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 M	ARG1
kusano 2b45e8
#define N	ARG2
kusano 2b45e8
#define K	ARG3
kusano 2b45e8
#define A	ARG4
kusano 2b45e8
#define B	ARG5
kusano 2b45e8
#define C	ARG6
kusano 2b45e8
#define LDC	%r10
kusano 2b45e8
	
kusano 2b45e8
#define I	%r12
kusano 2b45e8
#define J	%r13
kusano 2b45e8
#define AO	%r14
kusano 2b45e8
#define BO	%r15
kusano 2b45e8
#define	CO	%rbp
kusano 2b45e8
kusano 2b45e8
#define OFFSET	48 + STACKSIZE(%rsp)
kusano 2b45e8
kusano 2b45e8
#define STACKSIZE 64
kusano 2b45e8
kusano 2b45e8
#define KK	%r11
kusano 2b45e8
#define AORIG	 48(%rsp)
kusano 2b45e8
	
kusano 2b45e8
#ifdef OPTERON
kusano 2b45e8
#define PREFETCH	prefetch
kusano 2b45e8
#define PREFETCHW	prefetchw
kusano 2b45e8
#else
kusano 2b45e8
#define PREFETCH	prefetcht0
kusano 2b45e8
#define PREFETCHW	prefetcht0
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#define PREFETCHSIZE (5 + 4 * 10)
kusano 2b45e8
kusano 2b45e8
#ifndef CONJ
kusano 2b45e8
#define ADD1	  faddp
kusano 2b45e8
#define ADD2	  fsubrp
kusano 2b45e8
#define ADD3	  faddp
kusano 2b45e8
#define ADD4	  faddp
kusano 2b45e8
#elif defined(LN) || defined(LT)
kusano 2b45e8
#define ADD1	  faddp
kusano 2b45e8
#define ADD2	  faddp
kusano 2b45e8
#define ADD3	  fsubrp
kusano 2b45e8
#define ADD4	  faddp
kusano 2b45e8
#else
kusano 2b45e8
#define ADD1	  faddp
kusano 2b45e8
#define ADD2	  faddp
kusano 2b45e8
#define ADD3	  faddp
kusano 2b45e8
#define ADD4	  fsubrp
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	PROFCODE
kusano 2b45e8
	
kusano 2b45e8
	subq	$STACKSIZE, %rsp
kusano 2b45e8
	movq	%rbx,  0(%rsp)
kusano 2b45e8
	movq	%rbp,  8(%rsp)
kusano 2b45e8
	movq	%r12, 16(%rsp)
kusano 2b45e8
	movq	%r13, 24(%rsp)
kusano 2b45e8
	movq	%r14, 32(%rsp)
kusano 2b45e8
	movq	%r15, 40(%rsp)
kusano 2b45e8
kusano 2b45e8
	movq	40 + STACKSIZE(%rsp), LDC
kusano 2b45e8
kusano 2b45e8
	salq	$ZBASE_SHIFT, LDC
kusano 2b45e8
kusano 2b45e8
	addq	$8 * SIZE, A
kusano 2b45e8
	addq	$8 * SIZE, B
kusano 2b45e8
	
kusano 2b45e8
#ifdef LN
kusano 2b45e8
       movq	M, %rax
kusano 2b45e8
       salq	$ZBASE_SHIFT, %rax
kusano 2b45e8
       addq	%rax, C
kusano 2b45e8
       imulq	K, %rax
kusano 2b45e8
       addq	%rax, A
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef RT
kusano 2b45e8
       movq	N, %rax
kusano 2b45e8
       salq	$ZBASE_SHIFT, %rax
kusano 2b45e8
       imulq	K, %rax
kusano 2b45e8
       addq	%rax, B
kusano 2b45e8
kusano 2b45e8
       movq	N, %rax
kusano 2b45e8
       imulq	LDC, %rax
kusano 2b45e8
       addq	%rax, C
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef RN
kusano 2b45e8
       movq	OFFSET, %rax
kusano 2b45e8
       negq	%rax
kusano 2b45e8
       movq	%rax, KK
kusano 2b45e8
#endif	
kusano 2b45e8
kusano 2b45e8
#ifdef RT
kusano 2b45e8
       movq	N, %rax
kusano 2b45e8
       subq	OFFSET, %rax
kusano 2b45e8
       movq	%rax, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	cmpq	$0, M
kusano 2b45e8
	jle	.L999
kusano 2b45e8
	
