|
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 r4
|
|
kusano |
2b45e8 |
#define INCX r5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define NN r6
|
|
kusano |
2b45e8 |
#define XX r7
|
|
kusano |
2b45e8 |
#define PREA r8
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define FZERO 144(SP)
|
|
kusano |
2b45e8 |
#define FONE 148(SP)
|
|
kusano |
2b45e8 |
#define FMAX 152(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACKSIZE 160
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROLOGUE
|
|
kusano |
2b45e8 |
PROFCODE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addi SP, SP, -STACKSIZE
|
|
kusano |
2b45e8 |
li r10, 0
|
|
kusano |
2b45e8 |
lis r11, 0x3f80
|
|
kusano |
2b45e8 |
lis r12, 0x5fe0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
stfd f14, 0(SP)
|
|
kusano |
2b45e8 |
stfd f15, 8(SP)
|
|
kusano |
2b45e8 |
stfd f16, 16(SP)
|
|
kusano |
2b45e8 |
stfd f17, 24(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
stfd f18, 32(SP)
|
|
kusano |
2b45e8 |
stfd f19, 40(SP)
|
|
kusano |
2b45e8 |
stfd f20, 48(SP)
|
|
kusano |
2b45e8 |
stfd f21, 56(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
stfd f22, 64(SP)
|
|
kusano |
2b45e8 |
stfd f23, 72(SP)
|
|
kusano |
2b45e8 |
stfd f24, 80(SP)
|
|
kusano |
2b45e8 |
stfd f25, 88(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
stfd f26, 96(SP)
|
|
kusano |
2b45e8 |
stfd f27, 104(SP)
|
|
kusano |
2b45e8 |
stfd f28, 112(SP)
|
|
kusano |
2b45e8 |
stfd f29, 120(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
stfd f30, 128(SP)
|
|
kusano |
2b45e8 |
stfd f31, 136(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
stw r10, FZERO
|
|
kusano |
2b45e8 |
stw r11, FONE
|
|
kusano |
2b45e8 |
stw r12, FMAX
|
|
kusano |
2b45e8 |
stw r10, 4 + FMAX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
lfs f1, FZERO
|
|
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 |
slwi INCX, INCX, BASE_SHIFT
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
li PREA, L1_PREFETCHSIZE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
cmpwi cr0, N, 0
|
|
kusano |
2b45e8 |
ble- LL(9999)
|
|
kusano |
2b45e8 |
cmpwi cr0, INCX, 0
|
|
kusano |
2b45e8 |
ble- LL(9999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mr NN, N
|
|
kusano |
2b45e8 |
mr XX, X
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f1, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
add X, X, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fabs f0, f1
|
|
kusano |
2b45e8 |
fabs f2, f1
|
|
kusano |
2b45e8 |
fabs f3, f1
|
|
kusano |
2b45e8 |
fabs f4, f1
|
|
kusano |
2b45e8 |
fabs f5, f1
|
|
kusano |
2b45e8 |
fabs f6, f1
|
|
kusano |
2b45e8 |
fabs f7, f1
|
|
kusano |
2b45e8 |
fabs f1, f1
|
|
kusano |
2b45e8 |
subi N, N, 1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
cmpwi cr0, N, 0
|
|
kusano |
2b45e8 |
ble- LL(9999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
cmpwi cr0, INCX, SIZE
|
|
kusano |
2b45e8 |
bne- cr0, LL(1000)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
srawi. r0, N, 4
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq- cr0, LL(50)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f24, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f25, 1 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f26, 2 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f27, 3 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f28, 4 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f29, 5 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f30, 6 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f31, 7 * SIZE(X)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fabs f8, f24
|
|
kusano |
2b45e8 |
fabs f9, f25
|
|
kusano |
2b45e8 |
fabs f10, f26
|
|
kusano |
2b45e8 |
fabs f11, f27
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f24, 8 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f25, 9 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f26, 10 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f27, 11 * SIZE(X)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fabs f12, f28
|
|
kusano |
2b45e8 |
fabs f13, f29
|
|
kusano |
2b45e8 |
fabs f14, f30
|
|
