|
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 STACK 12
|
|
kusano |
2b45e8 |
#define ARGS 0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define M 4 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define N 8 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#ifdef DOUBLE
|
|
kusano |
2b45e8 |
#define BETA_R 16 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define BETA_I 24 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define C 48 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define LDC 52 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define BETA_R 16 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define BETA_I 20 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define C 40 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define LDC 44 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROLOGUE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pushl %ebp
|
|
kusano |
2b45e8 |
pushl %edi
|
|
kusano |
2b45e8 |
pushl %esi
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROFCODE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl M, %ebp
|
|
kusano |
2b45e8 |
movl N, %ecx
|
|
kusano |
2b45e8 |
movl LDC, %edx
|
|
kusano |
2b45e8 |
movl C, %edi
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FLD BETA_R
|
|
kusano |
2b45e8 |
FLD BETA_I
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
testl %ebp, %ebp # if n <= 0 goto End
|
|
kusano |
2b45e8 |
jle .L83
|
|
kusano |
2b45e8 |
testl %ecx, %ecx # if m <= 0 goto End
|
|
kusano |
2b45e8 |
jle .L83
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fld %st(1)
|
|
kusano |
2b45e8 |
fabs
|
|
kusano |
2b45e8 |
fld %st(1)
|
|
kusano |
2b45e8 |
fabs
|
|
kusano |
2b45e8 |
faddp %st, %st(1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, %edx
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ftst
|
|
kusano |
2b45e8 |
fnstsw %ax
|
|
kusano |
2b45e8 |
andb $68, %ah
|
|
kusano |
2b45e8 |
#ifndef C_SUN
|
|
kusano |
2b45e8 |
ffreep %st(0)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
.byte 0xdf
|
|
kusano |
2b45e8 |
.byte 0xc0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
je .L71
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L53:
|
|
kusano |
2b45e8 |
movl %edi, %esi # c_offset1 = c_offset
|
|
kusano |
2b45e8 |
addl %edx, %edi # c_offset += ldc
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl %ebp, %eax
|
|
kusano |
2b45e8 |
sarl $2, %eax
|
|
kusano |
2b45e8 |
jle .L56
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L57:
|
|
kusano |
2b45e8 |
#if defined(HAS_PREFETCH) && defined(PENTIUM3)
|
|
kusano |
2b45e8 |
prefetchnta 16 * SIZE(%esi)
|
|
kusano |
2b45e8 |
prefetchnta 24 * SIZE(%esi)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FSTU 0 * SIZE(%esi) # c_offset1
|
|
kusano |
2b45e8 |
FSTU 1 * SIZE(%esi)
|
|
kusano |
2b45e8 |
FSTU 2 * SIZE(%esi)
|
|
kusano |
2b45e8 |
FSTU 3 * SIZE(%esi)
|
|
kusano |
2b45e8 |
FSTU 4 * SIZE(%esi)
|
|
kusano |
2b45e8 |
FSTU 5 * SIZE(%esi)
|
|
kusano |
2b45e8 |
FSTU 6 * SIZE(%esi)
|
|
kusano |
2b45e8 |
FSTU 7 * SIZE(%esi)
|
|
kusano |
2b45e8 |
addl $8 * SIZE, %esi # c_offset1 += 8
|
|
kusano |
2b45e8 |
decl %eax # i--
|
|
kusano |
2b45e8 |
jg .L57
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L56:
|
|
kusano |
2b45e8 |
movl %ebp, %eax
|
|
kusano |
2b45e8 |
andl $3, %eax
|
|
kusano |
2b45e8 |
jle .L62
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L63:
|
|
kusano |
2b45e8 |
