seal 0.1.0

data/mpg123/src/libmpg123/compat.c

@@ -0,0 +1,138 @@
+/*
+	compat: Some compatibility functions.
+
+	The mpg123 code is determined to keep it's legacy. A legacy of old, old UNIX.
+	So anything possibly somewhat advanced should be considered to be put here, with proper #ifdef;-)
+
+	copyright 2007-8 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 Thomas Orgis, Windows Unicode stuff by JonY.
+*/
+
+#include "config.h"
+#include "compat.h"
+
+#ifdef _MSC_VER
+#include <io.h>
+#else
+#include <fcntl.h>
+#endif
+#include <sys/stat.h>
+
+#ifdef WANT_WIN32_UNICODE
+#include <wchar.h>
+#include <windows.h>
+#include <winnls.h>
+#endif
+
+#include "debug.h"
+
+/* A safe realloc also for very old systems where realloc(NULL, size) returns NULL. */
+void *safe_realloc(void *ptr, size_t size)
+{
+	if(ptr == NULL) return malloc(size);
+	else return realloc(ptr, size);
+}
+
+#ifndef HAVE_STRERROR
+const char *strerror(int errnum)
+{
+	extern int sys_nerr;
+	extern char *sys_errlist[];
+
+	return (errnum < sys_nerr) ?  sys_errlist[errnum]  :  "";
+}
+#endif
+
+#ifndef HAVE_STRDUP
+char *strdup(const char *src)
+{
+	char *dest;
+
+	if (!(dest = (char *) malloc(strlen(src)+1)))
+	return NULL;
+	else
+	return strcpy(dest, src);
+}
+#endif
+
+int compat_open(const char *filename, int flags)
+{
+	int ret;
+#if defined (WANT_WIN32_UNICODE)
+	wchar_t *frag = NULL;
+
+	ret = win32_utf8_wide(filename, &frag, NULL);
+	if ((frag == NULL) || (ret == 0)) goto fallback; /* Fallback to plain open when ucs-2 conversion fails */
+
+	ret = _wopen(frag, flags); /*Try _wopen */
+	if (ret != -1 ) goto open_ok; /* msdn says -1 means failure */
+
+fallback:
+#endif
+
+#if (defined(WIN32) && !defined (__CYGWIN__)) /* MSDN says POSIX function is deprecated beginning in Visual C++ 2005 */
+	ret = _open(filename, flags); /* Try plain old _open(), if it fails, do nothing */
+#else
+	/* On UNIX, we always add a default permission mask in case flags|O_CREAT. */
+	ret = open(filename, flags, S_IRUSR|S_IWUSR|S_IRGRP|S_IWGRP|S_IROTH|S_IWOTH);
+#endif
+
+#if defined (WANT_WIN32_UNICODE)
+open_ok:
+	free ((void *)frag); /* Freeing a NULL should be OK */
+#endif
+
+	return ret;
+}
+
+int compat_close(int infd)
+{
+#if (defined(WIN32) && !defined (__CYGWIN__)) /* MSDN says POSIX function is deprecated beginning in Visual C++ 2005 */
+	return _close(infd);
+#else
+	return close(infd);
+#endif
+}
+
+/* Windows Unicode stuff */
+
+#ifdef WANT_WIN32_UNICODE
+int win32_wide_utf8(const wchar_t * const wptr, char **mbptr, size_t * buflen)
+{
+  size_t len;
+  char *buf;
+  int ret = 0;
+
+  len = WideCharToMultiByte(CP_UTF8, 0, wptr, -1, NULL, 0, NULL, NULL); /* Get utf-8 string length */
+  buf = calloc(len + 1, sizeof (char)); /* Can we assume sizeof char always = 1? */
+
+  if(!buf) len = 0;
+  else {
+    if (len != 0) ret = WideCharToMultiByte(CP_UTF8, 0, wptr, -1, buf, len, NULL, NULL); /*Do actual conversion*/
+    buf[len] = '0'; /* Must terminate */
+  }
+  *mbptr = buf; /* Set string pointer to allocated buffer */
+  if(buflen != NULL) *buflen = (len) * sizeof (char); /* Give length of allocated memory if needed. */
+  return ret;
+}
+
+int win32_utf8_wide(const char *const mbptr, wchar_t **wptr, size_t *buflen)
+{
+  size_t len;
+  wchar_t *buf;
+  int ret = 0;
+
+  len = MultiByteToWideChar(CP_UTF8, MB_ERR_INVALID_CHARS, mbptr, -1, NULL, 0); /* Get converted size */
+  buf = calloc(len + 1, sizeof (wchar_t)); /* Allocate memory accordingly */
+
+  if(!buf) len = 0;
+  else {
+    if (len != 0) ret = MultiByteToWideChar (CP_UTF8, MB_ERR_INVALID_CHARS, mbptr, -1, buf, len); /* Do conversion */
+    buf[len] = L'0'; /* Must terminate */
+  }
+  *wptr = buf; /* Set string pointer to allocated buffer */
+  if (buflen != NULL) *buflen = len * sizeof (wchar_t); /* Give length of allocated memory if needed. */
+  return ret; /* Number of characters written */
+}
+#endif

data/mpg123/src/libmpg123/compat.h

@@ -0,0 +1,178 @@
+/*
+	compat: Some compatibility functions and header inclusions.