kusano |
2b45e8 |
fabs f15, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f28, 12 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f29, 13 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f30, 14 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f31, 15 * SIZE(X)
|
|
kusano |
2b45e8 |
bdz LL(20)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(10):
|
|
kusano |
2b45e8 |
fsub f16, f0, f8
|
|
kusano |
2b45e8 |
fsub f17, f1, f9
|
|
kusano |
2b45e8 |
fsub f18, f2, f10
|
|
kusano |
2b45e8 |
fsub f19, f3, f11
|
|
kusano |
2b45e8 |
fsub f20, f4, f12
|
|
kusano |
2b45e8 |
fsub f21, f5, f13
|
|
kusano |
2b45e8 |
fsub f22, f6, f14
|
|
kusano |
2b45e8 |
fsub f23, f7, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f0, f16, f0, f8
|
|
kusano |
2b45e8 |
fabs f8, f24
|
|
kusano |
2b45e8 |
fsel f1, f17, f1, f9
|
|
kusano |
2b45e8 |
fabs f9, f25
|
|
kusano |
2b45e8 |
fsel f2, f18, f2, f10
|
|
kusano |
2b45e8 |
fabs f10, f26
|
|
kusano |
2b45e8 |
fsel f3, f19, f3, f11
|
|
kusano |
2b45e8 |
fabs f11, f27
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f24, 16 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f25, 17 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f26, 18 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f27, 19 * SIZE(X)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f4, f20, f4, f12
|
|
kusano |
2b45e8 |
fabs f12, f28
|
|
kusano |
2b45e8 |
fsel f5, f21, f5, f13
|
|
kusano |
2b45e8 |
fabs f13, f29
|
|
kusano |
2b45e8 |
fsel f6, f22, f6, f14
|
|
kusano |
2b45e8 |
fabs f14, f30
|
|
kusano |
2b45e8 |
fsel f7, f23, f7, f15
|
|
kusano |
2b45e8 |
fabs f15, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f28, 20 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f29, 21 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f30, 22 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f31, 23 * SIZE(X)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsub f16, f0, f8
|
|
kusano |
2b45e8 |
fsub f17, f1, f9
|
|
kusano |
2b45e8 |
fsub f18, f2, f10
|
|
kusano |
2b45e8 |
fsub f19, f3, f11
|
|
kusano |
2b45e8 |
fsub f20, f4, f12
|
|
kusano |
2b45e8 |
fsub f21, f5, f13
|
|
kusano |
2b45e8 |
fsub f22, f6, f14
|
|
kusano |
2b45e8 |
fsub f23, f7, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f0, f16, f0, f8
|
|
kusano |
2b45e8 |
fabs f8, f24
|
|
kusano |
2b45e8 |
fsel f1, f17, f1, f9
|
|
kusano |
2b45e8 |
fabs f9, f25
|
|
kusano |
2b45e8 |
fsel f2, f18, f2, f10
|
|
kusano |
2b45e8 |
fabs f10, f26
|
|
kusano |
2b45e8 |
fsel f3, f19, f3, f11
|
|
kusano |
2b45e8 |
fabs f11, f27
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f24, 24 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f25, 25 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f26, 26 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f27, 27 * SIZE(X)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f4, f20, f4, f12
|
|
kusano |
2b45e8 |
fabs f12, f28
|
|
kusano |
2b45e8 |
fsel f5, f21, f5, f13
|
|
kusano |
2b45e8 |
fabs f13, f29
|
|
kusano |
2b45e8 |
fsel f6, f22, f6, f14
|
|
kusano |
2b45e8 |
fabs f14, f30
|
|
kusano |
2b45e8 |
fsel f7, f23, f7, f15
|
|
kusano |
2b45e8 |
fabs f15, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f28, 28 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f29, 29 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f30, 30 * SIZE(X)
|
|
kusano |
2b45e8 |
LFD f31, 31 * SIZE(X)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef POWER6
|
|
kusano |
2b45e8 |
L1_PREFETCH X, PREA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
addi X, X, 16 * SIZE
|
|
kusano |
2b45e8 |
#ifdef POWER6
|
|
kusano |
2b45e8 |
L1_PREFETCH X, PREA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
bdnz LL(10)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(20):
|
|
kusano |
2b45e8 |
fsub f16, f0, f8
|
|
kusano |
2b45e8 |
fsub f17, f1, f9
|
|
kusano |
2b45e8 |
fsub f18, f2, f10
|
|
kusano |
2b45e8 |
fsub f19, f3, f11
|
|
kusano |
2b45e8 |
fsub f20, f4, f12
|
|
kusano |
2b45e8 |
fsub f21, f5, f13
|
|
kusano |
2b45e8 |
fsub f22, f6, f14
|
|
kusano |
2b45e8 |
fsub f23, f7, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f0, f16, f0, f8
