2b45e8 |
#define INCX r7
|
|
kusano |
2b45e8 |
#define Y r8
|
|
kusano |
2b45e8 |
#define INCY r9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define YY r4
|
|
kusano |
2b45e8 |
#define INCX2 r5
|
|
kusano |
2b45e8 |
#define INCY2 r10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ALPHA f1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define A1 f0
|
|
kusano |
2b45e8 |
#define A2 f8
|
|
kusano |
2b45e8 |
#define A3 f2
|
|
kusano |
2b45e8 |
#define A4 f3
|
|
kusano |
2b45e8 |
#define A5 f4
|
|
kusano |
2b45e8 |
#define A6 f5
|
|
kusano |
2b45e8 |
#define A7 f6
|
|
kusano |
2b45e8 |
#define A8 f7
|
|
kusano |
2b45e8 |
#define A9 f25
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define B1 f9
|
|
kusano |
2b45e8 |
#define B2 f10
|
|
kusano |
2b45e8 |
#define B3 f11
|
|
kusano |
2b45e8 |
#define B4 f12
|
|
kusano |
2b45e8 |
#define B5 f13
|
|
kusano |
2b45e8 |
#define B6 f14
|
|
kusano |
2b45e8 |
#define B7 f15
|
|
kusano |
2b45e8 |
#define B8 f16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define C1 f17
|
|
kusano |
2b45e8 |
#define C2 f18
|
|
kusano |
2b45e8 |
#define C3 f19
|
|
kusano |
2b45e8 |
#define C4 f20
|
|
kusano |
2b45e8 |
#define C5 f21
|
|
kusano |
2b45e8 |
#define C6 f22
|
|
kusano |
2b45e8 |
#define C7 f23
|
|
kusano |
2b45e8 |
#define C8 f24
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROLOGUE
|
|
kusano |
2b45e8 |
PROFCODE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
li r10, -16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
stfpdux f14, SP, r10
|
|
kusano |
2b45e8 |
stfpdux f15, SP, r10
|
|
kusano |
2b45e8 |
stfpdux f16, SP, r10
|
|
kusano |
2b45e8 |
stfpdux f17, SP, r10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
stfpdux f18, SP, r10
|
|
kusano |
2b45e8 |
stfpdux f19, SP, r10
|
|
kusano |
2b45e8 |
stfpdux f20, SP, r10
|
|
kusano |
2b45e8 |
stfpdux f21, SP, r10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
stfpdux f22, SP, r10
|
|
kusano |
2b45e8 |
stfpdux f23, SP, r10
|
|
kusano |
2b45e8 |
stfpdux f24, SP, r10
|
|
kusano |
2b45e8 |
stfpdux f25, SP, r10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsmfp ALPHA, ALPHA
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
slwi INCX, INCX, BASE_SHIFT
|
|
kusano |
2b45e8 |
slwi INCY, INCY, BASE_SHIFT
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
add INCX2, INCX, INCX
|
|
kusano |
2b45e8 |
add INCY2, INCY, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
cmpwi cr0, N, 0
|
|
kusano |
2b45e8 |
ble LL(999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
cmpwi cr0, INCX, SIZE
|
|
kusano |
2b45e8 |
bne LL(100)
|
|
kusano |
2b45e8 |
cmpwi cr0, INCY, SIZE
|
|
kusano |
2b45e8 |
bne LL(100)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
andi. r0, Y, 2 * SIZE - 1
|
|
kusano |
2b45e8 |
beq LL(05)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD A1, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD B1, 0 * SIZE(Y)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addi X, X, SIZE
|
|
kusano |
2b45e8 |
addi Y, Y, SIZE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
addi N, N, -1
|
|
kusano |
2b45e8 |
STFD C1, -1 * SIZE(Y)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(05):
|
|
kusano |
2b45e8 |
andi. r0, X, 2 * SIZE - 1
|
|
kusano |
2b45e8 |
bne LL(20)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
sub X, X, INCX2
|
|
kusano |
2b45e8 |
sub Y, Y, INCY2
|
|
kusano |
2b45e8 |
mr YY, Y
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
srawi. r0, N, 4
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq- LL(15)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFPDUX A1, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
LFPDUX A2, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B2, Y, INCY2
|
|
kusano |
2b45e8 |
LFPDUX A3, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B3, Y, INCY2
|
|
kusano |
2b45e8 |
LFPDUX A4, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B4, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFPDUX A5, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B5, Y, INCY2
|
|
kusano |
2b45e8 |
LFPDUX A6, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B6, Y, INCY2
|
|
kusano |
2b45e8 |
LFPDUX A7, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B7, Y, INCY2
|
|
kusano |
2b45e8 |
LFPDUX A8, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B8, Y, INCY2
|
|
kusano |
2b45e8 |
bdz LL(13)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(12):
|
|
kusano |
2b45e8 |
fpmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
