|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
/* Copyright 2009, 2010 The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/* All rights reserved. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* Redistribution and use in source and binary forms, with or */
|
|
kusano |
2b45e8 |
/* without modification, are permitted provided that the following */
|
|
kusano |
2b45e8 |
/* conditions are met: */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 1. Redistributions of source code must retain the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 2. Redistributions in binary form must reproduce the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer in the documentation and/or other materials */
|
|
kusano |
2b45e8 |
/* provided with the distribution. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
|
|
kusano |
2b45e8 |
/* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
|
|
kusano |
2b45e8 |
/* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
|
|
kusano |
2b45e8 |
/* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
|
|
kusano |
2b45e8 |
/* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
|
|
kusano |
2b45e8 |
/* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
|
|
kusano |
2b45e8 |
/* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
|
|
kusano |
2b45e8 |
/* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
|
|
kusano |
2b45e8 |
/* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
|
|
kusano |
2b45e8 |
/* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
|
|
kusano |
2b45e8 |
/* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
|
|
kusano |
2b45e8 |
/* POSSIBILITY OF SUCH DAMAGE. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* The views and conclusions contained in the software and */
|
|
kusano |
2b45e8 |
/* documentation are those of the authors and should not be */
|
|
kusano |
2b45e8 |
/* interpreted as representing official policies, either expressed */
|
|
kusano |
2b45e8 |
/* or implied, of The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ASSEMBLER
|
|
kusano |
2b45e8 |
#include "common.h"
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACK 16
|
|
kusano |
2b45e8 |
#define ARGS 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define M 4 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define N 8 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define K 12 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define A 24 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define ARG_B 28 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define C 32 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define ARG_LDC 36 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
#define OFFSET 40 + STACK + ARGS(%esp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define J 0 + STACK(%esp)
|
|
kusano |
2b45e8 |
#define KK 4 + STACK(%esp)
|
|
kusano |
2b45e8 |
#define KKK 8 + STACK(%esp)
|
|
kusano |
2b45e8 |
#define AORIG 12 + STACK(%esp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(PENRYN) || defined(DUNNINGTON)
|
|
kusano |
2b45e8 |
#define PREFETCH prefetcht1
|
|
kusano |
2b45e8 |
#define PREFETCHSIZE 84
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef NEHALEM
|
|
kusano |
2b45e8 |
#define PREFETCH prefetcht1
|
|
kusano |
2b45e8 |
#define PREFETCHSIZE 84
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef ATOM
|
|
kusano |
2b45e8 |
#define PREFETCH prefetcht0
|
|
kusano |
2b45e8 |
#define PREFETCHSIZE 84
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef NANO
|
|
kusano |
2b45e8 |
#define PREFETCH prefetcht0
|
|
kusano |
2b45e8 |
#define PREFETCHSIZE (16 * 2)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define B %edi
|
|
kusano |
2b45e8 |
#define LDC %ebp
|
|
kusano |
2b45e8 |
#define AA %edx
|
|
kusano |
2b45e8 |
#define BB %ecx
|
|
kusano |
2b45e8 |
#define CO1 %esi
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ADD1 addps
|
|
kusano |
2b45e8 |
#define ADD2 addps
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROLOGUE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $ARGS, %esp
|
|
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 ARG_B, B
|
|
kusano |
2b45e8 |
movl ARG_LDC, LDC
|
|
kusano |
2b45e8 |
movl OFFSET, %eax
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
negl %eax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movl %eax, KK
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
testl %ebx, %ebx
|
|
kusano |
2b45e8 |
jle .L999
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-32 * SIZE, A
|
|
kusano |
2b45e8 |
subl $-32 * SIZE, B
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, LDC
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl M, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, %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 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
imull K, %eax
|
|
kusano |
2b45e8 |
addl %eax, B
|
|
kusano |
2b45e8 |
|
|
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 |
movl %eax, J
|
|
kusano |
2b45e8 |
sarl $1, J
|
|
kusano |
2b45e8 |
jle .L100
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L01:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl A, %eax
|
|
kusano |
2b45e8 |
movl %eax, 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 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $1 + ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
subl %eax, B
|
|
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
|
|
kusano |
2b45e8 |
#ifndef RT
|
|
kusano |
2b45e8 |
addl %eax, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
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 |
#ifdef LT
|
|
kusano |
2b45e8 |
movl OFFSET, %eax
|
|
kusano |
2b45e8 |
movl %eax, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
andl $1, %ebx
|
|
kusano |
2b45e8 |
jle .L30
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_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 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl B, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
sall $1 + ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -32 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm2, %xmm2
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
pxor %xmm6, %xmm6
|
|
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 .L42
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L41:
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -30 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -24 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -28 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -20 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -26 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -24 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -22 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -8 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -20 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -4 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -18 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps 0 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, AA
|
|
kusano |
2b45e8 |
subl $-32 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L41
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L42:
|
|
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 .L44
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L43:
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -30 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $2 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $4 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L43