|
|
kusano |
2b45e8 |
fabs f8, f24
|
|
kusano |
2b45e8 |
fsel f1, f17, f1, f9
|
|
kusano |
2b45e8 |
fabs f9, f25
|
|
kusano |
2b45e8 |
fsel f2, f18, f2, f10
|
|
kusano |
2b45e8 |
fabs f10, f26
|
|
kusano |
2b45e8 |
fsel f3, f19, f3, f11
|
|
kusano |
2b45e8 |
fabs f11, f27
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f4, f20, f4, f12
|
|
kusano |
2b45e8 |
fabs f12, f28
|
|
kusano |
2b45e8 |
fsel f5, f21, f5, f13
|
|
kusano |
2b45e8 |
fabs f13, f29
|
|
kusano |
2b45e8 |
fsel f6, f22, f6, f14
|
|
kusano |
2b45e8 |
fabs f14, f30
|
|
kusano |
2b45e8 |
fsel f7, f23, f7, f15
|
|
kusano |
2b45e8 |
fabs f15, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsub f16, f0, f8
|
|
kusano |
2b45e8 |
fsub f17, f1, f9
|
|
kusano |
2b45e8 |
fsub f18, f2, f10
|
|
kusano |
2b45e8 |
fsub f19, f3, f11
|
|
kusano |
2b45e8 |
fsub f20, f4, f12
|
|
kusano |
2b45e8 |
fsub f21, f5, f13
|
|
kusano |
2b45e8 |
fsub f22, f6, f14
|
|
kusano |
2b45e8 |
fsub f23, f7, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f0, f16, f0, f8
|
|
kusano |
2b45e8 |
fsel f1, f17, f1, f9
|
|
kusano |
2b45e8 |
fsel f2, f18, f2, f10
|
|
kusano |
2b45e8 |
fsel f3, f19, f3, f11
|
|
kusano |
2b45e8 |
fsel f4, f20, f4, f12
|
|
kusano |
2b45e8 |
fsel f5, f21, f5, f13
|
|
kusano |
2b45e8 |
fsel f6, f22, f6, f14
|
|
kusano |
2b45e8 |
fsel f7, f23, f7, f15
|
|
kusano |
2b45e8 |
addi X, X, 16 * SIZE
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(50):
|
|
kusano |
2b45e8 |
andi. r0, N, 15
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq LL(100)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(60):
|
|
kusano |
2b45e8 |
LFD f8, 0 * SIZE(X)
|
|
kusano |
2b45e8 |
addi X, X, 1 * SIZE
|
|
kusano |
2b45e8 |
fabs f8, f8
|
|
kusano |
2b45e8 |
fsub f16, f1, f8
|
|
kusano |
2b45e8 |
fsel f1, f16, f1, f8
|
|
kusano |
2b45e8 |
bdnz LL(60)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(100):
|
|
kusano |
2b45e8 |
fsub f8, f0, f1
|
|
kusano |
2b45e8 |
fsub f9, f2, f3
|
|
kusano |
2b45e8 |
fsub f10, f4, f5
|
|
kusano |
2b45e8 |
fsub f11, f6, f7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f0, f8, f0, f1
|
|
kusano |
2b45e8 |
fsel f2, f9, f2, f3
|
|
kusano |
2b45e8 |
fsel f4, f10, f4, f5
|
|
kusano |
2b45e8 |
fsel f6, f11, f6, f7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsub f8, f0, f2
|
|
kusano |
2b45e8 |
fsub f9, f4, f6
|
|
kusano |
2b45e8 |
fsel f0, f8, f0, f2
|
|
kusano |
2b45e8 |
fsel f4, f9, f4, f6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsub f8, f0, f4
|
|
kusano |
2b45e8 |
fsel f31, f8, f0, f4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
lfs f1, FZERO
|
|
kusano |
2b45e8 |
lfs f0, FONE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fcmpu cr0, f1, f31
|
|
kusano |
2b45e8 |
beq- cr0, LL(9999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fdiv f30, f0, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmr f0, f1
|
|
kusano |
2b45e8 |
fmr f2, f1
|
|
kusano |
2b45e8 |
fmr f3, f1
|
|
kusano |
2b45e8 |
fmr f4, f1
|
|
kusano |
2b45e8 |
fmr f5, f1
|
|
kusano |
2b45e8 |
fmr f6, f1
|
|
kusano |
2b45e8 |
fmr f7, f1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
srawi. r0, NN, 4
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq- cr0, LL(250)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f8, 0 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f9, 1 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f10, 2 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f11, 3 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f12, 4 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f13, 5 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f14, 6 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f15, 7 * SIZE(XX)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul f16, f30, f8
|
|
kusano |
2b45e8 |
fmul f17, f30, f9
|
|
kusano |
2b45e8 |
fmul f18, f30, f10
|
|
kusano |
2b45e8 |
fmul f19, f30, f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f8, 8 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f9, 9 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f10, 10 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f11, 11 * SIZE(XX)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul f20, f30, f12
