diff --git a/util/x86emu/include/x86emu/regs.h b/util/x86emu/include/x86emu/regs.h
index 718d22fc86..df5c6b2d90 100644
--- a/util/x86emu/include/x86emu/regs.h
+++ b/util/x86emu/include/x86emu/regs.h
@@ -231,6 +231,9 @@ struct i386_segment_regs {
 #define SYSMODE_PREFIX_REPNE    0x00000100
 #define SYSMODE_PREFIX_DATA     0x00000200
 #define SYSMODE_PREFIX_ADDR     0x00000400
+//phueper: for REP(E|NE) Instructions, we need to decide wether it should be using
+//the 32bit ECX register as or the 16bit CX register as count register
+#define SYSMODE_32BIT_REP       0x00000800
 #define SYSMODE_INTR_PENDING    0x10000000
 #define SYSMODE_EXTRN_INTR      0x20000000
 #define SYSMODE_HALTED          0x40000000
@@ -250,7 +253,8 @@ struct i386_segment_regs {
 						 SYSMODE_SEGOVR_GS      | \
 						 SYSMODE_SEGOVR_SS      | \
 						 SYSMODE_PREFIX_DATA    | \
-						 SYSMODE_PREFIX_ADDR)
+						 SYSMODE_PREFIX_ADDR    | \
+						 SYSMODE_32BIT_REP)
 
 #define  INTR_SYNCH           0x1
 #define  INTR_ASYNCH          0x2
@@ -274,9 +278,9 @@ typedef struct {
      */
     u32                         mode;
     volatile int                intr;   /* mask of pending interrupts */
-	int                         debug;
+    volatile int                         debug;
 #ifdef DEBUG
-	int                         check;
+    int                         check;
     u16                         saved_ip;
     u16                         saved_cs;
     int                         enc_pos;
diff --git a/util/x86emu/include/x86emu/x86emu.h b/util/x86emu/include/x86emu/x86emu.h
index be05eb6337..8eb4efb397 100644
--- a/util/x86emu/include/x86emu/x86emu.h
+++ b/util/x86emu/include/x86emu/x86emu.h
@@ -189,6 +189,8 @@ void 	X86EMU_halt_sys(void);
 #define DEBUG_TRACECALL_REGS_F  0x004000
 #define DEBUG_DECODE_NOPRINT_F  0x008000 
 #define DEBUG_SAVE_IP_CS_F      0x010000
+#define DEBUG_TRACEJMP_F        0x020000
+#define DEBUG_TRACEJMP_REGS_F   0x040000
 #define DEBUG_SYS_F             (DEBUG_SVC_F|DEBUG_FS_F|DEBUG_PROC_F)
 
 void 	X86EMU_trace_regs(void);
diff --git a/util/x86emu/x86emu/debug.c b/util/x86emu/x86emu/debug.c
index 1a7e9cb52b..0c531eae2b 100644
--- a/util/x86emu/x86emu/debug.c
+++ b/util/x86emu/x86emu/debug.c
@@ -52,7 +52,11 @@ static int      parse_line (char *s, int *ps, int *n);
 void X86EMU_trace_regs (void)
 {
     if (DEBUG_TRACE()) {
-        x86emu_dump_regs();
+	if (M.x86.mode & (SYSMODE_PREFIX_DATA | SYSMODE_PREFIX_ADDR)) {
+	        x86emu_dump_xregs();
+	} else {
+	        x86emu_dump_regs();
+	}
     }
     if (DEBUG_DECODE() && ! DEBUG_DECODE_NOPRINT()) {
         printk("%04x:%04x ",M.x86.saved_cs, M.x86.saved_ip);
@@ -185,7 +189,7 @@ static void print_encoded_bytes (u16 s, u16 o)
     for (i=0; i< M.x86.enc_pos; i++) {
         sprintf(buf1+2*i,"%02x", fetch_data_byte_abs(s,o+i));
     }
-    printk("%-20s",buf1);
+    printk("%-20s ",buf1);
 }
 
 static void print_decoded_instruction (void)
diff --git a/util/x86emu/x86emu/debug.h b/util/x86emu/x86emu/debug.h
index d786a3d17f..93fa5b008e 100644
--- a/util/x86emu/x86emu/debug.h
+++ b/util/x86emu/x86emu/debug.h
@@ -78,6 +78,8 @@
 # define DEBUG_SYSINT()        	(M.x86.debug & DEBUG_SYSINT_F)
 # define DEBUG_TRACECALL()     	(M.x86.debug & DEBUG_TRACECALL_F)
 # define DEBUG_TRACECALLREGS() 	(M.x86.debug & DEBUG_TRACECALL_REGS_F)
+# define DEBUG_TRACEJMP()       (M.x86.debug & DEBUG_TRACEJMP_F)
+# define DEBUG_TRACEJMPREGS()   (M.x86.debug & DEBUG_TRACEJMP_REGS_F)
 # define DEBUG_SYS()           	(M.x86.debug & DEBUG_SYS_F)
 # define DEBUG_MEM_TRACE()     	(M.x86.debug & DEBUG_MEM_TRACE_F)
 # define DEBUG_IO_TRACE()      	(M.x86.debug & DEBUG_IO_TRACE_F)
@@ -96,6 +98,8 @@
 # define DEBUG_SYSINT()        	0
 # define DEBUG_TRACECALL()     	0
 # define DEBUG_TRACECALLREGS() 	0
+# define DEBUG_TRACEJMP()       0
+# define DEBUG_TRACEJMPREGS()   0
 # define DEBUG_SYS()           	0
 # define DEBUG_MEM_TRACE()     	0
 # define DEBUG_IO_TRACE()      	0
@@ -169,14 +173,20 @@
 		x86emu_dump_regs();                                     \
 	if (DEBUG_TRACECALL())                                     	\
 		printk("%04x:%04x: CALL %s%04x:%04x\n", u , v, s, w, x);
-# define RETURN_TRACE(n,u,v)                                    \
+# define RETURN_TRACE(u,v,w,x,s)                                    \
 	if (DEBUG_TRACECALLREGS())									\
 		x86emu_dump_regs();                                     \
 	if (DEBUG_TRACECALL())                                     	\
-		printk("%04x:%04x: %s\n",u,v,n);
+		printk("%04x:%04x: RET %s %04x:%04x\n",u,v,s,w,x);
+# define  JMP_TRACE(u,v,w,x,s)                                 \
+   if (DEBUG_TRACEJMPREGS()) \
+      x86emu_dump_regs(); \
+   if (DEBUG_TRACEJMP()) \
+      printk("%04x:%04x: JMP %s%04x:%04x\n", u , v, s, w, x);
 #else
 # define CALL_TRACE(u,v,w,x,s)
-# define RETURN_TRACE(n,u,v)
+# define RETURN_TRACE(u,v,w,x,s)
+# define  JMP_TRACE(u,v,w,x,s)
 #endif
 
