|
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 |
#if !defined(HAVE_SSE2) || !defined(HAVE_MMX)
|
|
kusano |
2b45e8 |
#error You have to check your configuration.
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACK 16
|
|
kusano |
2b45e8 |
#define ARGS 0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACK_M 4 + STACK + ARGS(%esi)
|
|
kusano |
2b45e8 |
#define STACK_N 8 + STACK + ARGS(%esi)
|
|
kusano |
2b45e8 |
#define STACK_K 12 + STACK + ARGS(%esi)
|
|
kusano |
2b45e8 |
#define STACK_ALPHA 16 + STACK + ARGS(%esi)
|
|
kusano |
2b45e8 |
#define STACK_A 24 + STACK + ARGS(%esi)
|
|
kusano |
2b45e8 |
#define STACK_B 28 + STACK + ARGS(%esi)
|
|
kusano |
2b45e8 |
#define STACK_C 32 + STACK + ARGS(%esi)
|
|
kusano |
2b45e8 |
#define STACK_LDC 36 + STACK + ARGS(%esi)
|
|
kusano |
2b45e8 |
#define STACK_OFFT 40 + STACK + ARGS(%esi)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define K 16(%esp)
|
|
kusano |
2b45e8 |
#define N 20(%esp)
|
|
kusano |
2b45e8 |
#define M 24(%esp)
|
|
kusano |
2b45e8 |
#define A 28(%esp)
|
|
kusano |
2b45e8 |
#define C 32(%esp)
|
|
kusano |
2b45e8 |
#define J 36(%esp)
|
|
kusano |
2b45e8 |
#define OLD_STACK 40(%esp)
|
|
kusano |
2b45e8 |
#define OFFSET 44(%esp)
|
|
kusano |
2b45e8 |
#define KK 48(%esp)
|
|
kusano |
2b45e8 |
#define KKK 52(%esp)
|
|
kusano |
2b45e8 |
#define AORIG 56(%esp)
|
|
kusano |
2b45e8 |
#define BORIG 60(%esp)
|
|
kusano |
2b45e8 |
#define BUFFER 128(%esp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACK_ALIGN 4096
|
|
kusano |
2b45e8 |
#define STACK_OFFSET 1024
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define B %edi
|
|
kusano |
2b45e8 |
#define AA %edx
|
|
kusano |
2b45e8 |
#define BB %ecx
|
|
kusano |
2b45e8 |
#define LDC %ebp
|
|
kusano |
2b45e8 |
#define CO1 %esi
|
|
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 |
EMMS
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl %esp, %esi # save old stack
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $128 + LOCAL_BUFFER_SIZE + STACK_OFFSET, %esp
|
|
kusano |
2b45e8 |
andl $-STACK_ALIGN, %esp
|
|
kusano |
2b45e8 |
addl $STACK_OFFSET, %esp
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STACK_TOUCHING
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movd STACK_M, %mm0
|
|
kusano |
2b45e8 |
movl STACK_N, %eax
|
|
kusano |
2b45e8 |
movd STACK_K, %mm1
|
|
kusano |
2b45e8 |
movd STACK_A, %mm2
|
|
kusano |
2b45e8 |
movl STACK_B, B
|
|
kusano |
2b45e8 |
movd STACK_C, %mm3
|
|
kusano |
2b45e8 |
movl STACK_LDC, LDC
|
|
kusano |
2b45e8 |
movd STACK_OFFT, %mm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movd %mm1, K
|
|
kusano |
2b45e8 |
movl %eax, N
|
|
kusano |
2b45e8 |
movd %mm0, M
|
|
kusano |
2b45e8 |
movd %mm2, A
|
|
kusano |
2b45e8 |
movd %mm3, C
|
|
kusano |
2b45e8 |
movl %esi, OLD_STACK
|
|
kusano |
2b45e8 |
movd %mm4, OFFSET
|
|
kusano |
2b45e8 |
movd %mm4, KK
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, A
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, B
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
sall $BASE_SHIFT, LDC
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl M, %eax
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
addl %eax, C
|
|
kusano |
2b45e8 |
imull K, %eax
|
|
kusano |
2b45e8 |
addl %eax, A
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movl N, %eax
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
imull K, %eax
|
|
kusano |
2b45e8 |
addl %eax, B
|
|
kusano |
2b45e8 |
movl N, %eax
|
|
kusano |
2b45e8 |
imull LDC, %eax
|
|
kusano |
2b45e8 |
addl %eax, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
negl KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movl N, %eax
|
|
kusano |
2b45e8 |
subl OFFSET, %eax
|
|
kusano |
2b45e8 |
movl %eax, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl N, %eax
|
|
kusano |
2b45e8 |
sarl $1, %eax
|
|
kusano |
2b45e8 |
movl %eax, J
|
|
kusano |
2b45e8 |
jle .L100
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L01:
|
|
kusano |
2b45e8 |
/* Copying to Sub Buffer */
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl OFFSET, %eax
|
|
kusano |
2b45e8 |
addl M, %eax
|
|
kusano |
2b45e8 |
movl %eax, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $1 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
subl %eax, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
movl B, BORIG
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (B, %eax, 2), B
|
|
kusano |
2b45e8 |
leal (BB, %eax, 4), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movl OFFSET, %eax
|
|
kusano |
2b45e8 |
movl %eax, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarl $2, %eax
|
|
kusano |
2b45e8 |
jle .L03
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L02:
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
movddup -15 * SIZE(B), %xmm1
|
|
kusano |
2b45e8 |
movddup -14 * SIZE(B), %xmm2
|
|
kusano |
2b45e8 |
movddup -13 * SIZE(B), %xmm3
|
|
kusano |
2b45e8 |
movddup -12 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
movddup -11 * SIZE(B), %xmm5
|
|
kusano |
2b45e8 |
movddup -10 * SIZE(B), %xmm6
|
|
kusano |
2b45e8 |
movddup -9 * SIZE(B), %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm1, -14 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm2, -12 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm3, -10 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm4, -8 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm5, -6 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm6, -4 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm7, -2 * SIZE(BB)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $ 8 * SIZE, B
|
|
kusano |
2b45e8 |
addl $16 * SIZE, %ecx
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L02
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L03:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andl $3, %eax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L05
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L04:
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
movddup -15 * SIZE(B), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm1, -14 * SIZE(BB)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $2 * SIZE, B
|
|
kusano |
2b45e8 |
addl $4 * SIZE, %ecx
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L04
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L05:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl A, AA
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl A, %eax