FSTU 0 * SIZE(%esi)
|
|
kusano |
2b45e8 |
FSTU 1 * SIZE(%esi)
|
|
kusano |
2b45e8 |
addl $2 * SIZE,%esi
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L63
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L62:
|
|
kusano |
2b45e8 |
decl %ecx # j --
|
|
kusano |
2b45e8 |
jg .L53
|
|
kusano |
2b45e8 |
jmp .L83
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L71:
|
|
kusano |
2b45e8 |
movl %edi, %esi
|
|
kusano |
2b45e8 |
addl %edx, %edi # c_offset += ldc
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl %ebp, %eax
|
|
kusano |
2b45e8 |
sarl $1, %eax
|
|
kusano |
2b45e8 |
jle .L84
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L85:
|
|
kusano |
2b45e8 |
#if defined(HAS_PREFETCH) && defined(PENTIUM3)
|
|
kusano |
2b45e8 |
prefetchnta 16 * SIZE(%esi)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
fld %st(0)
|
|
kusano |
2b45e8 |
FMUL 0 * SIZE(%esi)
|
|
kusano |
2b45e8 |
fld %st(2)
|
|
kusano |
2b45e8 |
FMUL 1 * SIZE(%esi)
|
|
kusano |
2b45e8 |
faddp %st,%st(1)
|
|
kusano |
2b45e8 |
fld %st(2)
|
|
kusano |
2b45e8 |
FMUL 0 * SIZE(%esi)
|
|
kusano |
2b45e8 |
fld %st(2)
|
|
kusano |
2b45e8 |
FMUL 1 * SIZE(%esi)
|
|
kusano |
2b45e8 |
fsubrp %st,%st(1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FST 0 * SIZE(%esi)
|
|
kusano |
2b45e8 |
FST 1 * SIZE(%esi)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fld %st(0)
|
|
kusano |
2b45e8 |
FMUL 2 * SIZE(%esi)
|
|
kusano |
2b45e8 |
fld %st(2)
|
|
kusano |
2b45e8 |
FMUL 3 * SIZE(%esi)
|
|
kusano |
2b45e8 |
faddp %st,%st(1)
|
|
kusano |
2b45e8 |
fld %st(2)
|
|
kusano |
2b45e8 |
FMUL 2 * SIZE(%esi)
|
|
kusano |
2b45e8 |
fld %st(2)
|
|
kusano |
2b45e8 |
FMUL 3 * SIZE(%esi)
|
|
kusano |
2b45e8 |
fsubrp %st,%st(1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FST 2 * SIZE(%esi)
|
|
kusano |
2b45e8 |
FST 3 * SIZE(%esi)
|
|
kusano |
2b45e8 |
addl $4 * SIZE, %esi
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L85
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L84:
|
|
kusano |
2b45e8 |
movl %ebp, %eax
|
|
kusano |
2b45e8 |
andl $1, %eax
|
|
kusano |
2b45e8 |
jle .L74
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L75:
|
|
kusano |
2b45e8 |
#if defined(HAS_PREFETCH) && defined(PENTIUM3)
|
|
kusano |
2b45e8 |
prefetchnta 16 * SIZE(%esi)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
fld %st(0)
|
|
kusano |
2b45e8 |
FMUL 0 * SIZE(%esi)
|
|
kusano |
2b45e8 |
fld %st(2)
|
|
kusano |
2b45e8 |
FMUL 1 * SIZE(%esi)
|
|
kusano |
2b45e8 |
faddp %st,%st(1)
|
|
kusano |
2b45e8 |
fld %st(2)
|
|
kusano |
2b45e8 |
FMUL 0 * SIZE(%esi)
|
|
kusano |
2b45e8 |
fld %st(2)
|
|
kusano |
2b45e8 |
FMUL 1 * SIZE(%esi)
|
|
kusano |
2b45e8 |
fsubrp %st,%st(1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
FST 0 * SIZE(%esi)
|
|
kusano |
2b45e8 |
FST 1 * SIZE(%esi)
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L74:
|
|
kusano |
2b45e8 |
decl %ecx
|
|
kusano |
2b45e8 |
jg .L71
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L83:
|
|
kusano |
2b45e8 |
#ifndef C_SUN
|
|
kusano |
2b45e8 |
ffreep %st(0)
|
|
kusano |
2b45e8 |
ffreep %st(0)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
.byte 0xdf
|
|
kusano |
2b45e8 |
.byte 0xc0
|
|
kusano |
2b45e8 |
.byte 0xdf
|
|
kusano |
2b45e8 |
.byte 0xc0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
popl %esi
|
|
kusano |
2b45e8 |
popl %edi
|
|
kusano |
2b45e8 |
popl %ebp
|
|
kusano |
2b45e8 |
ret
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
EPILOGUE
|