remove MMX2 dependency (replace pshufw $78 with pswapd) so it can

be used on a K6-2/3+.
Patch by Zuxy Meng < zuxy PP meng AH gmail PP com >
Original thread:
Date: Jun 21, 2006 2:50 PM
Subject: [MPlayer-dev-eng] [PATCH] Saturation & PSWAPD bugfix in mp3lib/dct64_3dnow.c & mp3lib/dct64_k7.c


git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@18838 b3059339-0415-0410-9bf9-f77b7e298cf2

1 files changed, 32 insertions, 32 deletions

diff --git a/mp3lib/dct64_k7.c b/mp3lib/dct64_k7.c
index ea4cc0a282..1fca3592fd 100644
--- a/mp3lib/dct64_k7.c
+++ b/mp3lib/dct64_k7.c
@@ -32,8 +32,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	8(%%eax), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 120(%%eax), %%mm1\n\t"
-"	pshufw	$78, 112(%%eax), %%mm5\n\t"
+"	pswapd	120(%%eax), %%mm1\n\t"
+"	pswapd	112(%%eax), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, (%%edx)\n\t"
@@ -51,8 +51,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	24(%%eax), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 104(%%eax), %%mm1\n\t"
-"	pshufw	$78, 96(%%eax), %%mm5\n\t"
+"	pswapd	104(%%eax), %%mm1\n\t"
+"	pswapd	96(%%eax), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 16(%%edx)\n\t"
@@ -70,8 +70,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	40(%%eax), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 88(%%eax), %%mm1\n\t"
-"	pshufw	$78, 80(%%eax), %%mm5\n\t"
+"	pswapd	88(%%eax), %%mm1\n\t"
+"	pswapd	80(%%eax), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 32(%%edx)\n\t"
@@ -89,8 +89,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	56(%%eax), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 72(%%eax), %%mm1\n\t"
-"	pshufw	$78, 64(%%eax), %%mm5\n\t"
+"	pswapd	72(%%eax), %%mm1\n\t"
+"	pswapd	64(%%eax), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 48(%%edx)\n\t"
@@ -110,8 +110,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	8(%%edx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 56(%%edx), %%mm1\n\t"
-"	pshufw	$78, 48(%%edx), %%mm5\n\t"
+"	pswapd	56(%%edx), %%mm1\n\t"
+"	pswapd	48(%%edx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, (%%ecx)\n\t"
@@ -129,8 +129,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	24(%%edx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 40(%%edx), %%mm1\n\t"
-"	pshufw	$78, 32(%%edx), %%mm5\n\t"
+"	pswapd	40(%%edx), %%mm1\n\t"
+"	pswapd	32(%%edx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 16(%%ecx)\n\t"
@@ -150,8 +150,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	72(%%edx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 120(%%edx), %%mm1\n\t"
-"	pshufw	$78, 112(%%edx), %%mm5\n\t"
+"	pswapd	120(%%edx), %%mm1\n\t"
+"	pswapd	112(%%edx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 64(%%ecx)\n\t"
@@ -169,8 +169,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	88(%%edx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 104(%%edx), %%mm1\n\t"
-"	pshufw	$78, 96(%%edx), %%mm5\n\t"
+"	pswapd	104(%%edx), %%mm1\n\t"
+"	pswapd	96(%%edx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 80(%%ecx)\n\t"
@@ -193,8 +193,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	8(%%ecx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 24(%%ecx), %%mm1\n\t"
-"	pshufw	$78, 16(%%ecx), %%mm5\n\t"
+"	pswapd	24(%%ecx), %%mm1\n\t"
+"	pswapd	16(%%ecx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, (%%edx)\n\t"
@@ -212,8 +212,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	40(%%ecx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 56(%%ecx), %%mm1\n\t"
-"	pshufw	$78, 48(%%ecx), %%mm5\n\t"
+"	pswapd	56(%%ecx), %%mm1\n\t"
+"	pswapd	48(%%ecx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 32(%%edx)\n\t"
@@ -231,8 +231,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	72(%%ecx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 88(%%ecx), %%mm1\n\t"
-"	pshufw	$78, 80(%%ecx), %%mm5\n\t"
+"	pswapd	88(%%ecx), %%mm1\n\t"
+"	pswapd	80(%%ecx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 64(%%edx)\n\t"
@@ -250,8 +250,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	104(%%ecx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 120(%%ecx), %%mm1\n\t"
-"	pshufw	$78, 112(%%ecx), %%mm5\n\t"
+"	pswapd	120(%%ecx), %%mm1\n\t"
+"	pswapd	112(%%ecx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 96(%%edx)\n\t"
@@ -273,8 +273,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	16(%%edx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 8(%%edx), %%mm1\n\t"
-"	pshufw	$78, 24(%%edx), %%mm5\n\t"
+"	pswapd	8(%%edx), %%mm1\n\t"
+"	pswapd	24(%%edx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, (%%ecx)\n\t"
@@ -292,8 +292,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	48(%%edx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 40(%%edx), %%mm1\n\t"
-"	pshufw	$78, 56(%%edx), %%mm5\n\t"
+"	pswapd	40(%%edx), %%mm1\n\t"
+"	pswapd	56(%%edx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 32(%%ecx)\n\t"
@@ -311,8 +311,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	80(%%edx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 72(%%edx), %%mm1\n\t"
-"	pshufw	$78, 88(%%edx), %%mm5\n\t"
+"	pswapd	72(%%edx), %%mm1\n\t"
+"	pswapd	88(%%edx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 64(%%ecx)\n\t"
@@ -330,8 +330,8 @@ void dct64_MMX_3dnowex(real *a,real *b,real *c)
 "	movq	112(%%edx), %%mm4\n\t"
 "	movq	%%mm0, %%mm3\n\t"
 "	movq	%%mm4, %%mm7\n\t"
-"	pshufw	$78, 104(%%edx), %%mm1\n\t"
-"	pshufw	$78, 120(%%edx), %%mm5\n\t"
+"	pswapd	104(%%edx), %%mm1\n\t"
+"	pswapd	120(%%edx), %%mm5\n\t"
 "	pfadd	%%mm1, %%mm0\n\t"
 "	pfadd	%%mm5, %%mm4\n\t"
 "	movq	%%mm0, 96(%%ecx)\n\t"