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		20 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_INCX	24 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_Y		28 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_INCY	32 + STACK + ARGS(%esp)
kusano 2b45e8
kusano 2b45e8
#define M	%edx
kusano 2b45e8
#define X	%esi
kusano 2b45e8
#define Y	%edi
kusano 2b45e8
#define INCX	%ebx
kusano 2b45e8
#define INCY	%ecx
kusano 2b45e8
	
kusano 2b45e8
#include "l1param.h"
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
	pushl	%ebp
kusano 2b45e8
	pushl	%edi
kusano 2b45e8
	pushl	%esi
kusano 2b45e8
	pushl	%ebx
kusano 2b45e8
kusano 2b45e8
	movl	STACK_M,    M
kusano 2b45e8
	movl	STACK_X,    X
kusano 2b45e8
	movl	STACK_Y,    Y
kusano 2b45e8
	movl	STACK_INCX, INCX
kusano 2b45e8
	movl	STACK_INCY, INCY
kusano 2b45e8
kusano 2b45e8
	sall	$BASE_SHIFT, %ebx
kusano 2b45e8
	sall	$BASE_SHIFT, %ecx
kusano 2b45e8
kusano 2b45e8
	cmpl	$SIZE, INCX
kusano 2b45e8
	jne	.L50
kusano 2b45e8
	cmpl	$SIZE, INCY
kusano 2b45e8
	jne	.L50
kusano 2b45e8
kusano 2b45e8
	subl	$-32 * SIZE, X
kusano 2b45e8
	subl	$-32 * SIZE, Y
kusano 2b45e8
	
kusano 2b45e8
	cmpl	$3, M
kusano 2b45e8
	jle	.L16
kusano 2b45e8
kusano 2b45e8
	testl	$SIZE, Y
kusano 2b45e8
	je	.L05
kusano 2b45e8
kusano 2b45e8
	movss	-32 * SIZE(X), %xmm0
kusano 2b45e8
	movss	-32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, -32 * SIZE(X)
kusano 2b45e8
	movss	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	addl	$1 * SIZE, X
kusano 2b45e8
	addl	$1 * SIZE, Y
kusano 2b45e8
	decl	M
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L05:
kusano 2b45e8
	testl	$2 * SIZE, Y
kusano 2b45e8
	je	.L10
kusano 2b45e8
kusano 2b45e8
	movsd	-32 * SIZE(X), %xmm0
kusano 2b45e8
	movsd	-32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movlps	%xmm1, -32 * SIZE(X)
kusano 2b45e8
	movlps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	addl	$2 * SIZE, X
kusano 2b45e8
	addl	$2 * SIZE, Y
kusano 2b45e8
	subl	$2, M
kusano 2b45e8
	jle	.L19
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L10:
kusano 2b45e8
	cmpl	$3, M
kusano 2b45e8
	jle	.L16
kusano 2b45e8
kusano 2b45e8
	testl	$2 * SIZE, X
kusano 2b45e8
	jne	.L30
kusano 2b45e8
kusano 2b45e8
	testl	$1 * SIZE, X
kusano 2b45e8
	jne	.L20
kusano 2b45e8
kusano 2b45e8
	movl	M,  %eax
kusano 2b45e8
	sarl	$5, %eax
kusano 2b45e8
	jle	.L13
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L11:
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE +  0) - PREOFFSET(X)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-32 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-32 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-32 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-28 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-28 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-28 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE +  0) - PREOFFSET(Y)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-24 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-24 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-24 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-20 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-20 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-20 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-20 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE + 64) - PREOFFSET(X)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-16 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-16 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-16 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-16 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-12 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-12 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-12 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-12 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE + 64) - PREOFFSET(Y)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	 -8 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	 -8 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	 -8 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	 -8 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	 -4 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	 -4 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	 -4 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	 -4 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	subl	$-32 * SIZE, Y
kusano 2b45e8
	subl	$-32 * SIZE, X
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L11
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L13:
kusano 2b45e8
	testl	$16, M
kusano 2b45e8
	jle	.L14
kusano 2b45e8
kusano 2b45e8
	movaps	-32 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-32 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-32 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-28 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-28 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-28 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-24 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-24 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-24 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-20 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-20 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-20 * SIZE(Y)
kusano 2b45e8
 	movaps	%xmm1, 	-20 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$16 * SIZE, X
kusano 2b45e8
	addl	$16 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L14:
kusano 2b45e8
	testl	$8, M
kusano 2b45e8
	jle	.L15
kusano 2b45e8
kusano 2b45e8
	movaps	-32 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-32 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-32 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-28 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-28 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-28 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE, X
kusano 2b45e8
