Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/x86/izamax_sse.S

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	 0
kusano 2b45e8
	
kusano 2b45e8
#define STACK_M		 4 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_X		 8 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_INCX	12 + STACK + ARGS(%esp)
kusano 2b45e8
kusano 2b45e8
#define RET	%eax
kusano 2b45e8
#define	M	%ebx
kusano 2b45e8
#define X	%ecx
kusano 2b45e8
#define INCX	%edx
kusano 2b45e8
#define I	%esi
kusano 2b45e8
#define MM	%ebp
kusano 2b45e8
#define XX	%edi
kusano 2b45e8
#define TEMP	%ebx
kusano 2b45e8
	
kusano 2b45e8
#ifdef USE_MIN
kusano 2b45e8
#define maxps	minps
kusano 2b45e8
#define maxss	minss
kusano 2b45e8
#endif
kusano 2b45e8
	
kusano 2b45e8
#ifndef HAVE_SSE2
kusano 2b45e8
#define pxor	xorps
kusano 2b45e8
#define movsd	movlps
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#include "l1param.h"
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
kusano 2b45e8
	pushl	%ebp
kusano 2b45e8
	pushl	%edi
kusano 2b45e8
	pushl	%esi
kusano 2b45e8
	pushl	%ebx
kusano 2b45e8
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
	movl	STACK_M, M
kusano 2b45e8
	movl	STACK_X, X
kusano 2b45e8
	movl	STACK_INCX, INCX
kusano 2b45e8
kusano 2b45e8
#ifdef F_INTERFACE
kusano 2b45e8
	movl	(M), M
kusano 2b45e8
	movl	(INCX), INCX
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	pxor	%xmm0, %xmm0
kusano 2b45e8
	pxor	%xmm7, %xmm7
kusano 2b45e8
	xor	RET, RET
kusano 2b45e8
	testl	M, M
kusano 2b45e8
	jle	.L999
kusano 2b45e8
	testl	INCX, INCX
kusano 2b45e8
	jle	.L999
kusano 2b45e8
kusano 2b45e8
	sall	$ZBASE_SHIFT, INCX
kusano 2b45e8
	movl	M, MM
kusano 2b45e8
	movl	X, XX
kusano 2b45e8
kusano 2b45e8
#ifdef USE_ABS
kusano 2b45e8
#ifndef HAVE_SSE2
kusano 2b45e8
	subl	$8, %esp
kusano 2b45e8
	movl	$0x7fffffff, (%esp)
kusano 2b45e8
	movss	(%esp), %xmm7
kusano 2b45e8
	shufps	$0, %xmm7, %xmm7
kusano 2b45e8
	addl	$8, %esp
kusano 2b45e8
#else
kusano 2b45e8
	cmpeqps	%xmm7, %xmm7
kusano 2b45e8
	psrld	$1, %xmm7
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm0
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm1
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	decl	MM
kusano 2b45e8
	andps	%xmm7, %xmm0
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	addps	%xmm1,  %xmm0
kusano 2b45e8
	shufps	$0, %xmm0, %xmm0
kusano 2b45e8
	cmpl	$2 * SIZE, INCX
kusano 2b45e8
	jne	.L70
kusano 2b45e8
kusano 2b45e8
.L30:
kusano 2b45e8
	movl	MM,  I
kusano 2b45e8
	sarl	$3, I
kusano 2b45e8
	jle	.L35
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
	