|
|
kusano |
2b45e8 |
fmul f21, f30, f13
|
|
kusano |
2b45e8 |
fmul f22, f30, f14
|
|
kusano |
2b45e8 |
fmul f23, f30, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f12, 12 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f13, 13 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f14, 14 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f15, 15 * SIZE(XX)
|
|
kusano |
2b45e8 |
bdz LL(220)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(210):
|
|
kusano |
2b45e8 |
fmadd f0, f16, f16, f0
|
|
kusano |
2b45e8 |
fmul f16, f30, f8
|
|
kusano |
2b45e8 |
fmadd f1, f17, f17, f1
|
|
kusano |
2b45e8 |
fmul f17, f30, f9
|
|
kusano |
2b45e8 |
fmadd f2, f18, f18, f2
|
|
kusano |
2b45e8 |
fmul f18, f30, f10
|
|
kusano |
2b45e8 |
fmadd f3, f19, f19, f3
|
|
kusano |
2b45e8 |
fmul f19, f30, f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f8, 16 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f9, 17 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f10, 18 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f11, 19 * SIZE(XX)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd f4, f20, f20, f4
|
|
kusano |
2b45e8 |
fmul f20, f30, f12
|
|
kusano |
2b45e8 |
fmadd f5, f21, f21, f5
|
|
kusano |
2b45e8 |
fmul f21, f30, f13
|
|
kusano |
2b45e8 |
fmadd f6, f22, f22, f6
|
|
kusano |
2b45e8 |
fmul f22, f30, f14
|
|
kusano |
2b45e8 |
fmadd f7, f23, f23, f7
|
|
kusano |
2b45e8 |
fmul f23, f30, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f12, 20 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f13, 21 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f14, 22 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f15, 23 * SIZE(XX)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd f0, f16, f16, f0
|
|
kusano |
2b45e8 |
fmul f16, f30, f8
|
|
kusano |
2b45e8 |
fmadd f1, f17, f17, f1
|
|
kusano |
2b45e8 |
fmul f17, f30, f9
|
|
kusano |
2b45e8 |
fmadd f2, f18, f18, f2
|
|
kusano |
2b45e8 |
fmul f18, f30, f10
|
|
kusano |
2b45e8 |
fmadd f3, f19, f19, f3
|
|
kusano |
2b45e8 |
fmul f19, f30, f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f8, 24 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f9, 25 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f10, 26 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f11, 27 * SIZE(XX)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd f4, f20, f20, f4
|
|
kusano |
2b45e8 |
fmul f20, f30, f12
|
|
kusano |
2b45e8 |
fmadd f5, f21, f21, f5
|
|
kusano |
2b45e8 |
fmul f21, f30, f13
|
|
kusano |
2b45e8 |
fmadd f6, f22, f22, f6
|
|
kusano |
2b45e8 |
fmul f22, f30, f14
|
|
kusano |
2b45e8 |
fmadd f7, f23, f23, f7
|
|
kusano |
2b45e8 |
fmul f23, f30, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFD f12, 28 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f13, 29 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f14, 30 * SIZE(XX)
|
|
kusano |
2b45e8 |
LFD f15, 31 * SIZE(XX)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef POWER6
|
|
kusano |
2b45e8 |
L1_PREFETCH XX, PREA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
addi XX, XX, 16 * SIZE
|
|
kusano |
2b45e8 |
#ifdef POWER6
|
|
kusano |
2b45e8 |
L1_PREFETCH XX, PREA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
bdnz LL(210)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(220):
|
|
kusano |
2b45e8 |
fmadd f0, f16, f16, f0
|
|
kusano |
2b45e8 |
fmul f16, f30, f8
|
|
kusano |
2b45e8 |
fmadd f1, f17, f17, f1
|
|
kusano |
2b45e8 |
fmul f17, f30, f9
|
|
kusano |
2b45e8 |
fmadd f2, f18, f18, f2
|
|
kusano |
2b45e8 |
fmul f18, f30, f10
|
|
kusano |
2b45e8 |
fmadd f3, f19, f19, f3
|
|
kusano |
2b45e8 |
fmul f19, f30, f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd f4, f20, f20, f4
|
|
kusano |
2b45e8 |
fmul f20, f30, f12
|
|
kusano |
2b45e8 |
fmadd f5, f21, f21, f5
|
|
kusano |
2b45e8 |
fmul f21, f30, f13
|
|
kusano |
2b45e8 |
fmadd f6, f22, f22, f6
|
|
kusano |
2b45e8 |
fmul f22, f30, f14
|
|
kusano |
2b45e8 |
fmadd f7, f23, f23, f7
|
|
kusano |
2b45e8 |
fmul f23, f30, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd f0, f16, f16, f0