+	Basic standard C stuff, that may barely be above/around C89.
+
+	The mpg123 code is determined to keep it's legacy. A legacy of old, old UNIX.
+	It is envisioned to include this compat header instead of any of the "standard" headers, to catch compatibility issues.
+	So, don't include stdlib.h or string.h ... include compat.h.
+
+	copyright 2007-8 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 Thomas Orgis
+*/
+
+#ifndef MPG123_COMPAT_H
+#define MPG123_COMPAT_H
+
+#include "config.h"
+#include "intsym.h"
+
+#ifdef HAVE_STDLIB_H
+/* realloc, size_t */
+#include <stdlib.h>
+#endif
+
+#include        <stdio.h>
+#include        <math.h>
+
+#ifdef HAVE_SIGNAL_H
+#include <signal.h>
+#else
+#ifdef HAVE_SYS_SIGNAL_H
+#include <sys/signal.h>
+#endif
+#endif
+
+#ifdef HAVE_UNISTD_H
+#include <unistd.h>
+#endif
+
+/* Types, types, types. */
+/* Do we actually need these two in addition to sys/types.h? As replacement? */
+#ifdef HAVE_SYS_TYPES_H
+#include <sys/types.h>
+#endif
+#ifdef HAVE_INTTYPES_H
+#include <inttypes.h>
+#endif
+#ifdef HAVE_STDINT_H
+#include <stdint.h>
+#endif
+/* We want SIZE_MAX, etc. */
+#ifdef HAVE_LIMITS_H
+#include <limits.h>
+#endif
+ 
+#ifndef SIZE_MAX
+#define SIZE_MAX ((size_t)-1)
+#endif
+#ifndef ULONG_MAX
+#define ULONG_MAX ((unsigned long)-1)
+#endif
+
+#ifdef HAVE_STRING_H
+#include <string.h>
+#endif
+
+#ifdef OS2
+#include <float.h>
+#endif
+
+#ifdef HAVE_SYS_TIME_H
+#include <sys/time.h>
+#endif
+/* For select(), I need select.h according to POSIX 2001, else: sys/time.h sys/types.h unistd.h */
+#ifdef HAVE_SYS_SELECT_H
+#include <sys/select.h>
+#endif
+
+/* To parse big numbers... */
+#ifdef HAVE_ATOLL
+#define atobigint atoll
+#else
+#define atobigint atol
+#endif
+
+typedef unsigned char byte;
+
+/* A safe realloc also for very old systems where realloc(NULL, size) returns NULL. */
+void *safe_realloc(void *ptr, size_t size);
+#ifndef HAVE_STRERROR
+const char *strerror(int errnum);
+#endif
+
+#ifndef HAVE_STRDUP
+char *strdup(const char *s);
+#endif
+
+/* If we have the size checks enabled, try to derive some sane printfs.
+   Simple start: Use max integer type and format if long is not big enough.
+   I am hesitating to use %ll without making sure that it's there... */
+#if !(defined PLAIN_C89) && (defined SIZEOF_OFF_T) && (SIZEOF_OFF_T > SIZEOF_LONG) && (defined PRIiMAX)
+# define OFF_P PRIiMAX
+typedef intmax_t off_p;
+#else
+# define OFF_P "li"
+typedef long off_p;
+#endif
+
+#if !(defined PLAIN_C89) && (defined SIZEOF_SIZE_T) && (SIZEOF_SIZE_T > SIZEOF_LONG) && (defined PRIuMAX)
+# define SIZE_P PRIuMAX
+typedef uintmax_t size_p;
+#else
+# define SIZE_P "lu"
+typedef unsigned long size_p;
+#endif
+
+#if !(defined PLAIN_C89) && (defined SIZEOF_SSIZE_T) && (SIZEOF_SSIZE_T > SIZEOF_LONG) && (defined PRIiMAX)
+# define SSIZE_P PRIuMAX
+typedef intmax_t ssize_p;
+#else
+# define SSIZE_P "li"
+typedef long ssize_p;
+#endif
+
+/**
+ * Opening a file handle can be different.
