gcc 3.3 is a bit too smart here, It optimizes these variables away. However
movd to a mmx register only works from reg or mem.
	-jsaw

--- SDL-1.2.6/src/video/SDL_RLEaccel.c	2003-08-30 21:13:06.000000000 +0200
+++ SDL-1.2.6-rock/src/video/SDL_RLEaccel.c	2004-02-23 18:11:50.000000000 +0100
@@ -143,7 +143,7 @@
     do {							\
 	Uint32 *srcp = (Uint32 *)(from);			\
 	Uint32 *dstp = (Uint32 *)(to);				\
-        int i = 0x00FF00FF;					\
+        volatile int i = 0x00FF00FF;				\
         movd_m2r(*(&i), mm3);					\
         punpckldq_r2r(mm3, mm3);				\
         i = 0xFF000000;						\
@@ -213,10 +213,11 @@
 
 #define ALPHA_BLIT16_565MMX(to, from, length, bpp, alpha)	\
     do {						\
-        int i, n = 0;					\
+        volatile int i;					\
+	int n = 0;					\
 	Uint16 *srcp = (Uint16 *)(from);		\
 	Uint16 *dstp = (Uint16 *)(to);			\
-        Uint32 ALPHA = 0xF800;				\
+        volatile Uint32 ALPHA = 0xF800;			\
 	movd_m2r(*(&ALPHA), mm1);			\
         punpcklwd_r2r(mm1, mm1);			\
         punpcklwd_r2r(mm1, mm1);			\
@@ -229,7 +230,7 @@
         punpcklwd_r2r(mm7, mm7);			\
         punpcklwd_r2r(mm7, mm7);			\
 	alpha &= ~(1+2+4);				\
-        i = (Uint32)alpha | (Uint32)alpha << 16;	\
+        *(&i) = (Uint32)alpha | (Uint32)alpha << 16;	\
         movd_m2r(*(&i), mm0);				\
         punpckldq_r2r(mm0, mm0);			\
         ALPHA = alpha >> 3;				\
@@ -303,10 +304,11 @@
 
 #define ALPHA_BLIT16_555MMX(to, from, length, bpp, alpha)	\
     do {						\
-        int i, n = 0;					\
+        volatile int i;					\
+	int n = 0;					\
 	Uint16 *srcp = (Uint16 *)(from);		\
 	Uint16 *dstp = (Uint16 *)(to);			\
-        Uint32 ALPHA = 0x7C00;				\
+        volatile Uint32 ALPHA = 0x7C00;			\
 	movd_m2r(*(&ALPHA), mm1);			\
         punpcklwd_r2r(mm1, mm1);			\
         punpcklwd_r2r(mm1, mm1);			\
@@ -319,10 +321,10 @@
         punpcklwd_r2r(mm7, mm7);			\
         punpcklwd_r2r(mm7, mm7);			\
 	alpha &= ~(1+2+4);				\
-        i = (Uint32)alpha | (Uint32)alpha << 16;	\
+        *(&i) = (Uint32)alpha | (Uint32)alpha << 16;	\
         movd_m2r(*(&i), mm0);				\
         punpckldq_r2r(mm0, mm0);			\
-        i = ((int)(length) & 3);				\
+        i = ((int)(length) & 3);			\
         ALPHA = alpha >> 3;				\
 	for(; i > 0; --i) {				\
 	    Uint32 s = *srcp++;				\
@@ -524,7 +526,7 @@
     do {								\
 	Uint32 *srcp = (Uint32 *)(from);				\
 	Uint32 *dstp = (Uint32 *)(to);					\
-        int i = 0x00fefefe;						\
+        volatile int i = 0x00fefefe;					\
         movd_m2r(*(&i), mm4);						\
         punpckldq_r2r(mm4, mm4);					\
         i = 0x00010101;							\