|
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 INCXM1 r9
|
|
kusano |
2b45e8 |
#define PREX r8
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define FZERO f0
|
|
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 r0, 0
|
|
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 r0, 144(SP)
|
|
kusano |
2b45e8 |
lfs FZERO,144(SP)
|
|
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 |
fmr f1, FZERO
|
|
kusano |
2b45e8 |
slwi INCX, INCX, ZBASE_SHIFT
|
|
kusano |
2b45e8 |
fmr f2, FZERO
|
|
kusano |
2b45e8 |
fmr f3, FZERO
|
|
kusano |
2b45e8 |
subi INCXM1, INCX, SIZE
|
|
kusano |
2b45e8 |
fmr f4, FZERO
|
|
kusano |
2b45e8 |
sub X, X, INCXM1
|
|
kusano |
2b45e8 |
fmr f5, FZERO
|
|
kusano |
2b45e8 |
li PREX, 3 * 16 * SIZE
|
|
kusano |
2b45e8 |
fmr f6, FZERO
|
|
kusano |
2b45e8 |
cmpwi cr0, N, 0
|
|
kusano |
2b45e8 |
fmr f7, FZERO
|
|
kusano |
2b45e8 |
ble- LL(999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
cmpwi cr0, INCX, 0
|
|
kusano |
2b45e8 |
ble- LL(999)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
srawi. r0, N, 3
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq- LL(150)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDX f8, X, INCXM1
|
|
kusano |
2b45e8 |
LFDUX f9, X, INCX
|
|
kusano |
2b45e8 |
LFDX f10, X, INCXM1
|
|
kusano |
2b45e8 |
LFDUX f11, X, INCX
|
|
kusano |
2b45e8 |
LFDX f12, X, INCXM1
|
|
kusano |
2b45e8 |
LFDUX f13, X, INCX
|
|
kusano |
2b45e8 |
LFDX f14, X, INCXM1
|
|
kusano |
2b45e8 |
LFDUX f15, X, INCX
|
|
kusano |
2b45e8 |
fabs f16, f8
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDX f24, X, INCXM1
|
|
kusano |
2b45e8 |
fabs f17, f9
|
|
kusano |
2b45e8 |
LFDUX f25, X, INCX
|
|
kusano |
2b45e8 |
fabs f18, f10
|
|
kusano |
2b45e8 |
LFDX f26, X, INCXM1
|
|
kusano |
2b45e8 |
fabs f19, f11
|
|
kusano |
2b45e8 |
LFDUX f27, X, INCX
|
|
kusano |
2b45e8 |
fabs f20, f12
|
|
kusano |
2b45e8 |
LFDX f28, X, INCXM1
|
|
kusano |
2b45e8 |
fabs f21, f13
|
|
kusano |
2b45e8 |
LFDUX f29, X, INCX
|
|
kusano |
2b45e8 |
fabs f22, f14
|
|
kusano |
2b45e8 |
LFDX f30, X, INCXM1
|
|
kusano |
2b45e8 |
fabs f23, f15
|
|
kusano |
2b45e8 |
LFDUX f31, X, INCX
|
|
kusano |
2b45e8 |
bdz LL(120)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(110):
|
|
kusano |
2b45e8 |
LFDX f8, X, INCXM1
|
|
kusano |
2b45e8 |
FADD f0, f0, f16
|
|
kusano |
2b45e8 |
#ifdef PPCG4
|
|
kusano |
2b45e8 |
dcbt X, PREX
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
fabs f16, f24
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f9, X, INCX
|
|
kusano |
2b45e8 |
FADD f1, f1, f17
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f17, f25
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDX f10, X, INCXM1
|
|
kusano |
2b45e8 |
FADD f2, f2, f18
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f18, f26
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f11, X, INCX
|
|
kusano |
2b45e8 |
FADD f3, f3, f19
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f19, f27
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDX f12, X, INCXM1
|
|
kusano |
2b45e8 |
FADD f4, f4, f20
|
|
kusano |
2b45e8 |
#if defined(PPCG4) && defined(DOUBLE)
|
|
kusano |
2b45e8 |
dcbt X, PREX
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
fabs f20, f28
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f13, X, INCX
|
|
kusano |
2b45e8 |
FADD f5, f5, f21
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f21, f29
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDX f14, X, INCXM1
|
|
kusano |
2b45e8 |
FADD f6, f6, f22
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f22, f30