|
|
kusano |
2b45e8 |
movl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal (, LDC, 2), %eax
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
subl %eax, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movl C, CO1 # coffset = c
|
|
kusano |
2b45e8 |
#ifndef RT
|
|
kusano |
2b45e8 |
addl %eax, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
testl $1, %ebx
|
|
kusano |
2b45e8 |
jle .L30
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $0 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
subl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 1), AA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
sall $1 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
pxor %xmm6, %xmm6
|
|
kusano |
2b45e8 |
movsd -8 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
pxor %xmm7, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarl $3, %eax
|
|
kusano |
2b45e8 |
je .L52
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L51:
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulsd -14 * SIZE(BB), %xmm0
|
|
kusano |
2b45e8 |
addsd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
addsd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulsd -10 * SIZE(BB), %xmm0
|
|
kusano |
2b45e8 |
addsd %xmm1, %xmm6
|
|
kusano |
2b45e8 |
movsd 0 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
addsd %xmm0, %xmm7
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulsd -6 * SIZE(BB), %xmm0
|
|
kusano |
2b45e8 |
addsd %xmm3, %xmm4
|
|
kusano |
2b45e8 |
movsd -4 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
addsd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulsd -2 * SIZE(BB), %xmm0
|
|
kusano |
2b45e8 |
addsd %xmm3, %xmm6
|
|
kusano |
2b45e8 |
movsd 8 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
addsd %xmm0, %xmm7
|
|
kusano |
2b45e8 |
movsd -8 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
mulsd 2 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
addsd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movsd 4 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
addsd %xmm2, %xmm5
|
|
kusano |
2b45e8 |
movsd -11 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
mulsd 6 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
addsd %xmm1, %xmm6
|
|
kusano |
2b45e8 |
movsd 16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
addsd %xmm2, %xmm7
|
|
kusano |
2b45e8 |
movsd -10 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulsd 10 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
addsd %xmm3, %xmm4
|
|
kusano |
2b45e8 |
movsd 12 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
addsd %xmm2, %xmm5
|
|
kusano |
2b45e8 |
movsd -9 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulsd 14 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
addsd %xmm3, %xmm6
|
|
kusano |
2b45e8 |
movsd 24 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
addsd %xmm2, %xmm7
|
|
kusano |
2b45e8 |
movsd -4 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-8 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $32 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L51
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L52:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andl $7, %eax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
je .L54
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L53:
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulsd -14 * SIZE(BB), %xmm0
|
|
kusano |
2b45e8 |
addsd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
addsd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $1 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $4 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jg .L53
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L54:
|
|
kusano |
2b45e8 |
addsd %xmm6, %xmm4
|
|
kusano |
2b45e8 |
addsd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $1, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
subl $2, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 1), AA
|
|
kusano |
2b45e8 |
leal (B, %eax, 2), B
|
|
kusano |
2b45e8 |
leal (BB, %eax, 4), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(B), %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subsd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
subsd %xmm5, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm0
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
mulsd -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm4
|
|
kusano |
2b45e8 |
subsd %xmm4, %xmm1
|
|
kusano |
2b45e8 |
mulsd -13 * SIZE(B), %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
mulsd -13 * SIZE(B), %xmm1
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
subsd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulsd -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $1 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm0, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm1, 0 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addl $1 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd %xmm0, -16 * SIZE(B)
|
|
kusano |
2b45e8 |
movsd %xmm1, -15 * SIZE(B)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm0, -16 * SIZE(BB)
|
|
kusano |
2b45e8 |
movsd %xmm0, -15 * SIZE(BB)
|
|
kusano |
2b45e8 |
movsd %xmm1, -14 * SIZE(BB)
|
|
kusano |
2b45e8 |
movsd %xmm1, -13 * SIZE(BB)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd %xmm0, -16 * SIZE(AA)
|
|
kusano |
2b45e8 |
movsd %xmm1, -15 * SIZE(AA)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
leal (,%eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 1), AA
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $2 * SIZE, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $1, KK
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
sall $0 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L30:
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
testl $2, %ebx
|
|
kusano |
2b45e8 |
jle .L50
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $1 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
subl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 2), AA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
sall $1 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
pxor %xmm6, %xmm6
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
pxor %xmm7, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarl $3, %eax
|
|
kusano |
2b45e8 |
je .L32