+ * This function here is defined to take a path in native encoding (ISO8859 / UTF-8 / ...), or, when MS Windows Unicode support is enabled, an UTF-8 string that will be converted back to native UCS-2 (wide character) before calling the system's open function.
+ * @param[in] wptr Pointer to wide string.
+ * @param[in] mbptr Pointer to multibyte string.
+ * @return file descriptor (>=0) or error code.
+ */
+int compat_open(const char *filename, int flags);
+
+/**
+ * Closing a file handle can be platform specific.
+ * This function takes a file descriptor that is to be closed.
+ * @param[in] infd File descriptor to be closed.
+ * @return 0 if the file was successfully closed. A return value of -1 indicates an error.
+ */
+int compat_close(int infd);
+
+/* Those do make sense in a separate file, but I chose to include them in compat.c because that's the one source whose object is shared between mpg123 and libmpg123 -- and both need the functionality internally. */
+
+#ifdef WANT_WIN32_UNICODE
+/**
+ * win32_uni2mbc
+ * Converts a null terminated UCS-2 string to a multibyte (UTF-8) equivalent.
+ * Caller is supposed to free allocated buffer.
+ * @param[in] wptr Pointer to wide string.
+ * @param[out] mbptr Pointer to multibyte string.
+ * @param[out] buflen Optional parameter for length of allocated buffer.
+ * @return status of WideCharToMultiByte conversion.
+ *
+ * WideCharToMultiByte - http://msdn.microsoft.com/en-us/library/dd374130(VS.85).aspx
+ */
+int win32_wide_utf8(const wchar_t * const wptr, char **mbptr, size_t * buflen);
+
+/**
+ * win32_mbc2uni
+ * Converts a null terminated UTF-8 string to a UCS-2 equivalent.
+ * Caller is supposed to free allocated buffer.
+ * @param[out] mbptr Pointer to multibyte string.
+ * @param[in] wptr Pointer to wide string.
+ * @param[out] buflen Optional parameter for length of allocated buffer.
+ * @return status of WideCharToMultiByte conversion.
+ *
+ * MultiByteToWideChar - http://msdn.microsoft.com/en-us/library/dd319072(VS.85).aspx
+ */
+
+int win32_utf8_wide(const char *const mbptr, wchar_t **wptr, size_t *buflen);
+#endif
+
+/* That one comes from Tellie on OS/2, needed in resolver. */
+#ifdef __KLIBC__
+typedef int socklen_t;
+#endif
+
+#endif

data/mpg123/src/libmpg123/dct36_3dnow.S

@@ -0,0 +1,505 @@
+/*
+	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
+	subl $120,%esp
+	pushl %esi
+	pushl %ebx
+	movl 8(%ebp),%eax
+	movl 12(%ebp),%esi
+	movl 16(%ebp),%ecx
+	movl 20(%ebp),%edx
+	movl 24(%ebp),%ebx
+	leal -128(%ebp),%esp
+
+	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 ASM_NAME(COS9)+12,%mm2
+	punpckldq %mm2,%mm2
+	movd ASM_NAME(COS9)+24,%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 ASM_NAME(COS9)+4,%mm3
+	punpckldq %mm3,%mm3
+	pfmul %mm3,%mm2
+	pfadd %mm0,%mm2
+	movq 40(%eax),%mm3
+	movd ASM_NAME(COS9)+20,%mm4
+	punpckldq %mm4,%mm4
+	pfmul %mm4,%mm3
+	pfadd %mm3,%mm2
+	movq 56(%eax),%mm3
