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 RESULT		 4 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_N		 8 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_X		12 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_INCX	16 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_Y		20 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_INCY	24 + STACK + ARGS(%esp)
kusano 2b45e8
kusano 2b45e8
#include "l1param.h"
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
#define N	%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
	movl	STACK_N,     N
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
	testl	N, N
kusano 2b45e8
	jle	.L88
kusano 2b45e8
kusano 2b45e8
	sall	$ZBASE_SHIFT, INCX
kusano 2b45e8
	sall	$ZBASE_SHIFT, INCY
kusano 2b45e8
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
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	N,  %eax
kusano 2b45e8
	sarl	$1, %eax
kusano 2b45e8
	jle	.L15
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L16:
kusano 2b45e8
#ifdef PREFETCH
kusano 2b45e8
	PREFETCH (PREFETCHSIZE +  0) - PREOFFSET(X)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 1 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
	FLD	 1 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8
	FLD	 1 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
	FLD	 2 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
#ifdef PREFETCH
kusano 2b45e8
	PREFETCH (PREFETCHSIZE +  0) - PREOFFSET(Y)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	FLD	 2 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 3 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
	FLD	 3 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FLD	 2 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8
	FLD	 3 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8
	addl	$4 * SIZE, X
kusano 2b45e8
	addl	$4 * SIZE, Y
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L16
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L15:
kusano 2b45e8
	movl	N, %eax
kusano 2b45e8
	andl	$1,   %eax
kusano 2b45e8
	jle	.L27
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L22:
kusano 2b45e8
	FLD	 0 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 1 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
	FLD	 1 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8
	FLD	 1 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
	jmp	.L27
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L14:
kusano 2b45e8
	movl	N, %eax
kusano 2b45e8
	sarl	$1,   %eax
kusano 2b45e8
	jle	.L30
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
.L31:
kusano 2b45e8
	FLD	 0 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 1 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
	FLD	 1 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8
	FLD	 1 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(X)
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 1 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
	FLD	 1 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8
	FLD	 1 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
	addl	INCX, X
kusano 2b45e8
	addl	INCY, Y
kusano 2b45e8
kusano 2b45e8
	decl	%eax
kusano 2b45e8
	jg	.L31
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L30:
kusano 2b45e8
	movl	N,  %eax
kusano 2b45e8
	andl	$1, %eax
kusano 2b45e8
	jle	.L27
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L37:
kusano 2b45e8
	FLD	 0 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
kusano 2b45e8
	FLD	 1 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
	FLD	 1 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FLD	 0 * SIZE(Y)
kusano 2b45e8
	fmul	%st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
kusano 2b45e8
	FLD	 1 * SIZE(Y)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
	faddp	%st, %st(4)
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L27:
kusano 2b45e8
	movl	RESULT, %eax
kusano 2b45e8
kusano 2b45e8
#ifndef CONJ
kusano 2b45e8
	fsubp	%st, %st(3)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
#else
kusano 2b45e8
	faddp	%st, %st(3)
kusano 2b45e8
	fsubp	%st, %st(1)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	FST	1 * SIZE(%eax)
kusano 2b45e8
	FST	0 * SIZE(%eax)
kusano 2b45e8
kusano 2b45e8
	popl	%ebx
kusano 2b45e8
	popl	%esi
kusano 2b45e8
	popl	%edi
kusano 2b45e8
	ret
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L88:
kusano 2b45e8
	movl	RESULT, %eax
kusano 2b45e8
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
kusano 2b45e8
	FST	1 * SIZE(%eax)
kusano 2b45e8
	FST	0 * SIZE(%eax)
kusano 2b45e8
kusano 2b45e8
	popl	%ebx
kusano 2b45e8
	popl	%esi
kusano 2b45e8
	popl	%edi
kusano 2b45e8
	ret
kusano 2b45e8
kusano 2b45e8
	EPILOGUE