LFPDUX A1, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
fpmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
LFPDUX A2, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B2, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C3, ALPHA, A3, B3
|
|
kusano |
2b45e8 |
LFPDUX A3, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B3, Y, INCY2
|
|
kusano |
2b45e8 |
fpmadd C4, ALPHA, A4, B4
|
|
kusano |
2b45e8 |
LFPDUX A4, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B4, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C5, ALPHA, A5, B5
|
|
kusano |
2b45e8 |
LFPDUX A5, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B5, Y, INCY2
|
|
kusano |
2b45e8 |
fpmadd C6, ALPHA, A6, B6
|
|
kusano |
2b45e8 |
LFPDUX A6, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B6, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C7, ALPHA, A7, B7
|
|
kusano |
2b45e8 |
LFPDUX A7, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B7, Y, INCY2
|
|
kusano |
2b45e8 |
fpmadd C8, ALPHA, A8, B8
|
|
kusano |
2b45e8 |
LFPDUX A8, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B8, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C2, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C3, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C4, YY, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C5, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C6, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C7, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C8, YY, INCY2
|
|
kusano |
2b45e8 |
bdnz LL(12)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(13):
|
|
kusano |
2b45e8 |
fpmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fpmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
fpmadd C3, ALPHA, A3, B3
|
|
kusano |
2b45e8 |
fpmadd C4, ALPHA, A4, B4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C5, ALPHA, A5, B5
|
|
kusano |
2b45e8 |
fpmadd C6, ALPHA, A6, B6
|
|
kusano |
2b45e8 |
STFPDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
fpmadd C7, ALPHA, A7, B7
|
|
kusano |
2b45e8 |
STFPDUX C2, YY, INCY2
|
|
kusano |
2b45e8 |
fpmadd C8, ALPHA, A8, B8
|
|
kusano |
2b45e8 |
STFPDUX C3, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C4, YY, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C5, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C6, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C7, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C8, YY, INCY2
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(15):
|
|
kusano |
2b45e8 |
andi. r0, N, 15
|
|
kusano |
2b45e8 |
beq LL(999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
andi. r0, N, 8
|
|
kusano |
2b45e8 |
beq LL(16)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFPDUX A1, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
LFPDUX A2, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B2, Y, INCY2
|
|
kusano |
2b45e8 |
LFPDUX A3, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B3, Y, INCY2
|
|
kusano |
2b45e8 |
LFPDUX A4, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B4, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fpmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
fpmadd C3, ALPHA, A3, B3
|
|
kusano |
2b45e8 |
fpmadd C4, ALPHA, A4, B4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C2, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C3, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C4, YY, INCY2
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(16):
|
|
kusano |
2b45e8 |
andi. r0, N, 4
|
|
kusano |
2b45e8 |
beq LL(17)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFPDUX A1, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
LFPDUX A2, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B2, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fpmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C2, YY, INCY2
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(17):
|
|
kusano |
2b45e8 |
andi. r0, N, 2
|
|
kusano |
2b45e8 |
beq LL(18)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFPDUX A1, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(18):
|
|
kusano |
2b45e8 |
andi. r0, N, 1
|
|
kusano |
2b45e8 |
beq LL(999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX A1, X, INCX2
|
|
kusano |
2b45e8 |
LFDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
STFDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
b LL(999)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
/* X is unaliged */
