Index: liba52/imdct.c
===================================================================
RCS file: /cvsroot/mplayer/main/liba52/imdct.c,v
retrieving revision 1.25
diff -u -r1.25 imdct.c
--- liba52/imdct.c	26 Apr 2004 19:47:50 -0000	1.25
+++ liba52/imdct.c	2 May 2004 20:33:14 -0000
@@ -932,37 +932,34 @@
 
     /* 4-7. iterations */
     for (m=3; m < 7; m++) {
-	two_m = (1 << m);
-	two_m_plus_one = two_m<<1;
-	asm volatile(
-		"movl %0, %%esi				\n\t"
-		".balign 16				\n\t"
-		"1:					\n\t"
-		"xorl %%edi, %%edi			\n\t" // k
-		"leal (%%esi, %3), %%edx		\n\t"
-		"2:					\n\t"
-		"movaps (%%edx, %%edi), %%xmm1		\n\t"
-		"movaps (%4, %%edi, 2), %%xmm2		\n\t"
-		"mulps %%xmm1, %%xmm2			\n\t"
-		"shufps $0xB1, %%xmm1, %%xmm1		\n\t"
-		"mulps 16(%4, %%edi, 2), %%xmm1		\n\t"
-		"movaps (%%esi, %%edi), %%xmm0		\n\t"
-		"addps %%xmm2, %%xmm1			\n\t"
-		"movaps %%xmm1, %%xmm2			\n\t"
-		"addps %%xmm0, %%xmm1			\n\t"
-		"subps %%xmm2, %%xmm0			\n\t"
-		"movaps %%xmm1, (%%esi, %%edi)		\n\t"
-		"movaps %%xmm0, (%%edx, %%edi)		\n\t"
-		"addl $16, %%edi			\n\t"
-		"cmpl %3, %%edi				\n\t" //FIXME (opt) count against 0 
-		" jb 2b					\n\t"
-		"addl %2, %%esi				\n\t"
-		"cmpl %1, %%esi				\n\t"
-		" jb 1b					\n\t"
-		:: "g" (buf), "m" (buf+128), "m" (two_m_plus_one<<3), "r" (two_m<<3),
-		   "r" (sseW[m])
-		: "%esi", "%edi", "%edx"
-	);
+	int two_m8 = (8 << m);
+	unsigned long i = 0;
+	do {
+	    char *esi = (char*) buf + i;
+	    char *edx = esi + two_m8;
+	    unsigned long k = 0;
+	    do {
+		asm volatile(
+		    "movaps %0, %%xmm1			\n\t"
+		    "movaps %2, %%xmm2			\n\t"
+		    "mulps %%xmm1, %%xmm2		\n\t"
+		    "shufps $0xB1, %%xmm1, %%xmm1	\n\t"
+		    "mulps %3, %%xmm1			\n\t"
+		    "movaps %1, %%xmm0			\n\t"
+		    "addps %%xmm2, %%xmm1		\n\t"
+		    "movaps %%xmm1, %%xmm2		\n\t"
+		    "addps %%xmm0, %%xmm1		\n\t"
+		    "subps %%xmm2, %%xmm0		\n\t"
+		    "movaps %%xmm1, %1			\n\t"
+		    "movaps %%xmm0, %0			\n\t"
+		    :: "m" (edx[k]), "m" (esi[k]),
+		    "m" (((char*)sseW[m])[2*k]), "m" (((char*)sseW[m])[2*k+16])
+		    );
+		k += 16;
+	    } while (k < two_m8);
+	    i += 2 * two_m8;
+	} while (i < 8*128);
+
     }
 
     /* Post IFFT complex multiply  plus IFFT complex conjugate*/