kusano 2b45e8
.L31:
kusano 2b45e8
#ifdef PREFETCH
kusano 2b45e8
	PREFETCH (PREFETCHSIZE +  0) - PREOFFSET(XX)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movsd	 0 * SIZE(XX), %xmm1
kusano 2b45e8
	movhps	 2 * SIZE(XX), %xmm1
kusano 2b45e8
	movsd	 4 * SIZE(XX), %xmm2
kusano 2b45e8
	movhps	 6 * SIZE(XX), %xmm2
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm1, %xmm3
kusano 2b45e8
kusano 2b45e8
	shufps	$0x88, %xmm2, %xmm1
kusano 2b45e8
	shufps	$0xdd, %xmm2, %xmm3
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	addps	%xmm3,  %xmm1
kusano 2b45e8
	maxps	%xmm1,  %xmm0
kusano 2b45e8
kusano 2b45e8
	movsd	 8 * SIZE(XX), %xmm1
kusano 2b45e8
	movhps	10 * SIZE(XX), %xmm1
kusano 2b45e8
	movsd	12 * SIZE(XX), %xmm2
kusano 2b45e8
	movhps	14 * SIZE(XX), %xmm2
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm1, %xmm3
kusano 2b45e8
kusano 2b45e8
	shufps	$0x88, %xmm2, %xmm1
kusano 2b45e8
	shufps	$0xdd, %xmm2, %xmm3
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	addps	%xmm3,  %xmm1
kusano 2b45e8
	maxps	%xmm1,  %xmm0
kusano 2b45e8
kusano 2b45e8
	addl	$16 * SIZE, XX
kusano 2b45e8
	decl	I
kusano 2b45e8
	jg	.L31
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L35:
kusano 2b45e8
	andl	$7,  MM
kusano 2b45e8
	jle	.L40
kusano 2b45e8
kusano 2b45e8
	testl	$4, MM
kusano 2b45e8
	je	.L36
kusano 2b45e8
kusano 2b45e8
	movsd	 0 * SIZE(XX), %xmm1
kusano 2b45e8
	movhps	 2 * SIZE(XX), %xmm1
kusano 2b45e8
	movsd	 4 * SIZE(XX), %xmm2
kusano 2b45e8
	movhps	 6 * SIZE(XX), %xmm2
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm1, %xmm3
kusano 2b45e8
kusano 2b45e8
	shufps	$0x88, %xmm2, %xmm1
kusano 2b45e8
	shufps	$0xdd, %xmm2, %xmm3
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	addps	%xmm3,  %xmm1
kusano 2b45e8
	maxps	%xmm1,  %xmm0
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE, XX
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L36:
kusano 2b45e8
	testl	$2, MM
kusano 2b45e8
	je	.L37
kusano 2b45e8
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm1
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm2
kusano 2b45e8
	movss	2 * SIZE(XX), %xmm3
kusano 2b45e8
	movss	3 * SIZE(XX), %xmm4
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm2
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	andps	%xmm7, %xmm4
kusano 2b45e8
	addps	%xmm2,  %xmm1
kusano 2b45e8
	addps	%xmm4,  %xmm3
kusano 2b45e8
	maxss	%xmm1,  %xmm0
kusano 2b45e8
	maxss	%xmm3,  %xmm0
kusano 2b45e8
	addl	$4 * SIZE, XX
kusano 2b45e8
	ALIGN_3	
kusano 2b45e8
kusano 2b45e8
.L37:
kusano 2b45e8
	testl	$1, MM
kusano 2b45e8
	je	.L40
kusano 2b45e8
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm1
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm2
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm2
kusano 2b45e8
	addps	%xmm2,  %xmm1
kusano 2b45e8
	maxss	%xmm1,  %xmm0
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L40:
kusano 2b45e8
	movl	X, XX
kusano 2b45e8
	movl	M, MM
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, %xmm1
kusano 2b45e8
	movhlps %xmm0, %xmm0
kusano 2b45e8
	maxps	%xmm1, %xmm0
kusano 2b45e8
	movaps	%xmm0, %xmm1
kusano 2b45e8
	shufps	$1, %xmm0, %xmm0
kusano 2b45e8
	maxss	%xmm1, %xmm0
kusano 2b45e8
	shufps	$0, %xmm0, %xmm0
kusano 2b45e8
kusano 2b45e8
	movl	MM,  I
kusano 2b45e8
	sarl	$2, I
kusano 2b45e8
	jle	.L45
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
	