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L44:
|
|
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 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 1), AA
|
|
kusano |
2b45e8 |
leal (B, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm6
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm5, %xmm5
|
|
kusano |
2b45e8 |
pcmpeqb %xmm0, %xmm0
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm7, %xmm7
|
|
kusano |
2b45e8 |
psllq $63, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
shufps $0xb1, %xmm0, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm5
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm7
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm4
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm6
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm5
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm7
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm5, %xmm4
|
|
kusano |
2b45e8 |
addps %xmm7, %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
unpcklpd %xmm6, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd -32 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
movsd -30 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm1
|
|
kusano |
2b45e8 |
subps %xmm6, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm2, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm1, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
subps %xmm2, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(BB), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm5, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm5, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm5
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(BB), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm5, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm5, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm5
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm5, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm5, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm3, %xmm1
|
|
kusano |
2b45e8 |
subps %xmm2, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm1, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm1
|
|
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 |
movaps %xmm2, -32 * SIZE(BB)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlps %xmm2, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhps %xmm2, 0 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movlps %xmm1, -32 * SIZE(AA)
|
|
kusano |
2b45e8 |
movlps %xmm5, -30 * SIZE(AA)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlps %xmm1, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movlps %xmm5, 0 * 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(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 1), AA
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $1, KK
|
|
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 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L30:
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
sarl $1, %ebx
|
|
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 $1 + ZBASE_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 |
sall $1 + ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl B, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
sall $1 + ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm2, %xmm2
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
prefetcht0 -4 * SIZE(CO1)
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
prefetcht0 -4 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
pxor %xmm6, %xmm6
|
|
kusano |
2b45e8 |
pxor %xmm7, %xmm7
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
prefetcht0 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
prefetcht0 3 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
pxor %xmm6, %xmm6
|
|
kusano |
2b45e8 |
pxor %xmm7, %xmm7
|
|
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 |
.L11:
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
ADD1 %xmm3, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x1b, %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
ADD1 %xmm3, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x1b, %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movaps -24 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movaps -24 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
ADD1 %xmm3, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x1b, %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movaps -20 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movaps -20 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
ADD1 %xmm3, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x1b, %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movaps -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movaps -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 16) * SIZE(AA)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
ADD1 %xmm3, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x1b, %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movaps -12 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movaps -12 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
ADD1 %xmm3, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x1b, %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movaps -8 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movaps -8 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
ADD1 %xmm3, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x1b, %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movaps -4 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movaps -4 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm7
|
|
kusano |
2b45e8 |
subl $-32 * SIZE, BB
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
ADD1 %xmm3, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x1b, %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm5
|
|
kusano |
2b45e8 |
subl $-32 * SIZE, AA
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L11
|
|
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 # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
je .L14
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L13:
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm7
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
ADD1 %xmm3, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x1b, %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm3, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
ADD1 %xmm1, %xmm4
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $4 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $4 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L13
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L14:
|
|
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 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 2), AA
|
|
kusano |
2b45e8 |
leal (B, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
ADD2 %xmm2, %xmm7
|
|
kusano |
2b45e8 |
pcmpeqb %xmm0, %xmm0
|
|
kusano |
2b45e8 |
ADD1 %xmm3, %xmm6
|
|
kusano |
2b45e8 |
psllq $63, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm4
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
shufps $0xb1, %xmm0, %xmm0
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm5
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm7
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm0, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pxor %xmm1, %xmm5
|
|
kusano |
2b45e8 |
pxor %xmm1, %xmm7
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
haddps %xmm5, %xmm4
|
|
kusano |
2b45e8 |
haddps %xmm7, %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