kusano 2b45e8
	movq	N,   %rax
kusano 2b45e8
	movq	%rax, J
kusano 2b45e8
	testq	%rax, %rax
kusano 2b45e8
	jle	.L999
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L01:
kusano 2b45e8
#if defined(LT) || defined(RN)
kusano 2b45e8
	movq	A, AO
kusano 2b45e8
#else
kusano 2b45e8
	movq	A, AORIG
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef RT
kusano 2b45e8
	movq	K, %rax
kusano 2b45e8
	salq	$ZBASE_SHIFT, %rax
kusano 2b45e8
	subq	%rax, B
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef RT
kusano 2b45e8
	subq	LDC, C
kusano 2b45e8
#endif
kusano 2b45e8
	movq	C, CO
kusano 2b45e8
#ifndef RT
kusano 2b45e8
	addq	LDC, C
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef LN
kusano 2b45e8
	movq	OFFSET, %rax
kusano 2b45e8
	addq	M, %rax
kusano 2b45e8
	movq	%rax, KK
kusano 2b45e8
#endif	
kusano 2b45e8
kusano 2b45e8
#ifdef LT
kusano 2b45e8
	movq	OFFSET, %rax
kusano 2b45e8
	movq	%rax, KK
kusano 2b45e8
#endif
kusano 2b45e8
	movq	M,  I
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L11:
kusano 2b45e8
#ifdef LN
kusano 2b45e8
       movq	K, %rax
kusano 2b45e8
       salq	$ZBASE_SHIFT, %rax
kusano 2b45e8
       subq	%rax, AORIG
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(LN) || defined(RT)
kusano 2b45e8
	movq	KK, %rax
kusano 2b45e8
	salq	$ZBASE_SHIFT, %rax
kusano 2b45e8
	movq	AORIG, AO
kusano 2b45e8
	leaq	(AO, %rax, 1), AO
kusano 2b45e8
	leaq	(B,  %rax, 1), BO
kusano 2b45e8
#else
kusano 2b45e8
	movq	B, BO
kusano 2b45e8
#endif	
kusano 2b45e8
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
	fldz
kusano 2b45e8
kusano 2b45e8
#if   defined(HAVE_3DNOW)
kusano 2b45e8
	prefetchw	2 * SIZE(CO)
kusano 2b45e8
#elif defined(HAVE_SSE)
kusano 2b45e8
	prefetchnta	2 * SIZE(CO)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(LT) || defined(RN)
kusano 2b45e8
	movq	KK, %rax
kusano 2b45e8
#else
kusano 2b45e8
	movq	K,  %rax
kusano 2b45e8
	subq	KK, %rax
kusano 2b45e8
#endif
kusano 2b45e8
	sarq	$2, %rax
kusano 2b45e8
 	je	.L15
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L12:
kusano 2b45e8
	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(BO)
kusano 2b45e8
	fld	 %st(1)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	ADD1	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(BO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	 
kusano 2b45e8
	FLD	 -7 * SIZE(AO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	ADD2	 %st, %st(6)
kusano 2b45e8
	ADD3	 %st, %st(3)
kusano 2b45e8
	ADD4	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -6 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -6 * SIZE(BO)
kusano 2b45e8
	fld	 %st(1)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	ADD1	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -5 * SIZE(BO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	 
kusano 2b45e8
	FLD	 -5 * SIZE(AO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	ADD2	 %st, %st(6)
kusano 2b45e8
	ADD3	 %st, %st(3)
kusano 2b45e8
	ADD4	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	PREFETCH	(PREFETCHSIZE + 4) * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -4 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -4 * SIZE(BO)
kusano 2b45e8
	fld	 %st(1)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	ADD1	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -3 * SIZE(BO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	 
kusano 2b45e8
	FLD	 -3 * SIZE(AO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	ADD2	 %st, %st(6)
kusano 2b45e8
	ADD3	 %st, %st(3)
kusano 2b45e8
	ADD4	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -2 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -2 * SIZE(BO)
kusano 2b45e8
	fld	 %st(1)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	ADD1	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -1 * SIZE(BO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	 
kusano 2b45e8
	FLD	 -1 * SIZE(AO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	ADD2	 %st, %st(6)
kusano 2b45e8
	ADD3	 %st, %st(3)
kusano 2b45e8
	ADD4	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	addq	$8 * SIZE,AO
kusano 2b45e8
	addq	$8 * SIZE,BO
kusano 2b45e8
kusano 2b45e8
	decq	%rax
kusano 2b45e8
	jne	.L12
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L15:
kusano 2b45e8
#if defined(LT) || defined(RN)
kusano 2b45e8
	movq	KK, %rax
kusano 2b45e8
#else
kusano 2b45e8
	movq	K,  %rax
kusano 2b45e8
	subq	KK, %rax
kusano 2b45e8
#endif
kusano 2b45e8
	and	$3,  %rax
kusano 2b45e8
	je	.L18
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L16:
kusano 2b45e8
	FLD	 -8 * SIZE(AO)
kusano 2b45e8
kusano 2b45e8
	FLD	 -8 * SIZE(BO)
kusano 2b45e8
	fld	 %st(1)
kusano 2b45e8
	fmul	 %st(1), %st
kusano 2b45e8
	ADD1	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	FLD	 -7 * SIZE(BO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	 
kusano 2b45e8
	FLD	 -7 * SIZE(AO)
kusano 2b45e8
	fmul	 %st, %st(2)
kusano 2b45e8
	fmulp	 %st, %st(1)
kusano 2b45e8
kusano 2b45e8
	ADD2	 %st, %st(6)
kusano 2b45e8
	ADD3	 %st, %st(3)
kusano 2b45e8
	ADD4	 %st, %st(3)
kusano 2b45e8
kusano 2b45e8
	addq	$2 * SIZE,AO
kusano 2b45e8
	addq	$2 * SIZE,BO
kusano 2b45e8
kusano 2b45e8
	decq	%rax
kusano 2b45e8
	jne	 .L16
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L18:
kusano 2b45e8
	faddp	%st, %st(3)
kusano 2b45e8
	faddp	%st, %st(1)
kusano 2b45e8
	