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(20):
|
|
kusano |
2b45e8 |
LFD A1, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
addi X, X, SIZE
|
|
kusano |
2b45e8 |
sub X, X, INCX2
|
|
kusano |
2b45e8 |
sub Y, Y, INCY2
|
|
kusano |
2b45e8 |
mr YY, Y
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
srawi. r0, N, 4
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq- LL(25)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFXDUX A2, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
LFXDUX A3, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B2, Y, INCY2
|
|
kusano |
2b45e8 |
LFXDUX A4, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B3, Y, INCY2
|
|
kusano |
2b45e8 |
LFXDUX A5, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B4, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFXDUX A6, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B5, Y, INCY2
|
|
kusano |
2b45e8 |
LFXDUX A7, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B6, Y, INCY2
|
|
kusano |
2b45e8 |
fsmr A1, A2
|
|
kusano |
2b45e8 |
LFXDUX A8, X, INCX2
|
|
kusano |
2b45e8 |
fsmr A2, A3
|
|
kusano |
2b45e8 |
LFPDUX B7, Y, INCY2
|
|
kusano |
2b45e8 |
fsmr A3, A4
|
|
kusano |
2b45e8 |
LFXDUX A9, X, INCX2
|
|
kusano |
2b45e8 |
fsmr A4, A5
|
|
kusano |
2b45e8 |
LFPDUX B8, Y, INCY2
|
|
kusano |
2b45e8 |
bdz LL(23)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(22):
|
|
kusano |
2b45e8 |
fpmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fsmr A5, A6
|
|
kusano |
2b45e8 |
LFPDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
fpmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
LFXDUX A2, X, INCX2
|
|
kusano |
2b45e8 |
fsmr A6, A7
|
|
kusano |
2b45e8 |
LFPDUX B2, Y, INCY2
|
|
kusano |
2b45e8 |
fpmadd C3, ALPHA, A3, B3
|
|
kusano |
2b45e8 |
LFXDUX A3, X, INCX2
|
|
kusano |
2b45e8 |
fsmr A7, A8
|
|
kusano |
2b45e8 |
LFPDUX B3, Y, INCY2
|
|
kusano |
2b45e8 |
fpmadd C4, ALPHA, A4, B4
|
|
kusano |
2b45e8 |
LFXDUX A4, X, INCX2
|
|
kusano |
2b45e8 |
fsmr A8, A9
|
|
kusano |
2b45e8 |
LFPDUX B4, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C5, ALPHA, A5, B5
|
|
kusano |
2b45e8 |
LFXDUX A5, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B5, Y, INCY2
|
|
kusano |
2b45e8 |
fpmadd C6, ALPHA, A6, B6
|
|
kusano |
2b45e8 |
LFXDUX A6, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B6, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C7, ALPHA, A7, B7
|
|
kusano |
2b45e8 |
LFXDUX A7, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B7, Y, INCY2
|
|
kusano |
2b45e8 |
fpmadd C8, ALPHA, A8, B8
|
|
kusano |
2b45e8 |
LFXDUX A8, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B8, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmr A1, A9
|
|
kusano |
2b45e8 |
LFXDUX A9, X, INCX2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C2, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C3, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C4, YY, INCY2
|
|
kusano |
2b45e8 |
fsmr A1, A2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C5, YY, INCY2
|
|
kusano |
2b45e8 |
fsmr A2, A3
|
|
kusano |
2b45e8 |
STFPDUX C6, YY, INCY2
|
|
kusano |
2b45e8 |
fsmr A3, A4
|
|
kusano |
2b45e8 |
STFPDUX C7, YY, INCY2
|
|
kusano |
2b45e8 |
fsmr A4, A5
|
|
kusano |
2b45e8 |
STFPDUX C8, YY, INCY2
|
|
kusano |
2b45e8 |
bdnz LL(22)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(23):
|
|
kusano |
2b45e8 |
fpmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fsmr A5, A6
|
|
kusano |
2b45e8 |
fpmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
fsmr A6, A7
|
|
kusano |
2b45e8 |
fpmadd C3, ALPHA, A3, B3
|
|
kusano |
2b45e8 |
fsmr A7, A8
|
|
kusano |
2b45e8 |
fpmadd C4, ALPHA, A4, B4
|
|
kusano |
2b45e8 |
fsmr A8, A9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C5, ALPHA, A5, B5
|
|
kusano |
2b45e8 |
fpmadd C6, ALPHA, A6, B6
|
|
kusano |
2b45e8 |
fpmadd C7, ALPHA, A7, B7
|
|
kusano |
2b45e8 |
fpmadd C8, ALPHA, A8, B8
|
|
kusano |
2b45e8 |
fpmr A1, A9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C2, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C3, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C4, YY, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C5, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C6, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C7, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C8, YY, INCY2
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(25):
|
|
kusano |
2b45e8 |
andi. r0, N, 15
|
|
kusano |
2b45e8 |