	addl	$8 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L15:
kusano 2b45e8
	testl	$4, M
kusano 2b45e8
	jle	.L16
kusano 2b45e8
kusano 2b45e8
	movaps	-32 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm0, 	-32 * SIZE(Y)
kusano 2b45e8
	movaps	%xmm1, 	-32 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$4 * SIZE, X
kusano 2b45e8
	addl	$4 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L16:
kusano 2b45e8
	testl	$2, M
kusano 2b45e8
	jle	.L17
kusano 2b45e8
kusano 2b45e8
	movsd	-32 * SIZE(X), %xmm0
kusano 2b45e8
	movsd	-32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movlps	%xmm1, 	-32 * SIZE(X)
kusano 2b45e8
	addl	$2 * SIZE, X
kusano 2b45e8
	movlps	%xmm0, 	-32 * SIZE(Y)
kusano 2b45e8
	addl	$2 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L17:
kusano 2b45e8
	testl	$1, M
kusano 2b45e8
	jle	.L19
kusano 2b45e8
kusano 2b45e8
	movss	-32 * SIZE(X), %xmm0
kusano 2b45e8
	movss	-32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, 	-32 * SIZE(X)
kusano 2b45e8
	movss	%xmm0, 	-32 * SIZE(Y)
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L19:
kusano 2b45e8
	popl	%ebx
kusano 2b45e8
	popl	%esi
kusano 2b45e8
	popl	%edi
kusano 2b45e8
	popl	%ebp
kusano 2b45e8
	ret
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L20:
kusano 2b45e8
	movaps	 -33 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	 -32 * SIZE(Y), %xmm1
kusano 2b45e8
	
kusano 2b45e8
	movss	%xmm1, -32 * SIZE(X)
kusano 2b45e8
	PSHUFD2($0x39, %xmm1, %xmm3)
kusano 2b45e8
	movlps	%xmm3, -31 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	subl	$3, M
kusano 2b45e8
kusano 2b45e8
	movl	M,  %eax
kusano 2b45e8
	sarl	$5, %eax
kusano 2b45e8
	jle	.L23
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L21:
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE +  0) - PREOFFSET(X)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-29 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x39, %xmm0, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x93, %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -29 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-25 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x39, %xmm2, %xmm2
kusano 2b45e8
	movaps	%xmm2, -28 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x93, %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -25 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE +  0) - PREOFFSET(Y)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-21 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-20 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x39, %xmm0, %xmm0
kusano 2b45e8
	movaps	%xmm0, -24 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x93, %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -21 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-17 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-16 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x39, %xmm2, %xmm2
kusano 2b45e8
	movaps	%xmm2, -20 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x93, %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -17 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE + 64) - PREOFFSET(X)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-13 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-12 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x39, %xmm0, %xmm0
kusano 2b45e8
	movaps	%xmm0, -16 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x93, %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -13 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	 -9 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	 -8 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x39, %xmm2, %xmm2
kusano 2b45e8
	movaps	%xmm2, -12 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x93, %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3,  -9 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE + 64) - PREOFFSET(Y)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	 -5 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	 -4 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x39, %xmm0, %xmm0
kusano 2b45e8
	movaps	%xmm0,  -8 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x93, %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1,  -5 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	 -1 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	  0 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x39, %xmm2, %xmm2
kusano 2b45e8
	movaps	%xmm2,  -4 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x93, %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3,  -1 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	subl	$-32 * SIZE, X
kusano 2b45e8
	subl	$-32 * SIZE, Y
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L21
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L23:
kusano 2b45e8
	testl	$16, M
kusano 2b45e8
	jle	.L24
kusano 2b45e8
kusano 2b45e8
	movaps	-29 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x39, %xmm0, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x93, %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -29 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-25 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x39, %xmm2, %xmm2
kusano 2b45e8
	movaps	%xmm2, -28 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x93, %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -25 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-21 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-20 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x39, %xmm0, %xmm0
kusano 2b45e8
	movaps	%xmm0, -24 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x93, %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -21 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-17 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-16 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x39, %xmm2, %xmm2
kusano 2b45e8
	movaps	%xmm2, -20 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x93, %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -17 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$16 * SIZE, X