kusano 2b45e8
.L41:
kusano 2b45e8
#ifdef PREFETCH
kusano 2b45e8
	PREFETCH (PREFETCHSIZE +  0) - PREOFFSET(XX)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movsd	0 * SIZE(XX), %xmm1
kusano 2b45e8
	movhps	2 * SIZE(XX), %xmm1
kusano 2b45e8
	movsd	4 * SIZE(XX), %xmm2
kusano 2b45e8
	movhps	6 * SIZE(XX), %xmm2
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm1, %xmm3
kusano 2b45e8
kusano 2b45e8
	shufps	$0x88, %xmm2, %xmm1
kusano 2b45e8
	shufps	$0xdd, %xmm2, %xmm3
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	addps	%xmm3,  %xmm1
kusano 2b45e8
kusano 2b45e8
	cmpeqps	%xmm0, %xmm1
kusano 2b45e8
	movmskps %xmm1, TEMP
kusano 2b45e8
	testl	 $15, TEMP
kusano 2b45e8
	jne	 .L43
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE, XX
kusano 2b45e8
	addl	$4, RET
kusano 2b45e8
	decl	I
kusano 2b45e8
	jg	.L41
kusano 2b45e8
	jmp	.L45
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L43:
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm1
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm2
kusano 2b45e8
	movss	2 * SIZE(XX), %xmm3
kusano 2b45e8
	movss	3 * SIZE(XX), %xmm4
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm2
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	andps	%xmm7, %xmm4
kusano 2b45e8
kusano 2b45e8
	addps	%xmm2,  %xmm1
kusano 2b45e8
	addps	%xmm4,  %xmm3
kusano 2b45e8
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm1
kusano 2b45e8
	je	.L999
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm3
kusano 2b45e8
	je	.L999
kusano 2b45e8
kusano 2b45e8
	movss	4 * SIZE(XX), %xmm1
kusano 2b45e8
	movss	5 * SIZE(XX), %xmm2
kusano 2b45e8
	movss	6 * SIZE(XX), %xmm3
kusano 2b45e8
	movss	7 * SIZE(XX), %xmm4
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm2
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	andps	%xmm7, %xmm4
kusano 2b45e8
kusano 2b45e8
	addps	%xmm2,  %xmm1
kusano 2b45e8
	addps	%xmm4,  %xmm3
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE, XX
kusano 2b45e8
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm1
kusano 2b45e8
	je	.L999
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm3
kusano 2b45e8
	je	.L999
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L45:
kusano 2b45e8
	testl	$2, MM
kusano 2b45e8
	je	.L47
kusano 2b45e8
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm1
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm2
kusano 2b45e8
	movss	2 * SIZE(XX), %xmm3
kusano 2b45e8
	movss	3 * SIZE(XX), %xmm4
kusano 2b45e8
	addl	$4 * SIZE, XX
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm2
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	andps	%xmm7, %xmm4
kusano 2b45e8
	addps	%xmm2,  %xmm1
kusano 2b45e8
	addps	%xmm4,  %xmm3
kusano 2b45e8
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm1
kusano 2b45e8
	je	.L999
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm3
kusano 2b45e8
	je	.L999
kusano 2b45e8
	ALIGN_3	
kusano 2b45e8
kusano 2b45e8
.L47:
kusano 2b45e8
	incl	RET
kusano 2b45e8
	jmp	.L999
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L70:
kusano 2b45e8
	movl	MM,  I
kusano 2b45e8
	sarl	$3, I
kusano 2b45e8
	jle	.L75
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
	
kusano 2b45e8
.L71:
kusano 2b45e8
#ifdef PREFETCH
kusano 2b45e8
	PREFETCH (PREFETCHSIZE +  0) - PREOFFSET(XX)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movsd	0 * SIZE(XX), %xmm1
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movhps	0 * SIZE(XX), %xmm1
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movsd	0 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movhps	0 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm1, %xmm3
kusano 2b45e8
kusano 2b45e8
	shufps	$0x88, %xmm2, %xmm1
kusano 2b45e8
	shufps	$0xdd, %xmm2, %xmm3
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	addps	%xmm3,  %xmm1
kusano 2b45e8
	maxps	%xmm1,  %xmm0
kusano 2b45e8
kusano 2b45e8
	movsd	0 * SIZE(XX), %xmm1
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movhps	0 * SIZE(XX), %xmm1
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movsd	0 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movhps	0 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm1, %xmm3
kusano 2b45e8
kusano 2b45e8
	shufps	$0x88, %xmm2, %xmm1
kusano 2b45e8
	shufps	$0xdd, %xmm2, %xmm3
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	addps	%xmm3,  %xmm1
kusano 2b45e8
	maxps	%xmm1,  %xmm0
kusano 2b45e8
	decl	I
kusano 2b45e8
	jg	.L71
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L75:
kusano 2b45e8
	andl	$7,  MM
kusano 2b45e8
	jle	.L80
kusano 2b45e8
kusano 2b45e8
	testl	$4, MM
kusano 2b45e8
	je	.L76
kusano 2b45e8
kusano 2b45e8
	movsd	0 * SIZE(XX), %xmm1
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movhps	0 * SIZE(XX), %xmm1
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movsd	0 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movhps	0 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm1, %xmm3
kusano 2b45e8
kusano 2b45e8
	shufps	$0x88, %xmm2, %xmm1
kusano 2b45e8
	shufps	$0xdd, %xmm2, %xmm3
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	addps	%xmm3,  %xmm1
kusano 2b45e8
	maxps	%xmm1,  %xmm0
kusano 2b45e8
	ALIGN_3	
kusano 2b45e8
kusano 2b45e8
.L76:
kusano 2b45e8
	testl	$2, MM
kusano 2b45e8
	je	.L77
kusano 2b45e8
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm1
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm3
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm4
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm2
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	andps	%xmm7, %xmm4
kusano 2b45e8
	addps	%xmm2,  %xmm1
kusano 2b45e8
	addps	%xmm4,  %xmm3
kusano 2b45e8
	maxss	%xmm1,  %xmm0
kusano 2b45e8
	maxss	%xmm3,  %xmm0
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
	
