/*

	dct64_3dnow.s: Replacement of dct36() with AMD's 3DNow! SIMD operations support

	copyright ?-2006 by the mpg123 project - free software under the terms of the LGPL 2.1

	see COPYING and AUTHORS files in distribution or http://mpg123.org

	initially written by Syuuhei Kashiyama

	This code based 'dct36_3dnow.s' by Syuuhei Kashiyama

	<squash@mb.kcom.ne.jp>,only two types of changes have been made:

	- remove PREFETCH instruction for speedup

	- change function name for support 3DNow! automatic detect

	You can find Kashiyama's original 3dnow! support patch

	(for mpg123-0.59o) at

	http://user.ecc.u-tokyo.ac.jp/~g810370/linux-simd/ (Japanese).

	by KIMURA Takuhiro <kim@hannah.ipc.miyakyo-u.ac.jp> - until 31.Mar.1999

	                   <kim@comtec.co.jp>               - after  1.Apr.1999

	Replacement of dct36() with AMD's 3DNow! SIMD operations support

	Syuuhei Kashiyama <squash@mb.kcom.ne.jp>

	The author of this program disclaim whole expressed or implied

	warranties with regard to this program, and in no event shall the

	author of this program liable to whatever resulted from the use of

	this program. Use it at your own risk.

*/

#include "mangle.h"

	.globl ASM_NAME(dct36_3dnow)

/*	.type	 ASM_NAME(dct36_3dnow),@function */

ASM_NAME(dct36_3dnow):

	pushl %ebp

	movl %esp,%ebp

#if defined(PIC) && defined(__APPLE__)

	sub $4,%esp

#endif

	pushl %esi

	pushl %ebx

	#undef _EBX_

	#define _EBX_ %edi

	PREPARE_GOT

	GET_GOT

#if defined(PIC) && defined(__APPLE__)

	#define _COS9_ 0(%edi)

	#define _tfcos36_ 0(%eax)

	mov GLOBAL_VAR_PTR(tfcos36), %eax

	mov GLOBAL_VAR_PTR(COS9), %edi

	mov %eax, -4(%ebp)

#else

	#define _COS9_ GLOBAL_VAR(COS9)

	#define _tfcos36_ GLOBAL_VAR(tfcos36)

#endif

	movl 8(%ebp),%eax

	movl 12(%ebp),%esi

	movl 16(%ebp),%ecx

	movl 20(%ebp),%edx

	movl 24(%ebp),%ebx

	femms

	movq (%eax),%mm0

	movq 4(%eax),%mm1

	pfadd %mm1,%mm0

	movq %mm0,4(%eax)

	psrlq $32,%mm1

	movq 12(%eax),%mm2

	punpckldq %mm2,%mm1

	pfadd %mm2,%mm1

	movq %mm1,12(%eax)

	psrlq $32,%mm2

	movq 20(%eax),%mm3

	punpckldq %mm3,%mm2

	pfadd %mm3,%mm2

	movq %mm2,20(%eax)

	psrlq $32,%mm3

	movq 28(%eax),%mm4

	punpckldq %mm4,%mm3

	pfadd %mm4,%mm3

	movq %mm3,28(%eax)

	psrlq $32,%mm4

	movq 36(%eax),%mm5

	punpckldq %mm5,%mm4

	pfadd %mm5,%mm4

	movq %mm4,36(%eax)

	psrlq $32,%mm5

	movq 44(%eax),%mm6

	punpckldq %mm6,%mm5

	pfadd %mm6,%mm5

	movq %mm5,44(%eax)

	psrlq $32,%mm6

	movq 52(%eax),%mm7

	punpckldq %mm7,%mm6

	pfadd %mm7,%mm6

	movq %mm6,52(%eax)

	psrlq $32,%mm7

	movq 60(%eax),%mm0

	punpckldq %mm0,%mm7

	pfadd %mm0,%mm7

	movq %mm7,60(%eax)

	psrlq $32,%mm0

	movd 68(%eax),%mm1

	pfadd %mm1,%mm0

	movd %mm0,68(%eax)

	movd 4(%eax),%mm0

	movd 12(%eax),%mm1

	punpckldq %mm1,%mm0

	punpckldq 20(%eax),%mm1

	pfadd %mm1,%mm0

	movd %mm0,12(%eax)