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L31:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(BB), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(BB), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm6
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm7
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulpd -6 * SIZE(BB), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm4
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE(BB), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm6
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm7
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
mulpd 2 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm5
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
mulpd 6 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm6
|
|
kusano |
2b45e8 |
movapd 16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm7
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulpd 10 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm4
|
|
kusano |
2b45e8 |
movapd 12 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm5
|
|
kusano |
2b45e8 |
movapd -2 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulpd 14 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm6
|
|
kusano |
2b45e8 |
movapd 24 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm7
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $ 32 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L32:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andl $7, %eax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
je .L34
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L33:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(BB), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $2 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $4 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L33
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L34:
|
|
kusano |
2b45e8 |
addpd %xmm6, %xmm4
|
|
kusano |
2b45e8 |
addpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $2, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
subl $2, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 2), AA
|
|
kusano |
2b45e8 |
leal (B, %eax, 2), B
|
|
kusano |
2b45e8 |
leal (BB, %eax, 4), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
unpcklpd %xmm5, %xmm4
|
|
kusano |
2b45e8 |
unpckhpd %xmm5, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(B), %xmm2
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(B), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm2
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm5, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movddup -13 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movddup -14 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movddup -15 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movddup -13 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
movddup -15 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm4
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm1
|
|
kusano |
2b45e8 |
movddup -13 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movddup -13 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm1
|
|
kusano |
2b45e8 |
movddup -14 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $2 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd %xmm2, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm3, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm2, 0 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
movhpd %xmm3, 1 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd %xmm0, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm0, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm1, 0 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
movhpd %xmm1, 1 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addl $2 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd %xmm2, -16 * SIZE(B)
|
|
kusano |
2b45e8 |
movapd %xmm3, -14 * SIZE(B)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup %xmm2, %xmm0
|
|
kusano |
2b45e8 |
movddup %xmm3, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
unpckhpd %xmm2, %xmm2
|
|
kusano |
2b45e8 |
unpckhpd %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm2, -14 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm1, -12 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm3, -10 * SIZE(BB)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(AA)
|
|
kusano |
2b45e8 |
movapd %xmm1, -14 * SIZE(AA)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
leal (,%eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 2), AA
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $4 * SIZE, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $2, KK
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
sall $1 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L50:
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
sarl $2, %ebx # i = (m >> 2)
|
|
kusano |
2b45e8 |
jle .L99
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L10:
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $2 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
subl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 4), AA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
sall $1 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
pxor %xmm6, %xmm6
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
prefetcht2 -3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
pxor %xmm7, %xmm7
|
|
kusano |
2b45e8 |
prefetcht2 -3 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
prefetcht2 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
pxor %xmm7, %xmm7
|
|
kusano |
2b45e8 |
prefetcht2 3 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarl $3, %eax
|
|
kusano |
2b45e8 |
je .L15
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L12:
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm6
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm6
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm6
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd -2 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
movapd -2 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm6
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm6
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd 6 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movapd 6 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movapd 16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm6