shufps $0xd8, %xmm4, %xmm4
|
|
kusano |
2b45e8 |
shufps $0xd8, %xmm6, %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps %xmm4, %xmm5
|
|
kusano |
2b45e8 |
shufps $0xe4, %xmm6, %xmm4
|
|
kusano |
2b45e8 |
shufps $0xe4, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movaps %xmm4, %xmm5
|
|
kusano |
2b45e8 |
unpcklpd %xmm6, %xmm4
|
|
kusano |
2b45e8 |
unpckhpd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
subps %xmm5, %xmm3
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm1
|
|
kusano |
2b45e8 |
subps %xmm6, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm3, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm3
|
|
kusano |
2b45e8 |
addps %xmm4, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm3, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm3, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm1
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
subps %xmm1, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm2, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm2, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm2, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm1
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm3
|
|
kusano |
2b45e8 |
subps %xmm1, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm3, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm3
|
|
kusano |
2b45e8 |
addps %xmm4, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm1, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
subps %xmm2, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(BB), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm5, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm5, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm5
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(BB), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm5, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm5, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm5
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm5, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm5, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm3, %xmm1
|
|
kusano |
2b45e8 |
subps %xmm2, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm1, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm3, %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 |
movaps %xmm2, -32 * SIZE(BB)
|
|
kusano |
2b45e8 |
movaps %xmm3, -28 * SIZE(BB)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlps %xmm2, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movlps %xmm3, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhps %xmm2, 0 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
movhps %xmm3, 2 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movaps %xmm1, -32 * SIZE(AA)
|
|
kusano |
2b45e8 |
movaps %xmm5, -28 * SIZE(AA)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlps %xmm1, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhps %xmm1, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlps %xmm5, 0 * SIZE(CO1, LDC)
|
|
kusano |
2b45e8 |
movhps %xmm5, 2 * 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(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 2), AA
|
|
kusano |
2b45e8 |
leal (BB, %eax, 2), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $2, KK
|
|
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 |
sall $1 + ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decl %ebx
|
|
kusano |
2b45e8 |
jg .L10
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L99:
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $1 + ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl BB, 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_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L100:
|
|
kusano |
2b45e8 |
movl N, %eax
|
|
kusano |
2b45e8 |
andl $1, %eax
|
|
kusano |
2b45e8 |
jle .L999
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl A, %eax
|
|
kusano |
2b45e8 |
movl %eax, 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 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
subl %eax, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
subl LDC, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movl C, CO1
|
|
kusano |
2b45e8 |
#ifndef RT
|
|
kusano |
2b45e8 |
addl LDC, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
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 |
#ifdef LT
|
|
kusano |
2b45e8 |
movl OFFSET, %eax
|
|
kusano |
2b45e8 |
movl %eax, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
andl $1, %ebx
|
|
kusano |
2b45e8 |
jle .L130
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_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 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl B, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -32 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm2, %xmm2
|
|
kusano |
2b45e8 |
movsd -32 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
pxor %xmm6, %xmm6
|
|
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 .L142
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L141:
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movsd -30 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -30 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movsd -28 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -28 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movsd -26 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -26 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movsd -24 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -24 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movsd -22 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -22 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movsd -20 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -20 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movsd -18 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -18 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, AA
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L141
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L142:
|
|
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 .L144
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L143:
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movsd -30 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movsd -30 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $2 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $2 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L143
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L144:
|
|
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 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 1), AA
|
|
kusano |
2b45e8 |
leal (B, %eax, 1), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm5, %xmm5
|
|
kusano |
2b45e8 |
pcmpeqb %xmm0, %xmm0
|
|
kusano |
2b45e8 |
psllq $63, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
shufps $0xb1, %xmm0, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm5
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm5, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd -32 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd -32 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm2, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(RN) || defined(RT)
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm1, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm1
|
|
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 |