	psrlq $32,%mm0

	movd %mm0,20(%eax)

	psrlq $32,%mm1

	movd 28(%eax),%mm2

	punpckldq %mm2,%mm1

	punpckldq 36(%eax),%mm2

	pfadd %mm2,%mm1

	movd %mm1,28(%eax)

	psrlq $32,%mm1

	movd %mm1,36(%eax)

	psrlq $32,%mm2

	movd 44(%eax),%mm3

	punpckldq %mm3,%mm2

	punpckldq 52(%eax),%mm3

	pfadd %mm3,%mm2

	movd %mm2,44(%eax)

	psrlq $32,%mm2

	movd %mm2,52(%eax)

	psrlq $32,%mm3

	movd 60(%eax),%mm4

	punpckldq %mm4,%mm3

	punpckldq 68(%eax),%mm4

	pfadd %mm4,%mm3

	movd %mm3,60(%eax)

	psrlq $32,%mm3

	movd %mm3,68(%eax)

	movq 24(%eax),%mm0

	movq 48(%eax),%mm1

	movd 12+_COS9_,%mm2

	punpckldq %mm2,%mm2

	movd 24+_COS9_,%mm3

	punpckldq %mm3,%mm3

	pfmul %mm2,%mm0

	pfmul %mm3,%mm1

	pushl %eax

	movl $1,%eax

	movd %eax,%mm7

	pi2fd %mm7,%mm7

	popl %eax

	movq 8(%eax),%mm2

	movd 4+_COS9_,%mm3

	punpckldq %mm3,%mm3

	pfmul %mm3,%mm2

	pfadd %mm0,%mm2

	movq 40(%eax),%mm3

	movd 20+_COS9_,%mm4

	punpckldq %mm4,%mm4

	pfmul %mm4,%mm3

	pfadd %mm3,%mm2

	movq 56(%eax),%mm3

	movd 28+_COS9_,%mm4

	punpckldq %mm4,%mm4

	pfmul %mm4,%mm3

	pfadd %mm3,%mm2

	movq (%eax),%mm3

	movq 16(%eax),%mm4

	movd 8+_COS9_,%mm5

	punpckldq %mm5,%mm5

	pfmul %mm5,%mm4

	pfadd %mm4,%mm3

	movq 32(%eax),%mm4

	movd 16+_COS9_,%mm5

	punpckldq %mm5,%mm5

	pfmul %mm5,%mm4

	pfadd %mm4,%mm3

	pfadd %mm1,%mm3

	movq 64(%eax),%mm4

	movd 32+_COS9_,%mm5

	punpckldq %mm5,%mm5

	pfmul %mm5,%mm4

	pfadd %mm4,%mm3

	movq %mm2,%mm4

	pfadd %mm3,%mm4

	movq %mm7,%mm5

#if defined(PIC) && defined(__APPLE__)

	mov -4(%ebp),%eax

#endif

	punpckldq 0+_tfcos36_,%mm5

	pfmul %mm5,%mm4

	movq %mm4,%mm5

	pfacc %mm5,%mm5

	movd 108(%edx),%mm6

	punpckldq 104(%edx),%mm6

	pfmul %mm6,%mm5

	movd %mm5,36(%ecx)

	psrlq $32,%mm5

	movd %mm5,32(%ecx)

	movq %mm4,%mm6

	punpckldq %mm6,%mm5

	pfsub %mm6,%mm5

	punpckhdq %mm5,%mm5

	movd 32(%edx),%mm6

	punpckldq 36(%edx),%mm6

	pfmul %mm6,%mm5

	movd 32(%esi),%mm6

	punpckldq 36(%esi),%mm6

	pfadd %mm6,%mm5

	movd %mm5,1024(%ebx)

	psrlq $32,%mm5

	movd %mm5,1152(%ebx)

	movq %mm3,%mm4

	pfsub %mm2,%mm4

	movq %mm7,%mm5

	punpckldq 32+_tfcos36_,%mm5

#if defined(PIC) && defined(__APPLE__)