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd 10 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
movapd 10 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm6
|
|
kusano |
2b45e8 |
movapd 12 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd 12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd 14 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
movapd 14 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
subl $-32 * SIZE, BB
|
|
kusano |
2b45e8 |
movapd 24 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
subl $-32 * SIZE, AA
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm6
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L12
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L15:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andl $7, %eax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
je .L18
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L16:
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm6
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $4 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $4 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L16
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L18:
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $4, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
subl $2, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 4), AA
|
|
kusano |
2b45e8 |
leal (B, %eax, 2), B
|
|
kusano |
2b45e8 |
leal (BB, %eax, 4), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
unpcklpd %xmm5, %xmm4
|
|
kusano |
2b45e8 |
unpckhpd %xmm5, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm6, %xmm1
|
|
kusano |
2b45e8 |
unpcklpd %xmm7, %xmm6
|
|
kusano |
2b45e8 |
unpckhpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(B), %xmm2
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(B), %xmm3
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(B), %xmm5
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(B), %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm2
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
subpd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
subpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm6, %xmm1
|
|
kusano |
2b45e8 |
subpd %xmm5, %xmm2
|
|
kusano |
2b45e8 |
subpd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movddup -1 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm7
|
|
kusano |
2b45e8 |
movddup -2 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movddup -3 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movddup -4 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup -6 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movddup -7 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movddup -8 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup -11 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movddup -12 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movddup -15 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movddup -14 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movddup -13 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup -11 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movddup -10 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movddup -9 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup -6 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movddup -5 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup -1 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm7
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup -15 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
movapd %xmm4, %xmm5
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm4
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm5
|
|
kusano |
2b45e8 |
subpd %xmm5, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup -13 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movddup -13 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup -14 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
movapd %xmm4, %xmm5
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
subpd %xmm5, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $4 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd %xmm2, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm3, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm5, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm7, 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movhpd %xmm2, 0 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
movhpd %xmm3, 1 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
movhpd %xmm5, 2 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
movhpd %xmm7, 3 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd %xmm0, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm0, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm1, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm1, 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm2, 0 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
movhpd %xmm2, 1 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
movsd %xmm3, 2 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
movhpd %xmm3, 3 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addl $4 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd %xmm2, -16 * SIZE(B)
|
|
kusano |
2b45e8 |
movapd %xmm3, -14 * SIZE(B)
|
|
kusano |
2b45e8 |
movapd %xmm5, -12 * SIZE(B)
|
|
kusano |
2b45e8 |
movapd %xmm7, -10 * SIZE(B)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup %xmm2, %xmm0
|
|
kusano |
2b45e8 |
movddup %xmm3, %xmm1
|
|
kusano |
2b45e8 |
movddup %xmm5, %xmm4
|
|
kusano |
2b45e8 |
movddup %xmm7, %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
unpckhpd %xmm2, %xmm2
|
|
kusano |
2b45e8 |
unpckhpd %xmm3, %xmm3
|
|
kusano |
2b45e8 |
unpckhpd %xmm5, %xmm5
|
|
kusano |
2b45e8 |
unpckhpd %xmm7, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm2, -14 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm1, -12 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm3, -10 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm4, -8 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm5, -6 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm6, -4 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm7, -2 * SIZE(BB)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(AA)
|
|
kusano |
2b45e8 |
movapd %xmm1, -14 * SIZE(AA)
|
|
kusano |
2b45e8 |
movapd %xmm2, -12 * SIZE(AA)
|
|
kusano |
2b45e8 |