kusano 2b45e8
	addl	$16 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L24:
kusano 2b45e8
	testl	$8, M
kusano 2b45e8
	jle	.L25
kusano 2b45e8
kusano 2b45e8
	movaps	-29 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x39, %xmm0, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x93, %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -29 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-25 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x39, %xmm2, %xmm2
kusano 2b45e8
	movaps	%xmm2, -28 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x93, %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -25 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE, X
kusano 2b45e8
	addl	$8 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L25:
kusano 2b45e8
	testl	$4, M
kusano 2b45e8
	jle	.L26
kusano 2b45e8
kusano 2b45e8
	movaps	-29 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x39, %xmm0, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x93, %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -29 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm3, %xmm1
kusano 2b45e8
kusano 2b45e8
	addl	$4 * SIZE, X
kusano 2b45e8
	addl	$4 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L26:
kusano 2b45e8
	PSHUFD2($0x39, %xmm0, %xmm2)
kusano 2b45e8
	PSHUFD1($0xff, %xmm0)
kusano 2b45e8
kusano 2b45e8
	movlps	%xmm2, -32 * SIZE(Y)
kusano 2b45e8
	movss	%xmm0, -30 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	testl	$2, M
kusano 2b45e8
	jle	.L27
kusano 2b45e8
kusano 2b45e8
	movsd	-29 * SIZE(X), %xmm0
kusano 2b45e8
	movsd	-29 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movlps	%xmm0, -29 * SIZE(Y)
kusano 2b45e8
	movlps	%xmm1, -29 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$2 * SIZE, X
kusano 2b45e8
	addl	$2 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L27:
kusano 2b45e8
	testl	$1, M
kusano 2b45e8
	jle	.L29
kusano 2b45e8
kusano 2b45e8
	movss	-29 * SIZE(X), %xmm0
kusano 2b45e8
	movss	-29 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, -29 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, -29 * SIZE(X)
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L29:
kusano 2b45e8
	popl	%ebx
kusano 2b45e8
	popl	%esi
kusano 2b45e8
	popl	%edi
kusano 2b45e8
	popl	%ebp
kusano 2b45e8
	ret
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L30:
kusano 2b45e8
	testl	$1 * SIZE, X
kusano 2b45e8
	jne	.L40
kusano 2b45e8
kusano 2b45e8
	movhps	 -32 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	 -32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movlps	%xmm1, -32 * SIZE(X)
kusano 2b45e8
	subl	$2, M
kusano 2b45e8
kusano 2b45e8
	movl	M,  %eax
kusano 2b45e8
	sarl	$5, %eax
kusano 2b45e8
	jle	.L33
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L31:
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE +  0) - PREOFFSET(X)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-30 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -30 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-26 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -28 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -26 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE +  0) - PREOFFSET(Y)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-22 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-20 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -24 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -22 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-18 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-16 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -20 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -18 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE + 64) - PREOFFSET(X)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-14 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-12 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -16 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -14 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-10 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	 -8 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -12 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -10 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE + 64) - PREOFFSET(Y)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	 -6 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	 -4 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0,  -8 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1,  -6 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	 -2 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	  0 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2,  -4 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3,  -2 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	subl	$-32 * SIZE, X
kusano 2b45e8
	subl	$-32 * SIZE, Y
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L31
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L33:
kusano 2b45e8
	testl	$16, M
kusano 2b45e8
	jle	.L34
kusano 2b45e8
kusano 2b45e8
	movaps	-30 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -30 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-26 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -28 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -26 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-22 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-20 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -24 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -22 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-18 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-16 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -20 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -18 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$16 * SIZE, X
kusano 2b45e8
	addl	$16 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L34:
kusano 2b45e8
	testl	$8, M