	mov 8(%ebp),%eax

#endif

	pfmul %mm5,%mm4

	movq %mm4,%mm5

	pfacc %mm5,%mm5

	movd 140(%edx),%mm6

	punpckldq 72(%edx),%mm6

	pfmul %mm6,%mm5

	movd %mm5,68(%ecx)

	psrlq $32,%mm5

	movd %mm5,0(%ecx)

	movq %mm4,%mm6

	punpckldq %mm6,%mm5

	pfsub %mm6,%mm5

	punpckhdq %mm5,%mm5

	movd 0(%edx),%mm6

	punpckldq 68(%edx),%mm6

	pfmul %mm6,%mm5

	movd 0(%esi),%mm6

	punpckldq 68(%esi),%mm6

	pfadd %mm6,%mm5

	movd %mm5,0(%ebx)

	psrlq $32,%mm5

	movd %mm5,2176(%ebx)

	movq 8(%eax),%mm2

	movq 40(%eax),%mm3

	pfsub %mm3,%mm2

	movq 56(%eax),%mm3

	pfsub %mm3,%mm2

	movd 12+_COS9_,%mm3

	punpckldq %mm3,%mm3

	pfmul %mm3,%mm2

	movq 16(%eax),%mm3

	movq 32(%eax),%mm4

	pfsub %mm4,%mm3

	movq 64(%eax),%mm4

	pfsub %mm4,%mm3

	movd 24+_COS9_,%mm4

	punpckldq %mm4,%mm4

	pfmul %mm4,%mm3

	movq 48(%eax),%mm4

	pfsub %mm4,%mm3

	movq (%eax),%mm4

	pfadd %mm4,%mm3

	movq %mm2,%mm4

	pfadd %mm3,%mm4

	movq %mm7,%mm5

#if defined(PIC) && defined(__APPLE__)

	mov -4(%ebp),%eax

#endif

	punpckldq 4+_tfcos36_,%mm5

	pfmul %mm5,%mm4

	movq %mm4,%mm5

	pfacc %mm5,%mm5

	movd 112(%edx),%mm6

	punpckldq 100(%edx),%mm6

	pfmul %mm6,%mm5

	movd %mm5,40(%ecx)

	psrlq $32,%mm5

	movd %mm5,28(%ecx)

	movq %mm4,%mm6

	punpckldq %mm6,%mm5

	pfsub %mm6,%mm5

	punpckhdq %mm5,%mm5

	movd 28(%edx),%mm6

	punpckldq 40(%edx),%mm6

	pfmul %mm6,%mm5

	movd 28(%esi),%mm6

	punpckldq 40(%esi),%mm6

	pfadd %mm6,%mm5

	movd %mm5,896(%ebx)

	psrlq $32,%mm5

	movd %mm5,1280(%ebx)

	movq %mm3,%mm4

	pfsub %mm2,%mm4

	movq %mm7,%mm5

	punpckldq 28+_tfcos36_,%mm5

#if defined(PIC) && defined(__APPLE__)

	mov 8(%ebp),%eax

#endif

	pfmul %mm5,%mm4

	movq %mm4,%mm5

	pfacc %mm5,%mm5

	movd 136(%edx),%mm6

	punpckldq 76(%edx),%mm6

	pfmul %mm6,%mm5

	movd %mm5,64(%ecx)

	psrlq $32,%mm5

	movd %mm5,4(%ecx)

	movq %mm4,%mm6

	punpckldq %mm6,%mm5

	pfsub %mm6,%mm5

	punpckhdq %mm5,%mm5

	movd 4(%edx),%mm6

	punpckldq 64(%edx),%mm6

	pfmul %mm6,%mm5

	movd 4(%esi),%mm6

	punpckldq 64(%esi),%mm6

	pfadd %mm6,%mm5

	movd %mm5,128(%ebx)

	psrlq $32,%mm5

	movd %mm5,2048(%ebx)