movapd %xmm3, -10 * SIZE(AA)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
leal (,%eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 4), AA
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $8 * SIZE, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $4, KK
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $4, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
sall $2 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decl %ebx # i --
|
|
kusano |
2b45e8 |
jg .L10
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L99:
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (B, %eax, 2), B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
leal (,%eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (B, %eax, 2), B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
addl $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
subl $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decl J # j --
|
|
kusano |
2b45e8 |
jg .L01
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L100:
|
|
kusano |
2b45e8 |
movl N, %eax
|
|
kusano |
2b45e8 |
testl $1, %eax
|
|
kusano |
2b45e8 |
jle .L999
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L101:
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl OFFSET, %eax
|
|
kusano |
2b45e8 |
addl M, %eax
|
|
kusano |
2b45e8 |
movl %eax, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $0 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
subl %eax, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
movl B, BORIG
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (B, %eax, 1), B
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movl OFFSET, %eax
|
|
kusano |
2b45e8 |
movl %eax, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarl $3, %eax
|
|
kusano |
2b45e8 |
jle .L103
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L102:
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
movddup -15 * SIZE(B), %xmm1
|
|
kusano |
2b45e8 |
movddup -14 * SIZE(B), %xmm2
|
|
kusano |
2b45e8 |
movddup -13 * SIZE(B), %xmm3
|
|
kusano |
2b45e8 |
movddup -12 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
movddup -11 * SIZE(B), %xmm5
|
|
kusano |
2b45e8 |
movddup -10 * SIZE(B), %xmm6
|
|
kusano |
2b45e8 |
movddup -9 * SIZE(B), %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm1, -14 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm2, -12 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm3, -10 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm4, -8 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm5, -6 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm6, -4 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm7, -2 * SIZE(BB)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $ 8 * SIZE, B
|
|
kusano |
2b45e8 |
addl $16 * SIZE, %ecx
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jne .L102
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L103:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andl $7, %eax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L105
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L104:
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(BB)
|
|
kusano |
2b45e8 |
addl $1 * SIZE, B
|
|
kusano |
2b45e8 |
addl $2 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L104
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L105:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl A, AA
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl A, %eax
|
|
kusano |
2b45e8 |
movl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
subl LDC, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movl C, CO1 # coffset = c
|
|
kusano |
2b45e8 |
#ifndef RT
|
|
kusano |
2b45e8 |
addl LDC, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
testl $1, %ebx
|
|
kusano |
2b45e8 |
jle .L130
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $0 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
subl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 1), AA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
sall $0 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
movsd -8 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarl $3, %eax
|
|
kusano |
2b45e8 |
je .L152
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L151:
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addsd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addsd %xmm1, %xmm5
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addsd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movsd -10 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movsd -8 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addsd %xmm1, %xmm5
|
|
kusano |
2b45e8 |
movsd -0 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
movsd -11 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addsd %xmm3, %xmm4
|
|
kusano |
2b45e8 |
movsd -6 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
movsd -10 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addsd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
movsd -4 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
movsd -9 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addsd %xmm3, %xmm4
|
|
kusano |
2b45e8 |
movsd -2 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
movsd -4 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addsd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
movsd 8 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $ -8 * SIZE, AA
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, BB
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L151
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L152:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andl $7, %eax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
je .L154
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L153:
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addsd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $1 * SIZE, AA # aoffset += 8
|
|
kusano |
2b45e8 |
addl $2 * SIZE, BB # boffset1 += 8
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jg .L153
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L154:
|
|
kusano |
2b45e8 |
addsd %xmm5, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
subl $1, %eax
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 1), AA
|
|
kusano |
2b45e8 |
leal (B, %eax, 1), B
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subsd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
