|
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 |
#ifdef ATOM
|
|
kusano |
2b45e8 |
#define PREFETCH prefetchnta
|
|
kusano |
2b45e8 |
#define PREFETCHW prefetcht0
|
|
kusano |
2b45e8 |
#define PREFETCHSIZE (8 * 6)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACKSIZE 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define M 4 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
#define N 8 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
#define ALPHA_R 16 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
#define ALPHA_I 24 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
#define A 32 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
#define STACK_LDA 36 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
#define STACK_X 40 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
#define STACK_INCX 44 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
#define Y 48 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
#define STACK_INCY 52 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
#define BUFFER 56 + STACKSIZE(%esp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define I %eax
|
|
kusano |
2b45e8 |
#define J %ebx
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define INCX %ecx
|
|
kusano |
2b45e8 |
#define INCY J
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define A1 %esi
|
|
kusano |
2b45e8 |
#define X %edx
|
|
kusano |
2b45e8 |
#define Y1 %edi
|
|
kusano |
2b45e8 |
#define LDA %ebp
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if !defined(CONJ) && !defined(XCONJ)
|
|
kusano |
2b45e8 |
#define ADD1 addsd
|
|
kusano |
2b45e8 |
#define ADD2 addsd
|
|
kusano |
2b45e8 |
#define ADD3 subsd
|
|
kusano |
2b45e8 |
#define ADD4 addsd
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(CONJ) && !defined(XCONJ)
|
|
kusano |
2b45e8 |
#define ADD1 addsd
|
|
kusano |
2b45e8 |
#define ADD2 addsd
|
|
kusano |
2b45e8 |
#define ADD3 addsd
|
|
kusano |
2b45e8 |
#define ADD4 subsd
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if !defined(CONJ) && defined(XCONJ)
|
|
kusano |
2b45e8 |
#define ADD1 addsd
|
|
kusano |
2b45e8 |
#define ADD2 subsd
|
|
kusano |
2b45e8 |
#define ADD3 addsd
|
|
kusano |
2b45e8 |
#define ADD4 addsd
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(CONJ) && defined(XCONJ)
|
|
kusano |
2b45e8 |
#define ADD1 addsd
|
|
kusano |
2b45e8 |
#define ADD2 subsd
|
|
kusano |
2b45e8 |
#define ADD3 subsd
|
|
kusano |
2b45e8 |
#define ADD4 subsd
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROLOGUE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pushl %ebp
|
|
kusano |
2b45e8 |
pushl %edi
|
|
kusano |
2b45e8 |
pushl %esi
|
|
kusano |
2b45e8 |
pushl %ebx
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROFCODE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl STACK_LDA, LDA
|
|
kusano |
2b45e8 |
movl STACK_X, X
|
|
kusano |
2b45e8 |
movl STACK_INCX, INCX
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, INCX
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, LDA
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, A
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
cmpl $0, N
|
|
kusano |
2b45e8 |
jle .L999
|
|
kusano |
2b45e8 |
cmpl $0, M
|
|
kusano |
2b45e8 |
jle .L999
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl BUFFER, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl N, J
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pxor %xmm7, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl M, %eax
|
|
kusano |
2b45e8 |
addl $8, %eax
|
|
kusano |
2b45e8 |
sarl $3, %eax
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L01:
|
|
kusano |
2b45e8 |
movapd %xmm7, 0 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movapd %xmm7, 2 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movapd %xmm7, 4 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movapd %xmm7, 6 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movapd %xmm7, 8 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movapd %xmm7, 10 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movapd %xmm7, 12 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movapd %xmm7, 14 * SIZE(Y1)
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, Y1
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L01
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L10:
|
|
kusano |
2b45e8 |
movl BUFFER, Y1
|
|
kusano |
2b45e8 |
addl $16 * SIZE, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl A, A1
|
|
kusano |
2b45e8 |
addl LDA, A
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd 0 * SIZE(X), %xmm6
|
|
kusano |
2b45e8 |
movsd 1 * SIZE(X), %xmm7
|
|
kusano |
2b45e8 |
addl INCX, X
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
mulsd ALPHA_R, %xmm6
|
|
kusano |
2b45e8 |
mulsd ALPHA_I, %xmm2
|
|
kusano |
2b45e8 |
movapd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
mulsd ALPHA_I, %xmm3
|
|
kusano |
2b45e8 |
mulsd ALPHA_R, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef XCONJ
|
|
kusano |
2b45e8 |
subsd %xmm3, %xmm6
|
|
kusano |
2b45e8 |
addsd %xmm2, %xmm7
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addsd %xmm3, %xmm6
|
|
kusano |
2b45e8 |
subsd %xmm2, %xmm7
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl M, I
|
|
kusano |
2b45e8 |
sarl $2, I
|
|
kusano |
2b45e8 |
jle .L15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decl I
|
|
kusano |
2b45e8 |
jle .L14
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L13:
|
|
kusano |
2b45e8 |
#ifdef PREFETCH
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(A1)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -16 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movlpd %xmm1, -15 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
movsd -11 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -14 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movlpd %xmm1, -13 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -11 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
movsd -10 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
movsd -9 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -12 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -10 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movlpd %xmm1, -11 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -9 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
movsd -8 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
movsd -7 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -10 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -8 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movlpd %xmm1, -9 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -7 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-8 * SIZE, A1