|
|
kusano |
2b45e8 |
fmadd f1, f17, f17, f1
|
|
kusano |
2b45e8 |
fmadd f2, f18, f18, f2
|
|
kusano |
2b45e8 |
fmadd f3, f19, f19, f3
|
|
kusano |
2b45e8 |
fmadd f4, f20, f20, f4
|
|
kusano |
2b45e8 |
fmadd f5, f21, f21, f5
|
|
kusano |
2b45e8 |
fmadd f6, f22, f22, f6
|
|
kusano |
2b45e8 |
fmadd f7, f23, f23, f7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addi XX, XX, 16 * SIZE
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(250):
|
|
kusano |
2b45e8 |
andi. r0, NN, 15
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq- cr0, LL(270)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(260):
|
|
kusano |
2b45e8 |
LFD f8, 0 * SIZE(XX)
|
|
kusano |
2b45e8 |
addi XX, XX, 1 * SIZE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul f16, f30, f8
|
|
kusano |
2b45e8 |
fmadd f0, f16, f16, f0
|
|
kusano |
2b45e8 |
bdnz LL(260)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(270):
|
|
kusano |
2b45e8 |
fadd f0, f0, f1
|
|
kusano |
2b45e8 |
fadd f2, f2, f3
|
|
kusano |
2b45e8 |
fadd f4, f4, f5
|
|
kusano |
2b45e8 |
fadd f6, f6, f7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fadd f0, f0, f2
|
|
kusano |
2b45e8 |
fadd f4, f4, f6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fadd f0, f0, f4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsqrt f0, f0
|
|
kusano |
2b45e8 |
fmul f1, f31, f0
|
|
kusano |
2b45e8 |
b LL(9999)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(1000):
|
|
kusano |
2b45e8 |
sub X, X, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
srawi. r0, N, 4
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq- LL(1050)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f24, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f25, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f26, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f27, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f28, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f29, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f30, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f31, X, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fabs f8, f24
|
|
kusano |
2b45e8 |
fabs f9, f25
|
|
kusano |
2b45e8 |
fabs f10, f26
|
|
kusano |
2b45e8 |
fabs f11, f27
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f24, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f25, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f26, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f27, X, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fabs f12, f28
|
|
kusano |
2b45e8 |
fabs f13, f29
|
|
kusano |
2b45e8 |
fabs f14, f30
|
|
kusano |
2b45e8 |
fabs f15, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f28, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f29, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f30, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f31, X, INCX
|
|
kusano |
2b45e8 |
bdz LL(1020)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(1010):
|
|
kusano |
2b45e8 |
fsub f16, f0, f8
|
|
kusano |
2b45e8 |
fsub f17, f1, f9
|
|
kusano |
2b45e8 |
fsub f18, f2, f10
|
|
kusano |
2b45e8 |
fsub f19, f3, f11
|
|
kusano |
2b45e8 |
fsub f20, f4, f12
|
|
kusano |
2b45e8 |
fsub f21, f5, f13
|
|
kusano |
2b45e8 |
fsub f22, f6, f14
|
|
kusano |
2b45e8 |
fsub f23, f7, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f0, f16, f0, f8
|
|
kusano |
2b45e8 |
fabs f8, f24
|
|
kusano |
2b45e8 |
fsel f1, f17, f1, f9
|
|
kusano |
2b45e8 |
fabs f9, f25
|
|
kusano |
2b45e8 |
fsel f2, f18, f2, f10
|
|
kusano |
2b45e8 |
fabs f10, f26
|
|
kusano |
2b45e8 |
fsel f3, f19, f3, f11
|
|
kusano |
2b45e8 |
fabs f11, f27
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f24, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f25, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f26, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f27, X, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f4, f20, f4, f12
|
|
kusano |
2b45e8 |
fabs f12, f28
|
|
kusano |
2b45e8 |
fsel f5, f21, f5, f13
|
|
kusano |
2b45e8 |
fabs f13, f29
|
|
kusano |
2b45e8 |
fsel f6, f22, f6, f14
|
|
kusano |
2b45e8 |
fabs f14, f30