beq LL(999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
andi. r0, N, 8
|
|
kusano |
2b45e8 |
beq LL(26)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFXDUX A2, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
LFXDUX A3, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B2, Y, INCY2
|
|
kusano |
2b45e8 |
LFXDUX A4, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B3, Y, INCY2
|
|
kusano |
2b45e8 |
LFXDUX A5, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B4, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsmr A1, A2
|
|
kusano |
2b45e8 |
fsmr A2, A3
|
|
kusano |
2b45e8 |
fsmr A3, A4
|
|
kusano |
2b45e8 |
fsmr A4, A5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fpmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fpmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
fpmadd C3, ALPHA, A3, B3
|
|
kusano |
2b45e8 |
fpmadd C4, ALPHA, A4, B4
|
|
kusano |
2b45e8 |
fpmr A1, A5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C2, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C3, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C4, YY, INCY2
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(26):
|
|
kusano |
2b45e8 |
andi. r0, N, 4
|
|
kusano |
2b45e8 |
beq LL(27)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFXDUX A2, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
LFXDUX A3, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B2, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsmr A1, A2
|
|
kusano |
2b45e8 |
fsmr A2, A3
|
|
kusano |
2b45e8 |
fpmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fpmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
fpmr A1, A3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
STFPDUX C2, YY, INCY2
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(27):
|
|
kusano |
2b45e8 |
andi. r0, N, 2
|
|
kusano |
2b45e8 |
beq LL(28)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFXDUX A2, X, INCX2
|
|
kusano |
2b45e8 |
LFPDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsmr A1, A2
|
|
kusano |
2b45e8 |
fpmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fpmr A1, A2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFPDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(28):
|
|
kusano |
2b45e8 |
andi. r0, N, 1
|
|
kusano |
2b45e8 |
beq LL(999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX B1, Y, INCY2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
STFDUX C1, YY, INCY2
|
|
kusano |
2b45e8 |
b LL(999)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
####
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(100):
|
|
kusano |
2b45e8 |
sub X, X, INCX
|
|
kusano |
2b45e8 |
sub Y, Y, INCY
|
|
kusano |
2b45e8 |
mr YY, Y
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
srawi. r0, N, 3
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq- LL(115)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX A1, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B1, Y, INCY
|
|
kusano |
2b45e8 |
LFDUX A2, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B2, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX A3, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B3, Y, INCY
|
|
kusano |
2b45e8 |
LFDUX A4, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B4, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX A5, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B5, Y, INCY
|
|
kusano |
2b45e8 |
LFDUX A6, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B6, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX A7, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B7, Y, INCY
|
|
kusano |
2b45e8 |
LFDUX A8, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B8, Y, INCY
|
|
kusano |
2b45e8 |
bdz LL(113)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(112):
|
|
kusano |
2b45e8 |
fmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
LFDUX A1, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B1, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
LFDUX A2, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B2, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C3, ALPHA, A3, B3
|
|
kusano |
2b45e8 |
LFDUX A3, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B3, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C4, ALPHA, A4, B4
|
|
kusano |
2b45e8 |
LFDUX A4, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B4, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C5, ALPHA, A5, B5
|
|
kusano |
2b45e8 |
LFDUX A5, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B5, Y, INCY
|
|