|
|
kusano |
2b45e8 |
subl $-8 * SIZE, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $1, I
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jg .L13
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L14:
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -16 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movlpd %xmm1, -15 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
movsd -11 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -14 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movlpd %xmm1, -13 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -11 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
movsd -10 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
movsd -9 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -12 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -10 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movlpd %xmm1, -11 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -9 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -10 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -8 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movlpd %xmm1, -9 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -7 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-8 * SIZE, A1
|
|
kusano |
2b45e8 |
subl $-8 * SIZE, Y1
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L15:
|
|
kusano |
2b45e8 |
testl $2, M
|
|
kusano |
2b45e8 |
je .L17
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -16 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movlpd %xmm1, -15 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -14 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movlpd %xmm1, -13 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movsd -11 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $4 * SIZE, A1
|
|
kusano |
2b45e8 |
addl $4 * SIZE, Y1
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L17:
|
|
kusano |
2b45e8 |
testl $1, M
|
|
kusano |
2b45e8 |
je .L19
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(A1), %xmm2
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(A1), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm2, %xmm0
|
|
kusano |
2b45e8 |
mulsd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
ADD2 %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD3 %xmm3, %xmm0
|
|
kusano |
2b45e8 |
ADD4 %xmm5, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, -16 * SIZE(Y1)
|
|
kusano |
2b45e8 |
movlpd %xmm1, -15 * SIZE(Y1)
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L19:
|
|
kusano |
2b45e8 |
decl J
|
|
kusano |
2b45e8 |
jg .L10
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L990:
|
|
kusano |
2b45e8 |
movl Y, Y1
|
|
kusano |
2b45e8 |
movl BUFFER, X
|
|
kusano |
2b45e8 |
movl STACK_INCY, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl Y1, A1
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, INCY
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl M, %eax
|
|
kusano |
2b45e8 |
sarl $2, %eax
|
|
kusano |
2b45e8 |
jle .L994
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L992:
|
|
kusano |
2b45e8 |
movsd 0 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movsd 1 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
addl INCY, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd 0 * SIZE(Y1), %xmm2
|
|
kusano |
2b45e8 |
movsd 1 * SIZE(Y1), %xmm3
|
|
kusano |
2b45e8 |
addl INCY, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd 0 * SIZE(Y1), %xmm4
|
|
kusano |
2b45e8 |
movsd 1 * SIZE(Y1), %xmm5
|
|
kusano |
2b45e8 |
addl INCY, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd 0 * SIZE(Y1), %xmm6
|
|
kusano |
2b45e8 |
movsd 1 * SIZE(Y1), %xmm7
|
|
kusano |
2b45e8 |
addl INCY, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsd 0 * SIZE(X), %xmm0
|
|
kusano |
2b45e8 |
addsd 1 * SIZE(X), %xmm1
|
|
kusano |
2b45e8 |
addsd 2 * SIZE(X), %xmm2
|
|
kusano |
2b45e8 |
addsd 3 * SIZE(X), %xmm3
|
|
kusano |
2b45e8 |
addsd 4 * SIZE(X), %xmm4
|
|
kusano |
2b45e8 |
addsd 5 * SIZE(X), %xmm5
|
|
kusano |
2b45e8 |
addsd 6 * SIZE(X), %xmm6
|
|
kusano |
2b45e8 |
addsd 7 * SIZE(X), %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, 0 * SIZE(A1)
|
|
kusano |
2b45e8 |
movlpd %xmm1, 1 * SIZE(A1)
|
|
kusano |
2b45e8 |
addl INCY, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm2, 0 * SIZE(A1)
|
|
kusano |
2b45e8 |
movlpd %xmm3, 1 * SIZE(A1)
|
|
kusano |
2b45e8 |
addl INCY, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm4, 0 * SIZE(A1)
|
|
kusano |
2b45e8 |
movlpd %xmm5, 1 * SIZE(A1)
|
|
kusano |
2b45e8 |
addl INCY, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm6, 0 * SIZE(A1)
|
|
kusano |
2b45e8 |
movlpd %xmm7, 1 * SIZE(A1)
|
|
kusano |
2b45e8 |
addl INCY, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $8 * SIZE, X
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L992
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L994:
|
|
kusano |
2b45e8 |
testl $2, M
|
|
kusano |
2b45e8 |
jle .L996
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd 0 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movsd 1 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
addl INCY, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd 0 * SIZE(Y1), %xmm2
|
|
kusano |
2b45e8 |
movsd 1 * SIZE(Y1), %xmm3
|
|
kusano |
2b45e8 |
addl INCY, Y1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsd 0 * SIZE(X), %xmm0
|
|
kusano |
2b45e8 |
addsd 1 * SIZE(X), %xmm1
|
|
kusano |
2b45e8 |
addsd 2 * SIZE(X), %xmm2
|
|
kusano |
2b45e8 |
addsd 3 * SIZE(X), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, 0 * SIZE(A1)
|
|
kusano |
2b45e8 |
movlpd %xmm1, 1 * SIZE(A1)
|
|
kusano |
2b45e8 |
addl INCY, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm2, 0 * SIZE(A1)
|
|
kusano |
2b45e8 |
movlpd %xmm3, 1 * SIZE(A1)
|
|
kusano |
2b45e8 |
addl INCY, A1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $4 * SIZE, X
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L996:
|
|
kusano |
2b45e8 |
testl $1, M
|
|
kusano |
2b45e8 |
jle .L999
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd 0 * SIZE(Y1), %xmm0
|
|
kusano |
2b45e8 |
movsd 1 * SIZE(Y1), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsd 0 * SIZE(X), %xmm0
|
|
kusano |
2b45e8 |
addsd 1 * SIZE(X), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, 0 * SIZE(A1)
|
|
kusano |
2b45e8 |
movlpd %xmm1, 1 * SIZE(A1)
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L999:
|
|
kusano |
2b45e8 |
popl %ebx
|
|
kusano |
2b45e8 |
popl %esi
|
|
kusano |
2b45e8 |
popl %edi
|
|
kusano |
2b45e8 |
popl %ebp
|
|
kusano |
2b45e8 |
ret
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
EPILOGUE
|