	movq 8(%eax),%mm2

	movd 20+_COS9_,%mm3

	punpckldq %mm3,%mm3

	pfmul %mm3,%mm2

	pfsub %mm0,%mm2

	movq 40(%eax),%mm3

	movd 28+_COS9_,%mm4

	punpckldq %mm4,%mm4

	pfmul %mm4,%mm3

	pfsub %mm3,%mm2

	movq 56(%eax),%mm3

	movd 4+_COS9_,%mm4

	punpckldq %mm4,%mm4

	pfmul %mm4,%mm3

	pfadd %mm3,%mm2

	movq (%eax),%mm3

	movq 16(%eax),%mm4

	movd 32+_COS9_,%mm5

	punpckldq %mm5,%mm5

	pfmul %mm5,%mm4

	pfsub %mm4,%mm3

	movq 32(%eax),%mm4

	movd 8+_COS9_,%mm5

	punpckldq %mm5,%mm5

	pfmul %mm5,%mm4

	pfsub %mm4,%mm3

	pfadd %mm1,%mm3

	movq 64(%eax),%mm4

	movd 16+_COS9_,%mm5

	punpckldq %mm5,%mm5

	pfmul %mm5,%mm4

	pfadd %mm4,%mm3

	movq %mm2,%mm4

	pfadd %mm3,%mm4

	movq %mm7,%mm5

#if defined(PIC) && defined(__APPLE__)

	mov -4(%ebp),%eax

#endif

	punpckldq 8+_tfcos36_,%mm5

	pfmul %mm5,%mm4

	movq %mm4,%mm5

	pfacc %mm5,%mm5

	movd 116(%edx),%mm6

	punpckldq 96(%edx),%mm6

	pfmul %mm6,%mm5

	movd %mm5,44(%ecx)

	psrlq $32,%mm5

	movd %mm5,24(%ecx)

	movq %mm4,%mm6

	punpckldq %mm6,%mm5

	pfsub %mm6,%mm5

	punpckhdq %mm5,%mm5

	movd 24(%edx),%mm6

	punpckldq 44(%edx),%mm6

	pfmul %mm6,%mm5

	movd 24(%esi),%mm6

	punpckldq 44(%esi),%mm6

	pfadd %mm6,%mm5

	movd %mm5,768(%ebx)

	psrlq $32,%mm5

	movd %mm5,1408(%ebx)

	movq %mm3,%mm4

	pfsub %mm2,%mm4

	movq %mm7,%mm5

	punpckldq 24+_tfcos36_,%mm5

#if defined(PIC) && defined(__APPLE__)

	mov 8(%ebp),%eax

#endif

	pfmul %mm5,%mm4

	movq %mm4,%mm5

	pfacc %mm5,%mm5

	movd 132(%edx),%mm6

	punpckldq 80(%edx),%mm6

	pfmul %mm6,%mm5

	movd %mm5,60(%ecx)

	psrlq $32,%mm5

	movd %mm5,8(%ecx)

	movq %mm4,%mm6

	punpckldq %mm6,%mm5

	pfsub %mm6,%mm5

	punpckhdq %mm5,%mm5

	movd 8(%edx),%mm6

	punpckldq 60(%edx),%mm6

	pfmul %mm6,%mm5

	movd 8(%esi),%mm6

	punpckldq 60(%esi),%mm6

	pfadd %mm6,%mm5

	movd %mm5,256(%ebx)

	psrlq $32,%mm5

	movd %mm5,1920(%ebx)

	movq 8(%eax),%mm2

	movd 28+_COS9_,%mm3

	punpckldq %mm3,%mm3

	pfmul %mm3,%mm2

	pfsub %mm0,%mm2

	movq 40(%eax),%mm3

	movd 4+_COS9_,%mm4

	punpckldq %mm4,%mm4

	pfmul %mm4,%mm3

	pfadd %mm3,%mm2

	movq 56(%eax),%mm3

	movd 20+_COS9_,%mm4

	punpckldq %mm4,%mm4

	pfmul %mm4,%mm3

	pfsub %mm3,%mm2

	movq (%eax),%mm3

	movq 16(%eax),%mm4

	movd 16+_COS9_,%mm5

	punpckldq %mm5,%mm5

	pfmul %mm5,%mm4

	pfsub %mm4,%mm3

	movq 32(%eax),%mm4

	movd 32+_COS9_,%mm5

	punpckldq %mm5,%mm5

	pfmul %mm5,%mm4

	pfadd %mm4,%mm3

	pfadd %mm1,%mm3

	movq 64(%eax),%mm4

	movd 8+_COS9_,%mm5

	punpckldq %mm5,%mm5

	pfmul %mm5,%mm4

	pfsub %mm4,%mm3

	movq %mm2,%mm4

	pfadd %mm3,%mm4

	movq %mm7,%mm5

#if defined(PIC) && defined(__APPLE__)