kusano |
2b45e8 |
fmadd C6, ALPHA, A6, B6
|
|
kusano |
2b45e8 |
LFDUX A6, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B6, Y, INCY
|
|
kusano |
2b45e8 |
fmadd C7, ALPHA, A7, B7
|
|
kusano |
2b45e8 |
LFDUX A7, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B7, Y, INCY
|
|
kusano |
2b45e8 |
fmadd C8, ALPHA, A8, B8
|
|
kusano |
2b45e8 |
LFDUX A8, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B8, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFDUX C1, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C2, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C3, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C4, YY, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFDUX C5, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C6, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C7, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C8, YY, INCY
|
|
kusano |
2b45e8 |
bdnz LL(112)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(113):
|
|
kusano |
2b45e8 |
fmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
fmadd C3, ALPHA, A3, B3
|
|
kusano |
2b45e8 |
fmadd C4, ALPHA, A4, B4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C5, ALPHA, A5, B5
|
|
kusano |
2b45e8 |
fmadd C6, ALPHA, A6, B6
|
|
kusano |
2b45e8 |
STFDUX C1, YY, INCY
|
|
kusano |
2b45e8 |
fmadd C7, ALPHA, A7, B7
|
|
kusano |
2b45e8 |
STFDUX C2, YY, INCY
|
|
kusano |
2b45e8 |
fmadd C8, ALPHA, A8, B8
|
|
kusano |
2b45e8 |
STFDUX C3, YY, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFDUX C4, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C5, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C6, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C7, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C8, YY, INCY
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(115):
|
|
kusano |
2b45e8 |
andi. r0, N, 7
|
|
kusano |
2b45e8 |
beq LL(999)
|
|
kusano |
2b45e8 |
andi. r0, N, 4
|
|
kusano |
2b45e8 |
beq LL(117)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX A1, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B1, Y, INCY
|
|
kusano |
2b45e8 |
LFDUX A2, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B2, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX A3, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B3, Y, INCY
|
|
kusano |
2b45e8 |
LFDUX A4, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B4, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
fmadd C3, ALPHA, A3, B3
|
|
kusano |
2b45e8 |
fmadd C4, ALPHA, A4, B4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFDUX C1, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C2, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C3, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C4, YY, INCY
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(117):
|
|
kusano |
2b45e8 |
andi. r0, N, 2
|
|
kusano |
2b45e8 |
beq LL(118)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX A1, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B1, Y, INCY
|
|
kusano |
2b45e8 |
LFDUX A2, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B2, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
fmadd C2, ALPHA, A2, B2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STFDUX C1, YY, INCY
|
|
kusano |
2b45e8 |
STFDUX C2, YY, INCY
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(118):
|
|
kusano |
2b45e8 |
andi. r0, N, 1
|
|
kusano |
2b45e8 |
beq LL(999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX A1, X, INCX
|
|
kusano |
2b45e8 |
LFDUX B1, Y, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd C1, ALPHA, A1, B1
|
|
kusano |
2b45e8 |
STFDUX C1, YY, INCY
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(999):
|
|
kusano |
2b45e8 |
li r10, 16
|
|
kusano |
2b45e8 |
subi SP, SP, 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
lfpdux f25, SP, r10
|
|
kusano |
2b45e8 |
lfpdux f24, SP, r10
|
|
kusano |
2b45e8 |
lfpdux f23, SP, r10
|
|
kusano |
2b45e8 |
lfpdux f22, SP, r10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
lfpdux f21, SP, r10
|
|
kusano |
2b45e8 |
lfpdux f20, SP, r10
|
|
kusano |
2b45e8 |
lfpdux f19, SP, r10
|
|
kusano |
2b45e8 |
lfpdux f18, SP, r10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
lfpdux f17, SP, r10
|
|
kusano |
2b45e8 |
lfpdux f16, SP, r10
|
|
kusano |
2b45e8 |
lfpdux f15, SP, r10
|
|
kusano |
2b45e8 |
lfpdux f14, SP, r10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addi SP, SP, 16
|
|
kusano |
2b45e8 |
blr
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
EPILOGUE
|