|
|
kusano |
2b45e8 |
fsel f7, f23, f7, f15
|
|
kusano |
2b45e8 |
fabs f15, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f28, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f29, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f30, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f31, X, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsub f16, f0, f8
|
|
kusano |
2b45e8 |
fsub f17, f1, f9
|
|
kusano |
2b45e8 |
fsub f18, f2, f10
|
|
kusano |
2b45e8 |
fsub f19, f3, f11
|
|
kusano |
2b45e8 |
fsub f20, f4, f12
|
|
kusano |
2b45e8 |
fsub f21, f5, f13
|
|
kusano |
2b45e8 |
fsub f22, f6, f14
|
|
kusano |
2b45e8 |
fsub f23, f7, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f0, f16, f0, f8
|
|
kusano |
2b45e8 |
fabs f8, f24
|
|
kusano |
2b45e8 |
fsel f1, f17, f1, f9
|
|
kusano |
2b45e8 |
fabs f9, f25
|
|
kusano |
2b45e8 |
fsel f2, f18, f2, f10
|
|
kusano |
2b45e8 |
fabs f10, f26
|
|
kusano |
2b45e8 |
fsel f3, f19, f3, f11
|
|
kusano |
2b45e8 |
fabs f11, f27
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f24, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f25, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f26, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f27, X, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f4, f20, f4, f12
|
|
kusano |
2b45e8 |
fabs f12, f28
|
|
kusano |
2b45e8 |
fsel f5, f21, f5, f13
|
|
kusano |
2b45e8 |
fabs f13, f29
|
|
kusano |
2b45e8 |
fsel f6, f22, f6, f14
|
|
kusano |
2b45e8 |
fabs f14, f30
|
|
kusano |
2b45e8 |
fsel f7, f23, f7, f15
|
|
kusano |
2b45e8 |
fabs f15, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f28, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f29, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f30, X, INCX
|
|
kusano |
2b45e8 |
LFDUX f31, X, INCX
|
|
kusano |
2b45e8 |
bdnz LL(1010)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(1020):
|
|
kusano |
2b45e8 |
fsub f16, f0, f8
|
|
kusano |
2b45e8 |
fsub f17, f1, f9
|
|
kusano |
2b45e8 |
fsub f18, f2, f10
|
|
kusano |
2b45e8 |
fsub f19, f3, f11
|
|
kusano |
2b45e8 |
fsub f20, f4, f12
|
|
kusano |
2b45e8 |
fsub f21, f5, f13
|
|
kusano |
2b45e8 |
fsub f22, f6, f14
|
|
kusano |
2b45e8 |
fsub f23, f7, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f0, f16, f0, f8
|
|
kusano |
2b45e8 |
fabs f8, f24
|
|
kusano |
2b45e8 |
fsel f1, f17, f1, f9
|
|
kusano |
2b45e8 |
fabs f9, f25
|
|
kusano |
2b45e8 |
fsel f2, f18, f2, f10
|
|
kusano |
2b45e8 |
fabs f10, f26
|
|
kusano |
2b45e8 |
fsel f3, f19, f3, f11
|
|
kusano |
2b45e8 |
fabs f11, f27
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f4, f20, f4, f12
|
|
kusano |
2b45e8 |
fabs f12, f28
|
|
kusano |
2b45e8 |
fsel f5, f21, f5, f13
|
|
kusano |
2b45e8 |
fabs f13, f29
|
|
kusano |
2b45e8 |
fsel f6, f22, f6, f14
|
|
kusano |
2b45e8 |
fabs f14, f30
|
|
kusano |
2b45e8 |
fsel f7, f23, f7, f15
|
|
kusano |
2b45e8 |
fabs f15, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsub f16, f0, f8
|
|
kusano |
2b45e8 |
fsub f17, f1, f9
|
|
kusano |
2b45e8 |
fsub f18, f2, f10
|
|
kusano |
2b45e8 |
fsub f19, f3, f11
|
|
kusano |
2b45e8 |
fsub f20, f4, f12
|
|
kusano |
2b45e8 |
fsub f21, f5, f13
|
|
kusano |
2b45e8 |
fsub f22, f6, f14
|
|
kusano |
2b45e8 |
fsub f23, f7, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f0, f16, f0, f8
|
|
kusano |
2b45e8 |
fsel f1, f17, f1, f9
|
|
kusano |
2b45e8 |
fsel f2, f18, f2, f10
|
|
kusano |
2b45e8 |
fsel f3, f19, f3, f11
|
|
kusano |
2b45e8 |
fsel f4, f20, f4, f12
|
|
kusano |
2b45e8 |
fsel f5, f21, f5, f13
|
|
kusano |
2b45e8 |
fsel f6, f22, f6, f14
|
|
kusano |
2b45e8 |
fsel f7, f23, f7, f15
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(1050):
|
|
kusano |
2b45e8 |
andi. r0, N, 15
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq LL(1999)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(1060):
|
|
kusano |
2b45e8 |
LFDUX f8, X, INCX
|
|
kusano |
2b45e8 |
fabs f8, f8
|
|
kusano |
2b45e8 |
fsub f16, f1, f8
|
|
kusano |
2b45e8 |
fsel f1, f16, f1, f8
|
|
kusano |
2b45e8 |
bdnz LL(1060)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(1999):
|
|
kusano |
2b45e8 |
fsub f8, f0, f1
|
|
kusano |
2b45e8 |