movlps %xmm2, -32 * SIZE(BB)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlps %xmm2, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movlps %xmm1, -32 * SIZE(AA)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlps %xmm1, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addl $2 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 1), AA
|
|
kusano |
2b45e8 |
leal (BB, %eax, 1), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $1, KK
|
|
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 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L130:
|
|
kusano |
2b45e8 |
movl M, %ebx
|
|
kusano |
2b45e8 |
sarl $1, %ebx
|
|
kusano |
2b45e8 |
jle .L149
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L110:
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $1 + ZBASE_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 |
sall $1 + ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AA
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movl B, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movl KK, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm2, %xmm2
|
|
kusano |
2b45e8 |
movsd -32 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm3, %xmm3
|
|
kusano |
2b45e8 |
movhps -30 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
prefetcht0 -4 * SIZE(CO1)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
prefetcht0 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
pxor %xmm6, %xmm6
|
|
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 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L111:
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(AA)
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movaps -24 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movaps -20 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -24 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movaps -16 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 16) * SIZE(AA)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movaps -12 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -20 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movaps -8 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movaps -4 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xaa, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0xff, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movaps -16 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movaps 0 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subl $-32 * SIZE, AA
|
|
kusano |
2b45e8 |
subl $-16 * SIZE, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jne .L111
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
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 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L113:
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0x00, %xmm1, %xmm2
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
pshufd $0x55, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
movsd -30 * SIZE(BB), %xmm1
|
|
kusano |
2b45e8 |
mulps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(AA), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $4 * SIZE, AA
|
|
kusano |
2b45e8 |
addl $2 * SIZE, BB
|
|
kusano |
2b45e8 |
decl %eax
|
|
kusano |
2b45e8 |
jg .L113
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L114:
|
|
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 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 2), AA
|
|
kusano |
2b45e8 |
leal (B, %eax, 1), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm2, %xmm4
|
|
kusano |
2b45e8 |
addps %xmm3, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xb1, %xmm5, %xmm5
|
|
kusano |
2b45e8 |
pcmpeqb %xmm0, %xmm0
|
|
kusano |
2b45e8 |
psllq $63, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
shufps $0xb1, %xmm0, %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm5
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm5, %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movaps %xmm4, %xmm5
|
|
kusano |
2b45e8 |
unpcklpd %xmm6, %xmm4
|
|
kusano |
2b45e8 |
unpckhpd %xmm6, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd -32 * SIZE(BB), %xmm2
|
|
kusano |
2b45e8 |
movsd -30 * SIZE(BB), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
subps %xmm5, %xmm3
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm3, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm3
|
|
kusano |
2b45e8 |
addps %xmm4, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm3, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm3, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm1
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
subps %xmm1, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm2, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm2, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm2
|
|
kusano |
2b45e8 |
addps %xmm4, %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm2, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm2, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm1
|
|
kusano |
2b45e8 |
subps %xmm4, %xmm3
|
|
kusano |
2b45e8 |
subps %xmm1, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movaps -28 * SIZE(AA), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xee, %xmm5, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0xbb, %xmm5, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm3, %xmm4
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm4
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm4
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm3
|
|
kusano |
2b45e8 |
addps %xmm4, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(RN) || defined(RT)
|
|
kusano |
2b45e8 |
movaps -32 * SIZE(BB), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x44, %xmm4, %xmm6
|
|
kusano |
2b45e8 |
pshufd $0x11, %xmm4, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0xa0, %xmm1, %xmm3
|
|
kusano |
2b45e8 |
pshufd $0xf5, %xmm1, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorps %xmm0, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulps %xmm6, %xmm3
|
|
kusano |
2b45e8 |
mulps %xmm7, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addps %xmm3, %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 |
movlps %xmm2, -32 * SIZE(BB)
|
|
kusano |
2b45e8 |
movlps %xmm3, -30 * SIZE(BB)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlps %xmm2, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movlps %xmm3, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movaps %xmm1, -32 * SIZE(AA)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlps %xmm1, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhps %xmm1, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addl $4 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
subl KK, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
leal (AA, %eax, 2), AA
|
|
kusano |
2b45e8 |
leal (BB, %eax, 1), BB
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subl $2, KK
|
|
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 |
sall $1 + ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decl %ebx # i --
|
|
kusano |
2b45e8 |
jg .L110
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L149:
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movl K, %eax
|
|
kusano |
2b45e8 |
sall $ZBASE_SHIFT, %eax
|
|
kusano |
2b45e8 |
addl %eax, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movl BB, 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_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L999:
|
|
kusano |
2b45e8 |
popl %ebx
|
|
kusano |
2b45e8 |
popl %esi
|
|
kusano |
2b45e8 |
popl %edi
|
|
kusano |
2b45e8 |
popl %ebp
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addl $ARGS, %esp
|
|
kusano |
2b45e8 |
ret
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
EPILOGUE
|