kusano 2b45e8
.L77:
kusano 2b45e8
	testl	$1, MM
kusano 2b45e8
	je	.L80
kusano 2b45e8
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm1
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm2
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm2
kusano 2b45e8
	addps	%xmm2,  %xmm1
kusano 2b45e8
	maxss	%xmm1,  %xmm0
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L80:
kusano 2b45e8
	movl	X, XX
kusano 2b45e8
	movl	M, MM
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, %xmm1
kusano 2b45e8
	movhlps %xmm0, %xmm0
kusano 2b45e8
	maxps	%xmm1, %xmm0
kusano 2b45e8
	movaps	%xmm0, %xmm1
kusano 2b45e8
	shufps	$1, %xmm0, %xmm0
kusano 2b45e8
	maxss	%xmm1, %xmm0
kusano 2b45e8
	shufps	$0, %xmm0, %xmm0
kusano 2b45e8
kusano 2b45e8
	movl	MM,  I
kusano 2b45e8
	sarl	$2, I
kusano 2b45e8
	jle	.L85
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
	
kusano 2b45e8
.L81:
kusano 2b45e8
#ifdef PREFETCH
kusano 2b45e8
	PREFETCH (PREFETCHSIZE +  0) - PREOFFSET(XX)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movsd	0 * SIZE(XX), %xmm1
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movhps	0 * SIZE(XX), %xmm1
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movsd	0 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movhps	0 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm1, %xmm3
kusano 2b45e8
kusano 2b45e8
	shufps	$0x88, %xmm2, %xmm1
kusano 2b45e8
	shufps	$0xdd, %xmm2, %xmm3
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	addps	%xmm3,  %xmm1
kusano 2b45e8
kusano 2b45e8
	cmpeqps	%xmm0, %xmm1
kusano 2b45e8
	movmskps %xmm1, TEMP
kusano 2b45e8
	testl	 $15, TEMP
kusano 2b45e8
	jne	 .L83
kusano 2b45e8
kusano 2b45e8
	addl	$4, RET
kusano 2b45e8
	decl	I
kusano 2b45e8
	jg	.L81
kusano 2b45e8
	jmp	.L85
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L83:
kusano 2b45e8
	leal	(, INCX, 4), TEMP
kusano 2b45e8
	subl	TEMP, XX
kusano 2b45e8
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm1
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm3
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm4
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm2
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	andps	%xmm7, %xmm4
kusano 2b45e8
kusano 2b45e8
	addps	%xmm2,  %xmm1
kusano 2b45e8
	addps	%xmm4,  %xmm3
kusano 2b45e8
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm1
kusano 2b45e8
	je	.L999
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm3
kusano 2b45e8
	je	.L999
kusano 2b45e8
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm1
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm3
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm4
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm2
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	andps	%xmm7, %xmm4
kusano 2b45e8
kusano 2b45e8
	addps	%xmm2,  %xmm1
kusano 2b45e8
	addps	%xmm4,  %xmm3
kusano 2b45e8
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm1
kusano 2b45e8
	je	.L999
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm3
kusano 2b45e8
	je	.L999
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L85:
kusano 2b45e8
	testl	$2, MM
kusano 2b45e8
	je	.L87
kusano 2b45e8
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm1
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm2
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
	movss	0 * SIZE(XX), %xmm3
kusano 2b45e8
	movss	1 * SIZE(XX), %xmm4
kusano 2b45e8
	addl	INCX, XX
kusano 2b45e8
kusano 2b45e8
	andps	%xmm7, %xmm1
kusano 2b45e8
	andps	%xmm7, %xmm2
kusano 2b45e8
	andps	%xmm7, %xmm3
kusano 2b45e8
	andps	%xmm7, %xmm4
kusano 2b45e8
	addps	%xmm2, %xmm1
kusano 2b45e8
	addps	%xmm4, %xmm3
kusano 2b45e8
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm1
kusano 2b45e8
	je	.L999
kusano 2b45e8
	incl	RET
kusano 2b45e8
	comiss	%xmm0, %xmm3
kusano 2b45e8
	je	.L999
kusano 2b45e8
	ALIGN_3	
kusano 2b45e8
kusano 2b45e8
.L87:
kusano 2b45e8
	incl	RET
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
	ret
kusano 2b45e8
kusano 2b45e8
	EPILOGUE