	mov -4(%ebp),%eax

#endif

	punpckldq 12+_tfcos36_,%mm5

	pfmul %mm5,%mm4

	movq %mm4,%mm5

	pfacc %mm5,%mm5

	movd 120(%edx),%mm6

	punpckldq 92(%edx),%mm6

	pfmul %mm6,%mm5

	movd %mm5,48(%ecx)

	psrlq $32,%mm5

	movd %mm5,20(%ecx)

	movq %mm4,%mm6

	punpckldq %mm6,%mm5

	pfsub %mm6,%mm5

	punpckhdq %mm5,%mm5

	movd 20(%edx),%mm6

	punpckldq 48(%edx),%mm6

	pfmul %mm6,%mm5

	movd 20(%esi),%mm6

	punpckldq 48(%esi),%mm6

	pfadd %mm6,%mm5

	movd %mm5,640(%ebx)

	psrlq $32,%mm5

	movd %mm5,1536(%ebx)

	movq %mm3,%mm4

	pfsub %mm2,%mm4

	movq %mm7,%mm5

	punpckldq 20+_tfcos36_,%mm5

#if defined(PIC) && defined(__APPLE__)

	mov 8(%ebp),%eax

#endif

	pfmul %mm5,%mm4

	movq %mm4,%mm5

	pfacc %mm5,%mm5

	movd 128(%edx),%mm6

	punpckldq 84(%edx),%mm6

	pfmul %mm6,%mm5

	movd %mm5,56(%ecx)

	psrlq $32,%mm5

	movd %mm5,12(%ecx)

	movq %mm4,%mm6

	punpckldq %mm6,%mm5

	pfsub %mm6,%mm5

	punpckhdq %mm5,%mm5

	movd 12(%edx),%mm6

	punpckldq 56(%edx),%mm6

	pfmul %mm6,%mm5

	movd 12(%esi),%mm6

	punpckldq 56(%esi),%mm6

	pfadd %mm6,%mm5

	movd %mm5,384(%ebx)

	psrlq $32,%mm5

	movd %mm5,1792(%ebx)

	movq (%eax),%mm4

	movq 16(%eax),%mm3

	pfsub %mm3,%mm4

	movq 32(%eax),%mm3

	pfadd %mm3,%mm4

	movq 48(%eax),%mm3

	pfsub %mm3,%mm4

	movq 64(%eax),%mm3

	pfadd %mm3,%mm4

	movq %mm7,%mm5

#if defined(PIC) && defined(__APPLE__)

	mov -4(%ebp),%eax

#endif

	punpckldq 16+_tfcos36_,%mm5

	pfmul %mm5,%mm4

	movq %mm4,%mm5

	pfacc %mm5,%mm5

	movd 124(%edx),%mm6

	punpckldq 88(%edx),%mm6

	pfmul %mm6,%mm5

	movd %mm5,52(%ecx)

	psrlq $32,%mm5

	movd %mm5,16(%ecx)

	movq %mm4,%mm6

	punpckldq %mm6,%mm5

	pfsub %mm6,%mm5

	punpckhdq %mm5,%mm5

	movd 16(%edx),%mm6

	punpckldq 52(%edx),%mm6

	pfmul %mm6,%mm5

	movd 16(%esi),%mm6

	punpckldq 52(%esi),%mm6

	pfadd %mm6,%mm5

	movd %mm5,512(%ebx)

	psrlq $32,%mm5

	movd %mm5,1664(%ebx)

	femms

	RESTORE_GOT

	popl %ebx

	popl %esi

	movl %ebp,%esp

	popl %ebp

	ret

#if defined(PIC) && defined(__APPLE__)

	.section	__IMPORT,__pointers,non_lazy_symbol_pointers

L_tfcos36:

	.indirect_symbol	ASM_NAME(tfcos36)

	.long	0

L_COS9:

	.indirect_symbol	ASM_NAME(COS9)

	.long	0

#endif

NONEXEC_STACK