fsub f9, f2, f3
|
|
kusano |
2b45e8 |
fsub f10, f4, f5
|
|
kusano |
2b45e8 |
fsub f11, f6, f7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsel f0, f8, f0, f1
|
|
kusano |
2b45e8 |
fsel f2, f9, f2, f3
|
|
kusano |
2b45e8 |
fsel f4, f10, f4, f5
|
|
kusano |
2b45e8 |
fsel f6, f11, f6, f7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsub f8, f0, f2
|
|
kusano |
2b45e8 |
fsub f9, f4, f6
|
|
kusano |
2b45e8 |
fsel f0, f8, f0, f2
|
|
kusano |
2b45e8 |
fsel f4, f9, f4, f6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsub f8, f0, f4
|
|
kusano |
2b45e8 |
fsel f31, f8, f0, f4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
lfs f1, FZERO
|
|
kusano |
2b45e8 |
lfs f0, FONE
|
|
kusano |
2b45e8 |
lfd f2, FMAX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fcmpu cr0, f1, f31
|
|
kusano |
2b45e8 |
beq- cr0, LL(9999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fdiv f30, f0, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmr f0, f1
|
|
kusano |
2b45e8 |
fmr f2, f1
|
|
kusano |
2b45e8 |
fmr f3, f1
|
|
kusano |
2b45e8 |
fmr f4, f1
|
|
kusano |
2b45e8 |
fmr f5, f1
|
|
kusano |
2b45e8 |
fmr f6, f1
|
|
kusano |
2b45e8 |
fmr f7, f1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
sub XX, XX, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
srawi. r0, NN, 4
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq- cr0, LL(2150)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f8, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f9, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f10, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f11, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f12, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f13, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f14, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f15, XX, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul f16, f30, f8
|
|
kusano |
2b45e8 |
fmul f17, f30, f9
|
|
kusano |
2b45e8 |
fmul f18, f30, f10
|
|
kusano |
2b45e8 |
fmul f19, f30, f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f8, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f9, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f10, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f11, XX, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul f20, f30, f12
|
|
kusano |
2b45e8 |
fmul f21, f30, f13
|
|
kusano |
2b45e8 |
fmul f22, f30, f14
|
|
kusano |
2b45e8 |
fmul f23, f30, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f12, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f13, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f14, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f15, XX, INCX
|
|
kusano |
2b45e8 |
bdz LL(2120)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(2110):
|
|
kusano |
2b45e8 |
fmadd f0, f16, f16, f0
|
|
kusano |
2b45e8 |
fmul f16, f30, f8
|
|
kusano |
2b45e8 |
fmadd f1, f17, f17, f1
|
|
kusano |
2b45e8 |
fmul f17, f30, f9
|
|
kusano |
2b45e8 |
fmadd f2, f18, f18, f2
|
|
kusano |
2b45e8 |
fmul f18, f30, f10
|
|
kusano |
2b45e8 |
fmadd f3, f19, f19, f3
|
|
kusano |
2b45e8 |
fmul f19, f30, f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f8, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f9, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f10, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f11, XX, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd f4, f20, f20, f4
|
|
kusano |
2b45e8 |
fmul f20, f30, f12
|
|
kusano |
2b45e8 |
fmadd f5, f21, f21, f5
|
|
kusano |
2b45e8 |
fmul f21, f30, f13
|
|
kusano |
2b45e8 |
fmadd f6, f22, f22, f6
|
|
kusano |
2b45e8 |
fmul f22, f30, f14
|
|
kusano |
2b45e8 |
fmadd f7, f23, f23, f7
|
|
kusano |
2b45e8 |
fmul f23, f30, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f12, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f13, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f14, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f15, XX, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd f0, f16, f16, f0
|
|
kusano |
2b45e8 |
fmul f16, f30, f8
|
|
kusano |
2b45e8 |
fmadd f1, f17, f17, f1
|
|
kusano |
2b45e8 |
fmul f17, f30, f9
|
|