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f15, X, INCX
|
|
kusano |
2b45e8 |
FADD f7, f7, f23
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f23, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDX f24, X, INCXM1
|
|
kusano |
2b45e8 |
FADD f0, f0, f16
|
|
kusano |
2b45e8 |
#ifdef PPCG4
|
|
kusano |
2b45e8 |
dcbt X, PREX
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
fabs f16, f8
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f25, X, INCX
|
|
kusano |
2b45e8 |
FADD f1, f1, f17
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f17, f9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDX f26, X, INCXM1
|
|
kusano |
2b45e8 |
FADD f2, f2, f18
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f18, f10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f27, X, INCX
|
|
kusano |
2b45e8 |
FADD f3, f3, f19
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f19, f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDX f28, X, INCXM1
|
|
kusano |
2b45e8 |
FADD f4, f4, f20
|
|
kusano |
2b45e8 |
#if defined(PPCG4) && defined(DOUBLE)
|
|
kusano |
2b45e8 |
dcbt X, PREX
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
fabs f20, f12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f29, X, INCX
|
|
kusano |
2b45e8 |
FADD f5, f5, f21
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f21, f13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDX f30, X, INCXM1
|
|
kusano |
2b45e8 |
FADD f6, f6, f22
|
|
kusano |
2b45e8 |
nop
|
|
kusano |
2b45e8 |
fabs f22, f14
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LFDUX f31, X, INCX
|
|
kusano |
2b45e8 |
FADD f7, f7, f23
|
|
kusano |
2b45e8 |
fabs f23, f15
|
|
kusano |
2b45e8 |
bdnz LL(110)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(120):
|
|
kusano |
2b45e8 |
FADD f0, f0, f16
|
|
kusano |
2b45e8 |
fabs f16, f24
|
|
kusano |
2b45e8 |
FADD f1, f1, f17
|
|
kusano |
2b45e8 |
fabs f17, f25
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FADD f2, f2, f18
|
|
kusano |
2b45e8 |
fabs f18, f26
|
|
kusano |
2b45e8 |
FADD f3, f3, f19
|
|
kusano |
2b45e8 |
fabs f19, f27
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FADD f4, f4, f20
|
|
kusano |
2b45e8 |
fabs f20, f28
|
|
kusano |
2b45e8 |
FADD f5, f5, f21
|
|
kusano |
2b45e8 |
fabs f21, f29
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FADD f6, f6, f22
|
|
kusano |
2b45e8 |
fabs f22, f30
|
|
kusano |
2b45e8 |
FADD f7, f7, f23
|
|
kusano |
2b45e8 |
fabs f23, f31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FADD f0, f0, f16
|
|
kusano |
2b45e8 |
FADD f1, f1, f17
|
|
kusano |
2b45e8 |
FADD f2, f2, f18
|
|
kusano |
2b45e8 |
FADD f3, f3, f19
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FADD f4, f4, f20
|
|
kusano |
2b45e8 |
FADD f5, f5, f21
|
|
kusano |
2b45e8 |
FADD f6, f6, f22
|
|
kusano |
2b45e8 |
FADD f7, f7, f23
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(150):
|
|
kusano |
2b45e8 |
andi. r0, N, 7
|
|
kusano |
2b45e8 |
mtspr CTR, r0
|
|
kusano |
2b45e8 |
beq LL(999)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(160):
|
|
kusano |
2b45e8 |
LFDX f8, X, INCXM1
|
|
kusano |
2b45e8 |
LFDUX f9, X, INCX
|
|
kusano |
2b45e8 |
fabs f8, f8
|
|
kusano |
2b45e8 |
fabs f9, f9
|
|
kusano |
2b45e8 |
FADD f0, f0, f8
|
|
kusano |
2b45e8 |
FADD f1, f1, f9
|
|
kusano |
2b45e8 |
bdnz LL(160)
|
|
kusano |
2b45e8 |
.align 4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
LL(999):
|
|
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 |
FADD f1, f0, f4
|
|
kusano |
2b45e8 |
|
|
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
|