kusano 2b45e8
	fxch	%st(1)
kusano 2b45e8
kusano 2b45e8
#if defined(LN) || defined(RT)
kusano 2b45e8
	movq	KK, %rax
kusano 2b45e8
#ifdef LN
kusano 2b45e8
	subq	$1, %rax
kusano 2b45e8
#else
kusano 2b45e8
	subq	$1, %rax
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	salq	$ZBASE_SHIFT, %rax
kusano 2b45e8
kusano 2b45e8
	movq	AORIG, AO
kusano 2b45e8
	leaq	(AO, %rax, 1), AO
kusano 2b45e8
	leaq	(B,  %rax, 1), BO
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(LN) || defined(LT)
kusano 2b45e8
	FLD	-8 * SIZE(BO)
kusano 2b45e8
	fsubp	%st, %st(1)
kusano 2b45e8
	FLD	-7 * SIZE(BO)
kusano 2b45e8
	fsubp	%st, %st(2)
kusano 2b45e8
#else
kusano 2b45e8
	FLD	-8 * SIZE(AO)
kusano 2b45e8
	fsubp	%st, %st(1)
kusano 2b45e8
	FLD	-7 * SIZE(AO)
kusano 2b45e8
	fsubp	%st, %st(2)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(LN) || defined(LT)
kusano 2b45e8
	FLD	-8 * SIZE(AO)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	-8 * SIZE(AO)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	FLD	-7 * SIZE(AO)
kusano 2b45e8
	fmulp	%st, %st(3)
kusano 2b45e8
	FLD	-7 * SIZE(AO)
kusano 2b45e8
	fmulp	%st, %st(4)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(RN) || defined(RT)
kusano 2b45e8
	FLD	-8 * SIZE(BO)
kusano 2b45e8
	fmul	%st(1), %st
kusano 2b45e8
	FLD	-8 * SIZE(BO)
kusano 2b45e8
	fmul	%st(3), %st
kusano 2b45e8
	FLD	-7 * SIZE(BO)
kusano 2b45e8
	fmulp	%st, %st(3)
kusano 2b45e8
	FLD	-7 * SIZE(BO)
kusano 2b45e8
	fmulp	%st, %st(4)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifndef CONJ
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
	fsubp	%st, %st(2)
kusano 2b45e8
#else
kusano 2b45e8
	fsubp	%st, %st(2)
kusano 2b45e8
	faddp	%st, %st(2)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(LN) || defined(LT)
kusano 2b45e8
	fld	%st
kusano 2b45e8
	FST	-7 * SIZE(BO)
kusano 2b45e8
	fxch	%st(1)
kusano 2b45e8
	fld	%st
kusano 2b45e8
	FST	-8 * SIZE(BO)
kusano 2b45e8
#else
kusano 2b45e8
	fld	%st
kusano 2b45e8
	FST	-7 * SIZE(AO)
kusano 2b45e8
	fxch	%st(1)
kusano 2b45e8
	fld	%st
kusano 2b45e8
	FST	-8 * SIZE(AO)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef LN
kusano 2b45e8
	subq	$2 * SIZE, CO
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	FST	0 * SIZE(CO)
kusano 2b45e8
	FST	1 * SIZE(CO)
kusano 2b45e8
	