kusano 2b45e8
	jle	.L35
kusano 2b45e8
kusano 2b45e8
	movaps	-30 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -30 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-26 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -28 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm1, %xmm3
kusano 2b45e8
	movaps	%xmm3, -26 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE, X
kusano 2b45e8
	addl	$8 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L35:
kusano 2b45e8
	testl	$4, M
kusano 2b45e8
	jle	.L36
kusano 2b45e8
kusano 2b45e8
	movaps	-30 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	SHUFPD_1  %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	SHUFPD_1  %xmm3, %xmm1
kusano 2b45e8
	movaps	%xmm1, -30 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm3, %xmm1
kusano 2b45e8
kusano 2b45e8
	addl	$4 * SIZE, X
kusano 2b45e8
	addl	$4 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L36:
kusano 2b45e8
	movhps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	testl	$2, M
kusano 2b45e8
	jle	.L37
kusano 2b45e8
kusano 2b45e8
	movsd	-30 * SIZE(X), %xmm0
kusano 2b45e8
	movsd	-30 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movlps	%xmm0, -30 * SIZE(Y)
kusano 2b45e8
	movlps	%xmm1, -30 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$2 * SIZE, X
kusano 2b45e8
	addl	$2 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L37:
kusano 2b45e8
	testl	$1, M
kusano 2b45e8
	jle	.L39
kusano 2b45e8
kusano 2b45e8
	movss	-30 * SIZE(X), %xmm0
kusano 2b45e8
	movss	-30 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, -30 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, -30 * SIZE(X)
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L39:
kusano 2b45e8
	popl	%ebx
kusano 2b45e8
	popl	%esi
kusano 2b45e8
	popl	%edi
kusano 2b45e8
	popl	%ebp
kusano 2b45e8
	ret
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L40:
kusano 2b45e8
	movaps	 -35 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	 -32 * SIZE(Y), %xmm1
kusano 2b45e8
	
kusano 2b45e8
	movss	%xmm1, -32 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	subl	$3, M
kusano 2b45e8
kusano 2b45e8
	movl	M,  %eax
kusano 2b45e8
	sarl	$5, %eax
kusano 2b45e8
	jle	.L43
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L41:
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE +  0) - PREOFFSET(X)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-31 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x93, %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x39, %xmm1, %xmm1
kusano 2b45e8
	movaps	%xmm1, -31 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-27 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x93, %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -28 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x39, %xmm3, %xmm3
kusano 2b45e8
	movaps	%xmm3, -27 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE +  0) - PREOFFSET(Y)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-23 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-20 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x93, %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -24 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x39, %xmm1, %xmm1
kusano 2b45e8
	movaps	%xmm1, -23 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-19 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-16 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x93, %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -20 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x39, %xmm3, %xmm3
kusano 2b45e8
	movaps	%xmm3, -19 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE + 64) - PREOFFSET(X)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	-15 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-12 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x93, %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -16 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x39, %xmm1, %xmm1
kusano 2b45e8
	movaps	%xmm1, -15 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-11 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	 -8 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x93, %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -12 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x39, %xmm3, %xmm3
kusano 2b45e8
	movaps	%xmm3, -11 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCHW
kusano 2b45e8
	PREFETCHW (PREFETCHSIZE + 64) - PREOFFSET(Y)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	movaps	 -7 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	 -4 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x93, %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0,  -8 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x39, %xmm1, %xmm1
kusano 2b45e8
	movaps	%xmm1,  -7 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	 -3 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	  0 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x93, %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2,  -4 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x39, %xmm3, %xmm3
kusano 2b45e8
	movaps	%xmm3,  -3 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	subl	$-32 * SIZE, X
kusano 2b45e8
	subl	$-32 * SIZE, Y
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L41
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L43:
kusano 2b45e8
	testl	$16, M
kusano 2b45e8
	jle	.L44
kusano 2b45e8
kusano 2b45e8
	movaps	-31 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x93, %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x39, %xmm1, %xmm1
kusano 2b45e8
	movaps	%xmm1, -31 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-27 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x93, %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -28 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x39, %xmm3, %xmm3
kusano 2b45e8