+	movd ASM_NAME(COS9)+28,%mm4
+	punpckldq %mm4,%mm4
+	pfmul %mm4,%mm3
+	pfadd %mm3,%mm2
+	movq (%eax),%mm3
+	movq 16(%eax),%mm4
+	movd ASM_NAME(COS9)+8,%mm5
+	punpckldq %mm5,%mm5
+	pfmul %mm5,%mm4
+	pfadd %mm4,%mm3
+	movq 32(%eax),%mm4
+	movd ASM_NAME(COS9)+16,%mm5
+	punpckldq %mm5,%mm5
+	pfmul %mm5,%mm4
+	pfadd %mm4,%mm3
+	pfadd %mm1,%mm3
+	movq 64(%eax),%mm4
+	movd ASM_NAME(COS9)+32,%mm5
+	punpckldq %mm5,%mm5
+	pfmul %mm5,%mm4
+	pfadd %mm4,%mm3
+	movq %mm2,%mm4
+	pfadd %mm3,%mm4
+	movq %mm7,%mm5
+	punpckldq ASM_NAME(tfcos36)+0,%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 ASM_NAME(tfcos36)+32,%mm5
+	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 ASM_NAME(COS9)+12,%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 ASM_NAME(COS9)+24,%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
+	punpckldq ASM_NAME(tfcos36)+4,%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 ASM_NAME(tfcos36)+28,%mm5
+	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 ASM_NAME(COS9)+20,%mm3
+	punpckldq %mm3,%mm3
+	pfmul %mm3,%mm2
+	pfsub %mm0,%mm2
+	movq 40(%eax),%mm3
+	movd ASM_NAME(COS9)+28,%mm4
+	punpckldq %mm4,%mm4
+	pfmul %mm4,%mm3
+	pfsub %mm3,%mm2
+	movq 56(%eax),%mm3
+	movd ASM_NAME(COS9)+4,%mm4
+	punpckldq %mm4,%mm4
+	pfmul %mm4,%mm3
+	pfadd %mm3,%mm2
+	movq (%eax),%mm3
+	movq 16(%eax),%mm4
+	movd ASM_NAME(COS9)+32,%mm5
+	punpckldq %mm5,%mm5
+	pfmul %mm5,%mm4
+	pfsub %mm4,%mm3
+	movq 32(%eax),%mm4
+	movd ASM_NAME(COS9)+8,%mm5
+	punpckldq %mm5,%mm5
+	pfmul %mm5,%mm4
+	pfsub %mm4,%mm3
+	pfadd %mm1,%mm3
+	movq 64(%eax),%mm4
+	movd ASM_NAME(COS9)+16,%mm5
+	punpckldq %mm5,%mm5
+	pfmul %mm5,%mm4
+	pfadd %mm4,%mm3
+	movq %mm2,%mm4
+	pfadd %mm3,%mm4
+	movq %mm7,%mm5
+	punpckldq ASM_NAME(tfcos36)+8,%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 ASM_NAME(tfcos36)+24,%mm5
+	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 ASM_NAME(COS9)+28,%mm3
+	punpckldq %mm3,%mm3
+	pfmul %mm3,%mm2
+	pfsub %mm0,%mm2
+	movq 40(%eax),%mm3
+	movd ASM_NAME(COS9)+4,%mm4
+	punpckldq %mm4,%mm4
+	pfmul %mm4,%mm3
+	pfadd %mm3,%mm2
+	movq 56(%eax),%mm3
+	movd ASM_NAME(COS9)+20,%mm4
+	punpckldq %mm4,%mm4
+	pfmul %mm4,%mm3
+	pfsub %mm3,%mm2
+	movq (%eax),%mm3
+	movq 16(%eax),%mm4
+	movd ASM_NAME(COS9)+16,%mm5
+	punpckldq %mm5,%mm5
+	pfmul %mm5,%mm4
+	pfsub %mm4,%mm3
+	movq 32(%eax),%mm4
+	movd ASM_NAME(COS9)+32,%mm5
+	punpckldq %mm5,%mm5
+	pfmul %mm5,%mm4
+	pfadd %mm4,%mm3
+	pfadd %mm1,%mm3
+	movq 64(%eax),%mm4
+	movd ASM_NAME(COS9)+8,%mm5
+	punpckldq %mm5,%mm5
+	pfmul %mm5,%mm4
+	pfsub %mm4,%mm3
+	movq %mm2,%mm4
+	pfadd %mm3,%mm4
+	movq %mm7,%mm5
+	punpckldq ASM_NAME(tfcos36)+12,%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 ASM_NAME(tfcos36)+20,%mm5
+	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
+	punpckldq ASM_NAME(tfcos36)+16,%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
+	popl %ebx
+	popl %esi
+	movl %ebp,%esp
+	popl %ebp
+	ret
+
+NONEXEC_STACK