kusano 2b45e8
#ifndef LN
kusano 2b45e8
	addq	$2 * SIZE, CO
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(LT) || defined(RN)
kusano 2b45e8
	movq	K,  %rax
kusano 2b45e8
	subq	KK, %rax
kusano 2b45e8
	salq	$ZBASE_SHIFT, %rax
kusano 2b45e8
	leaq	(AO, %rax, 1), AO
kusano 2b45e8
	leaq	(BO, %rax, 1), BO
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef LN
kusano 2b45e8
	subq	$1, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef LT
kusano 2b45e8
	addq	$1, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef RT
kusano 2b45e8
       movq	K, %rax
kusano 2b45e8
       salq	$ZBASE_SHIFT, %rax
kusano 2b45e8
       addq	%rax, AORIG
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	decq	I
kusano 2b45e8
	jne	.L11
kusano 2b45e8
kusano 2b45e8
#ifdef LN
kusano 2b45e8
       movq	K, %rax
kusano 2b45e8
       salq	$ZBASE_SHIFT, %rax
kusano 2b45e8
       leaq	(B, %rax, 1), B
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(LT) || defined(RN)
kusano 2b45e8
	movq	BO, B
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef RN
kusano 2b45e8
	addq	$1, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#ifdef RT
kusano 2b45e8
	subq	$1, KK
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	decq	J
kusano 2b45e8
	jne	.L01
kusano 2b45e8
	ALIGN_4
kusano 2b45e8
kusano 2b45e8
.L999:
kusano 2b45e8
	movq	  0(%rsp), %rbx
kusano 2b45e8
	movq	  8(%rsp), %rbp
kusano 2b45e8
	movq	 16(%rsp), %r12
kusano 2b45e8
	movq	 24(%rsp), %r13
kusano 2b45e8
	movq	 32(%rsp), %r14
kusano 2b45e8
	movq	 40(%rsp), %r15
kusano 2b45e8
	addq	$STACKSIZE, %rsp
kusano 2b45e8
	ret
kusano 2b45e8
kusano 2b45e8
	EPILOGUE