mulsd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(RN) || defined(RT)
|
|
kusano |
2b45e8 |
mulsd -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $1 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm0, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addl $1 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd %xmm0, -16 * SIZE(B)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm0, -16 * SIZE(BB)
|
|
kusano |
2b45e8 |
movsd %xmm0, -15 * SIZE(BB)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd %xmm0, -16 * SIZE(AA)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
leal (,%eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 1), AA
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $1 * SIZE, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $1, KK
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
sall $0 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L130:
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
testl $2, %ebx
|
|
kusano |
2b45e8 |
jle .L150
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $1 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
subl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 2), AA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
sall $0 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarl $3, %eax
|
|
kusano |
2b45e8 |
je .L132
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L131:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm5
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm5
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm4
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
movapd -2 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm4
|
|
kusano |
2b45e8 |
movapd -2 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, AA
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, BB
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L131
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L132:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andl $7, %eax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
je .L134
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L133:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $2 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $2 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L133
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L134:
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $2, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
subl $1, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 2), AA
|
|
kusano |
2b45e8 |
leal (B, %eax, 1), B
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movapd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
unpckhpd %xmm2, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm5
|
|
kusano |
2b45e8 |
subsd %xmm5, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
unpcklpd %xmm2, %xmm0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movapd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
unpckhpd %xmm2, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
subsd %xmm5, %xmm2
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
unpcklpd %xmm2, %xmm0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(RN) || defined(RT)
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $2 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm0, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm0, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addl $2 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(B)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup %xmm0, %xmm1
|
|
kusano |
2b45e8 |
unpckhpd %xmm0, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm1, -16 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm0, -14 * SIZE(BB)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(AA)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
leal (,%eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 2), AA
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $2 * SIZE, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $2, KK
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
sall $1 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L150:
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
sarl $2, %ebx # i = (m >> 2)
|
|
kusano |
2b45e8 |
jle .L159
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L110:
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $2 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
subl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 4), AA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
sall $0 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
pxor %xmm6, %xmm6
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
pxor %xmm7, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarl $3, %eax
|
|
kusano |
2b45e8 |
je .L112
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L111:
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm4
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm6
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd -6 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm6
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm5
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm7
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm0
|
|
kusano |
2b45e8 |
mulpd 2 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm4
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm6
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm0
|
|
kusano |
2b45e8 |
mulpd 6 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
movapd 16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm7
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulpd 10 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm4
|
|
kusano |
2b45e8 |
movapd 12 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm6
|
|
kusano |
2b45e8 |
movapd -2 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulpd 14 * SIZE(AA), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm5
|
|
kusano |
2b45e8 |
movapd 24 * SIZE(AA), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm7
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $ 32 * SIZE, AA
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L111
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L112:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andl $7, %eax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
je .L114
