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	12
kusano 2b45e8
#define ARGS	 0
kusano 2b45e8
	
kusano 2b45e8
#define STACK_M		 4 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_ALPHA_R	16 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_ALPHA_I	32 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_X		48 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_INCX	52 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_Y		56 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_INCY	60 + STACK + ARGS(%esp)
kusano 2b45e8
kusano 2b45e8
#define M	%ebx
kusano 2b45e8
#define X	%esi
kusano 2b45e8
#define INCX	%ecx
kusano 2b45e8
#define Y	%edi
kusano 2b45e8
#define INCY	%edx
kusano 2b45e8
kusano 2b45e8
#ifndef CONJ
kusano 2b45e8
#define ADD1	fsubrp
kusano 2b45e8
#define ADD2	faddp
kusano 2b45e8
#else
kusano 2b45e8
#define ADD1	faddp
kusano 2b45e8
#define ADD2	fsubrp
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
kusano 2b45e8
	pushl	%edi
kusano 2b45e8
	pushl	%esi
kusano 2b45e8
	pushl	%ebx
kusano 2b45e8
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
#if defined(F_INTERFACE_GFORT) || defined(F_INTERFACE_G95)
kusano 2b45e8
	EMMS
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	FLD	STACK_ALPHA_I
kusano 2b45e8
	FLD	STACK_ALPHA_R
kusano 2b45e8
kusano 2b45e8
	movl	STACK_M,    M
kusano 2b45e8
	movl	STACK_X,    X
kusano 2b45e8
	movl	STACK_INCX, INCX
kusano 2b45e8
	movl	STACK_Y,    Y
kusano 2b45e8
	movl	STACK_INCY, INCY
kusano 2b45e8
kusano 2b45e8
	sall	$ZBASE_SHIFT, INCX
kusano 2b45e8
	sall	$ZBASE_SHIFT, INCY
kusano 2b45e8
	
kusano 2b45e8
	testl	M, M
kusano 2b45e8
	jle	.L40
kusano 2b45e8
kusano 2b45e8
	cmpl	$2 * SIZE, INCX
kusano 2b45e8
	jne	.L14
kusano 2b45e8
	cmpl	$2 * SIZE, INCY
kusano 2b45e8
	jne	.L14
kusano 2b45e8
kusano 2b45e8
	movl	M,  %eax
kusano 2b45e8
	sarl	$2, %eax
kusano 2b45e8
	jle	.L15
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L16:
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	ADD1	%st, %st(1)
kusano 2b45e8
	FLD	0 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	ADD2	%st, %st(1)
kusano 2b45e8
	FLD	1 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	1 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	2 * SIZE(X)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	3 * SIZE(X)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	ADD1	%st, %st(1)
kusano 2b45e8
	FLD	2 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	2 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	2 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	FLD	3 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	ADD2	%st, %st(1)
kusano 2b45e8
	FLD	3 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	3 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	4 * SIZE(X)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	5 * SIZE(X)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	ADD1	%st, %st(1)
kusano 2b45e8
	FLD	4 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	4 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	4 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	FLD	5 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	ADD2	%st, %st(1)
kusano 2b45e8
	FLD	5 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	5 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	6 * SIZE(X)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	7 * SIZE(X)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	ADD1	%st, %st(1)
kusano 2b45e8
	FLD	6 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	6 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	6 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	FLD	7 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	ADD2	%st, %st(1)
kusano 2b45e8
	FLD	7 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	7 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
#ifdef HAVE_3DNOW
kusano 2b45e8
	prefetch	20 * SIZE(X)
kusano 2b45e8
	prefetchw	20 * SIZE(Y)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE, X
kusano 2b45e8
	addl	$8 * SIZE, Y
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L16
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L15:
kusano 2b45e8
	movl	M,  %eax
kusano 2b45e8
	andl	$3, %eax
kusano 2b45e8
	jle	.L40
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L22:
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	ADD1	%st, %st(1)
kusano 2b45e8
	FLD	0 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	ADD2	%st, %st(1)
kusano 2b45e8
	FLD	1 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	1 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	addl	$2 * SIZE, X
kusano 2b45e8
	addl	$2 * SIZE, Y
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L22
kusano 2b45e8
	jmp	.L40
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L14:
kusano 2b45e8
	movl	M, %eax
kusano 2b45e8
	sarl	$2,   %eax
kusano 2b45e8
	jle	.L28
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L29:
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	ADD1	%st, %st(1)
kusano 2b45e8
	FLD	0 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	ADD2	%st, %st(1)
kusano 2b45e8
	FLD	1 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	1 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	ADD1	%st, %st(1)
kusano 2b45e8
	FLD	0 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	ADD2	%st, %st(1)
kusano 2b45e8
	FLD	1 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	1 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	ADD1	%st, %st(1)
kusano 2b45e8
	FLD	0 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	ADD2	%st, %st(1)
kusano 2b45e8
	FLD	1 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	1 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	ADD1	%st, %st(1)
kusano 2b45e8
	FLD	0 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	ADD2	%st, %st(1)
kusano 2b45e8
	FLD	1 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	1 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L29
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L28:
kusano 2b45e8
	movl	M,  %eax
kusano 2b45e8
	andl	$3, %eax
kusano 2b45e8
	jle	.L40
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L35:
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	ADD1	%st, %st(1)
kusano 2b45e8
	FLD	0 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	0 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	FLD	0 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	FLD	1 * SIZE(X)
kusano 2b45e8
	fmul	%st(2), %st
kusano 2b45e8
	ADD2	%st, %st(1)
kusano 2b45e8
	FLD	1 * SIZE(Y)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	FST	1 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L35
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L40:
kusano 2b45e8
	ffreep	%st(0)
kusano 2b45e8
	ffreep	%st(0)
kusano 2b45e8
	xorl	%eax,%eax
kusano 2b45e8
	popl	%ebx
kusano 2b45e8
	popl	%esi
kusano 2b45e8
	popl	%edi
kusano 2b45e8
	ret
kusano 2b45e8
kusano 2b45e8
	EPILOGUE