kusano |
2b45e8 |
fmadd f2, f18, f18, f2
|
|
kusano |
2b45e8 |
fmul f18, f30, f10
|
|
kusano |
2b45e8 |
fmadd f3, f19, f19, f3
|
|
kusano |
2b45e8 |
fmul f19, f30, f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f8, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f9, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f10, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f11, XX, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd f4, f20, f20, f4
|
|
kusano |
2b45e8 |
fmul f20, f30, f12
|
|
kusano |
2b45e8 |
fmadd f5, f21, f21, f5
|
|
kusano |
2b45e8 |
fmul f21, f30, f13
|
|
kusano |
2b45e8 |
fmadd f6, f22, f22, f6
|
|
kusano |
2b45e8 |
fmul f22, f30, f14
|
|
kusano |
2b45e8 |
fmadd f7, f23, f23, f7
|
|
kusano |
2b45e8 |
fmul f23, f30, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f12, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f13, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f14, XX, INCX
|
|
kusano |
2b45e8 |
LFDUX f15, XX, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
bdnz LL(2110)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(2120):
|
|
kusano |
2b45e8 |
fmadd f0, f16, f16, f0
|
|
kusano |
2b45e8 |
fmul f16, f30, f8
|
|
kusano |
2b45e8 |
fmadd f1, f17, f17, f1
|
|
kusano |
2b45e8 |
fmul f17, f30, f9
|
|
kusano |
2b45e8 |
fmadd f2, f18, f18, f2
|
|
kusano |
2b45e8 |
fmul f18, f30, f10
|
|
kusano |
2b45e8 |
fmadd f3, f19, f19, f3
|
|
kusano |
2b45e8 |
fmul f19, f30, f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd f4, f20, f20, f4
|
|
kusano |
2b45e8 |
fmul f20, f30, f12
|
|
kusano |
2b45e8 |
fmadd f5, f21, f21, f5
|
|
kusano |
2b45e8 |
fmul f21, f30, f13
|
|
kusano |
2b45e8 |
fmadd f6, f22, f22, f6
|
|
kusano |
2b45e8 |
fmul f22, f30, f14
|
|
kusano |
2b45e8 |
fmadd f7, f23, f23, f7
|
|
kusano |
2b45e8 |
fmul f23, f30, f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmadd f0, f16, f16, f0
|
|
kusano |
2b45e8 |
fmadd f1, f17, f17, f1
|
|
kusano |
2b45e8 |
fmadd f2, f18, f18, f2
|
|
kusano |
2b45e8 |
fmadd f3, f19, f19, f3
|
|
kusano |
2b45e8 |
fmadd f4, f20, f20, f4
|
|
kusano |
2b45e8 |
fmadd f5, f21, f21, f5
|
|
kusano |
2b45e8 |
fmadd f6, f22, f22, f6
|
|
kusano |
2b45e8 |
fmadd f7, f23, f23, f7
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(2150):
|
|
kusano |
2b45e8 |
andi. r0, NN, 15
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq- cr0, LL(2170)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(2160):
|
|
kusano |
2b45e8 |
LFDUX f8, XX, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fmul f16, f30, f8
|
|
kusano |
2b45e8 |
fmadd f0, f16, f16, f0
|
|
kusano |
2b45e8 |
bdnz LL(2160)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(2170):
|
|
kusano |
2b45e8 |
fadd f0, f0, f1
|
|
kusano |
2b45e8 |
fadd f2, f2, f3
|
|
kusano |
2b45e8 |
fadd f4, f4, f5
|
|
kusano |
2b45e8 |
fadd f6, f6, f7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fadd f0, f0, f2
|
|
kusano |
2b45e8 |
fadd f4, f4, f6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fadd f0, f0, f4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fsqrt f0, f0
|
|
kusano |
2b45e8 |
fmul f1, f31, f0
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(9999):
|
|
kusano |
2b45e8 |
lfd f14, 0(SP)
|
|
kusano |
2b45e8 |
lfd f15, 8(SP)
|
|
kusano |
2b45e8 |
lfd f16, 16(SP)
|
|
kusano |
2b45e8 |
lfd f17, 24(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
lfd f18, 32(SP)
|
|
kusano |
2b45e8 |
lfd f19, 40(SP)
|
|
kusano |
2b45e8 |
lfd f20, 48(SP)
|
|
kusano |
2b45e8 |
lfd f21, 56(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
lfd f22, 64(SP)
|
|
kusano |
2b45e8 |
lfd f23, 72(SP)
|
|
kusano |
2b45e8 |
lfd f24, 80(SP)
|
|
kusano |
2b45e8 |
lfd f25, 88(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
lfd f26, 96(SP)
|
|
kusano |
2b45e8 |
lfd f27, 104(SP)
|
|
kusano |
2b45e8 |
lfd f28, 112(SP)
|
|
kusano |
2b45e8 |
lfd f29, 120(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
lfd f30, 128(SP)
|
|
kusano |
2b45e8 |
lfd f31, 136(SP)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addi SP, SP, STACKSIZE
|
|
kusano |
2b45e8 |
blr
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
EPILOGUE
|