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L113:
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm4
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm6
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $4 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $2 * SIZE, BB
|
|
kusano |
2b45e8 |
subl $1, %eax
|
|
kusano |
2b45e8 |
jg .L113
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L114:
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm4
|
|
kusano |
2b45e8 |
addpd %xmm7, %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $4, %eax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
subl $1, %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl AORIG, AA
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
leal 16 * SIZE + BUFFER, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 4), AA
|
|
kusano |
2b45e8 |
leal (B, %eax, 1), B
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(B), %xmm0
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(B), %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
subpd %xmm6, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movapd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
unpckhpd %xmm2, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm3
|
|
kusano |
2b45e8 |
unpckhpd %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -1 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -2 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm3, %xmm5
|
|
kusano |
2b45e8 |
subsd %xmm5, %xmm1
|
|
kusano |
2b45e8 |
movsd -3 * SIZE(AA), %xmm6
|
|
kusano |
2b45e8 |
mulsd %xmm3, %xmm6
|
|
kusano |
2b45e8 |
subsd %xmm6, %xmm2
|
|
kusano |
2b45e8 |
movsd -4 * SIZE(AA), %xmm7
|
|
kusano |
2b45e8 |
mulsd %xmm3, %xmm7
|
|
kusano |
2b45e8 |
subsd %xmm7, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -6 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -7 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm1, %xmm5
|
|
kusano |
2b45e8 |
subsd %xmm5, %xmm2
|
|
kusano |
2b45e8 |
movsd -8 * SIZE(AA), %xmm6
|
|
kusano |
2b45e8 |
mulsd %xmm1, %xmm6
|
|
kusano |
2b45e8 |
subsd %xmm6, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -11 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -12 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm5
|
|
kusano |
2b45e8 |
subsd %xmm5, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
unpcklpd %xmm2, %xmm0
|
|
kusano |
2b45e8 |
unpcklpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movapd %xmm0, %xmm2
|
|
kusano |
2b45e8 |
unpckhpd %xmm2, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm1, %xmm3
|
|
kusano |
2b45e8 |
unpckhpd %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -15 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm5
|
|
kusano |
2b45e8 |
subsd %xmm5, %xmm2
|
|
kusano |
2b45e8 |
movsd -14 * SIZE(AA), %xmm6
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm6
|
|
kusano |
2b45e8 |
subsd %xmm6, %xmm1
|
|
kusano |
2b45e8 |
movsd -13 * SIZE(AA), %xmm7
|
|
kusano |
2b45e8 |
mulsd %xmm0, %xmm7
|
|
kusano |
2b45e8 |
subsd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -11 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm2
|
|
kusano |
2b45e8 |
movsd -10 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm5
|
|
kusano |
2b45e8 |
subsd %xmm5, %xmm1
|
|
kusano |
2b45e8 |
movsd -9 * SIZE(AA), %xmm6
|
|
kusano |
2b45e8 |
mulsd %xmm2, %xmm6
|
|
kusano |
2b45e8 |
subsd %xmm6, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -6 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm1
|
|
kusano |
2b45e8 |
movsd -5 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
mulsd %xmm1, %xmm5
|
|
kusano |
2b45e8 |
subsd %xmm5, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -1 * SIZE(AA), %xmm4
|
|
kusano |
2b45e8 |
mulsd %xmm4, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
unpcklpd %xmm2, %xmm0
|
|
kusano |
2b45e8 |
unpcklpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(RN) || defined(RT)
|
|
kusano |
2b45e8 |
movddup -16 * SIZE(B), %xmm4
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $4 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm0, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm0, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm1, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm1, 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addl $4 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(B)
|
|
kusano |
2b45e8 |
movapd %xmm1, -14 * SIZE(B)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movddup %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movddup %xmm1, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
unpckhpd %xmm0, %xmm0
|
|
kusano |
2b45e8 |
unpckhpd %xmm1, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm2, -16 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm0, -14 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm3, -12 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd %xmm1, -10 * SIZE(BB)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd %xmm0, -16 * SIZE(AA)
|
|
kusano |
2b45e8 |
movapd %xmm1, -14 * SIZE(AA)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
leal (,%eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 4), AA
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $4 * SIZE, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $4, KK
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addl $4, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
movl BORIG, B
|
|
kusano |
2b45e8 |
sall $2 + BASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
decl %ebx # i --
|
|
kusano |
2b45e8 |
jg .L110
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L159:
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
leal (, %eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (B, %eax, 1), B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
leal (,%eax, SIZE), %eax
|
|
kusano |
2b45e8 |
leal (B, %eax, 1), B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
addl $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
subl $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L999:
|
|
kusano |
2b45e8 |
movl OLD_STACK, %esp
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
EMMS
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
popl %ebx
|
|
kusano |
2b45e8 |
popl %esi
|
|
kusano |
2b45e8 |
popl %edi
|
|
kusano |
2b45e8 |
popl %ebp
|
|
kusano |
2b45e8 |
ret
|
|
kusano |
2b45e8 |
ALIGN_2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
EPILOGUE
|