	movaps	%xmm3, -27 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-23 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-20 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x93, %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -24 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x39, %xmm1, %xmm1
kusano 2b45e8
	movaps	%xmm1, -23 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-19 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-16 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x93, %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -20 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x39, %xmm3, %xmm3
kusano 2b45e8
	movaps	%xmm3, -19 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$16 * SIZE, X
kusano 2b45e8
	addl	$16 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L44:
kusano 2b45e8
	testl	$8, M
kusano 2b45e8
	jle	.L45
kusano 2b45e8
kusano 2b45e8
	movaps	-31 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x93, %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x39, %xmm1, %xmm1
kusano 2b45e8
	movaps	%xmm1, -31 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	-27 * SIZE(X), %xmm0
kusano 2b45e8
	movaps	-24 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, %xmm2
kusano 2b45e8
	shufps	$0x93, %xmm0, %xmm2
kusano 2b45e8
	movaps	%xmm2, -28 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, %xmm3
kusano 2b45e8
	shufps	$0x39, %xmm3, %xmm3
kusano 2b45e8
	movaps	%xmm3, -27 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE, X
kusano 2b45e8
	addl	$8 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L45:
kusano 2b45e8
	testl	$4, M
kusano 2b45e8
	jle	.L46
kusano 2b45e8
kusano 2b45e8
	movaps	-31 * SIZE(X), %xmm2
kusano 2b45e8
	movaps	-28 * SIZE(Y), %xmm3
kusano 2b45e8
kusano 2b45e8
	movss	%xmm2, %xmm0
kusano 2b45e8
	shufps	$0x93, %xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movss	%xmm3, %xmm1
kusano 2b45e8
	shufps	$0x39, %xmm1, %xmm1
kusano 2b45e8
	movaps	%xmm1, -31 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	movaps	%xmm2, %xmm0
kusano 2b45e8
	movaps	%xmm3, %xmm1
kusano 2b45e8
kusano 2b45e8
	addl	$4 * SIZE, X
kusano 2b45e8
	addl	$4 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L46:
kusano 2b45e8
	movsd	-31 * SIZE(X), %xmm2
kusano 2b45e8
kusano 2b45e8
	PSHUFD2($0x39, %xmm1, %xmm1)
kusano 2b45e8
	movlps	%xmm1, -31 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	PSHUFD1($0xff, %xmm0)
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movlps	%xmm2, -31 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	addl	$3 * SIZE, X
kusano 2b45e8
	addl	$3 * SIZE, Y
kusano 2b45e8
kusano 2b45e8
	testl	$2, M
kusano 2b45e8
	jle	.L47
kusano 2b45e8
kusano 2b45e8
	movsd	-32 * SIZE(X), %xmm0
kusano 2b45e8
	movsd	-32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movlps	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movlps	%xmm1, -32 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	addl	$2 * SIZE, X
kusano 2b45e8
	addl	$2 * SIZE, Y
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L47:
kusano 2b45e8
	testl	$1, M
kusano 2b45e8
	jle	.L49
kusano 2b45e8
kusano 2b45e8
	movss	-32 * SIZE(X), %xmm0
kusano 2b45e8
	movss	-32 * SIZE(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm0, -32 * SIZE(Y)
kusano 2b45e8
	movss	%xmm1, -32 * SIZE(X)
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L49:
kusano 2b45e8
	popl	%ebx
kusano 2b45e8
	popl	%esi
kusano 2b45e8
	popl	%edi
kusano 2b45e8
	popl	%ebp
kusano 2b45e8
	ret
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L50:
kusano 2b45e8
	movl	M,  %eax
kusano 2b45e8
	sarl	$3, %eax
kusano 2b45e8
	jle	.L55
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L51:
kusano 2b45e8
	movss	(X), %xmm0
kusano 2b45e8
	movss	(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, (X)
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	movss	%xmm0, (Y)
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	movss	(X), %xmm0
kusano 2b45e8
	movss	(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, (X)
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	movss	%xmm0, (Y)
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	movss	(X), %xmm0
kusano 2b45e8
	movss	(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, (X)
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	movss	%xmm0, (Y)
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	movss	(X), %xmm0
kusano 2b45e8
	movss	(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, (X)
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	movss	%xmm0, (Y)
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	movss	(X), %xmm0
kusano 2b45e8
	movss	(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, (X)
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	movss	%xmm0, (Y)
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	movss	(X), %xmm0
kusano 2b45e8
	movss	(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, (X)
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	movss	%xmm0, (Y)
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	movss	(X), %xmm0
kusano 2b45e8
	movss	(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, (X)
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	movss	%xmm0, (Y)
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	movss	(X), %xmm0
kusano 2b45e8
	movss	(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, (X)
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	movss	%xmm0, (Y)
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L51
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L55:
kusano 2b45e8
	movl	M,  %eax
kusano 2b45e8
	andl	$7, %eax
kusano 2b45e8
	jle	.L57
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L56:
kusano 2b45e8
	movss	(X), %xmm0
kusano 2b45e8
	movss	(Y), %xmm1
kusano 2b45e8
kusano 2b45e8
	movss	%xmm1, (X)
kusano 2b45e8
	movss	%xmm0, (Y)
kusano 2b45e8
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L56
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L57:
kusano 2b45e8
	popl	%ebx
kusano 2b45e8
	popl	%esi
kusano 2b45e8
	popl	%edi
kusano 2b45e8
	popl	%ebp
kusano 2b45e8
	ret
kusano 2b45e8
kusano 2b45e8
	EPILOGUE