 #ifdef DEBUG
diff --git a/util/x86emu/x86emu/ops.c b/util/x86emu/x86emu/ops.c
index 95f7c9e68d..772bcbb15a 100644
--- a/util/x86emu/x86emu/ops.c
+++ b/util/x86emu/x86emu/ops.c
@@ -1061,7 +1061,11 @@ void x86emuOp_push_byte_IMM(u8 X86EMU_UNUSED(op1))
     imm = (s8)fetch_byte_imm();
     DECODE_PRINTF2("PUSH\t%d\n", imm);
     TRACE_AND_STEP();
-    push_word(imm);
+    if (M.x86.mode & SYSMODE_PREFIX_DATA) {
+        push_long(imm);
+    } else {
+        push_word(imm);
+    }
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
 }
@@ -1256,8 +1260,10 @@ void x86emuOp_jump_near_cond(u8 op1)
     target = (u16)(M.x86.R_IP + (s16)offset);
     DECODE_PRINTF2("%x\n", target);
     TRACE_AND_STEP();
-    if (cond)
+    if (cond) {
         M.x86.R_IP = target;
+	JMP_TRACE(M.x86.saved_cs, M.x86.saved_ip, M.x86.R_CS, M.x86.R_IP, " NEAR COND ");
+    }
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
 }
@@ -2516,9 +2522,11 @@ void x86emuOp_movs_byte(u8 X86EMU_UNUSED(op1))
     count = 1;
     if (M.x86.mode & (SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE)) {
         /* dont care whether REPE or REPNE */
-        /* move them until CX is ZERO. */
-        count = M.x86.R_CX;
+        /* move them until (E)CX is ZERO. */
+        count = (M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX;
         M.x86.R_CX = 0;
+	if (M.x86.mode & SYSMODE_32BIT_REP)
+            M.x86.R_ECX = 0;
         M.x86.mode &= ~(SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE);
     }
     while (count--) {
@@ -2526,6 +2534,8 @@ void x86emuOp_movs_byte(u8 X86EMU_UNUSED(op1))
         store_data_byte_abs(M.x86.R_ES, M.x86.R_DI, val);
         M.x86.R_SI += inc;
         M.x86.R_DI += inc;
+        if (M.x86.intr & INTR_HALTED)
+            break;
     }
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
@@ -2559,9 +2569,11 @@ void x86emuOp_movs_word(u8 X86EMU_UNUSED(op1))
     count = 1;
     if (M.x86.mode & (SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE)) {
         /* dont care whether REPE or REPNE */
-        /* move them until CX is ZERO. */
-        count = M.x86.R_CX;
+        /* move them until (E)CX is ZERO. */
+        count = (M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX;
         M.x86.R_CX = 0;
+	if (M.x86.mode & SYSMODE_32BIT_REP)
+            M.x86.R_ECX = 0;
         M.x86.mode &= ~(SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE);
     }
     while (count--) {
@@ -2574,6 +2586,8 @@ void x86emuOp_movs_word(u8 X86EMU_UNUSED(op1))
         }
         M.x86.R_SI += inc;
         M.x86.R_DI += inc;
+        if (M.x86.intr & INTR_HALTED)
+            break;
     }
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
@@ -2598,16 +2612,21 @@ void x86emuOp_cmps_byte(u8 X86EMU_UNUSED(op1))
 
     if (M.x86.mode & (SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE)) {
         /* REPE  */
-        /* move them until CX is ZERO. */
-        while (M.x86.R_CX != 0) {
+        /* move them until (E)CX is ZERO. */
+        while (((M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX) != 0) {
             val1 = fetch_data_byte(M.x86.R_SI);
             val2 = fetch_data_byte_abs(M.x86.R_ES, M.x86.R_DI);
                      cmp_byte(val1, val2);
-            M.x86.R_CX -= 1;
+            if (M.x86.mode & SYSMODE_32BIT_REP)
+                M.x86.R_ECX -= 1;
+            else
+                M.x86.R_CX -= 1;
             M.x86.R_SI += inc;
             M.x86.R_DI += inc;
             if ( (M.x86.mode & SYSMODE_PREFIX_REPE) && (ACCESS_FLAG(F_ZF) == 0) ) break;
             if ( (M.x86.mode & SYSMODE_PREFIX_REPNE) && ACCESS_FLAG(F_ZF) ) break;
+            if (M.x86.intr & INTR_HALTED)
+                break;
         }
         M.x86.mode &= ~(SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE);
     } else {
@@ -2644,8 +2663,8 @@ void x86emuOp_cmps_word(u8 X86EMU_UNUSED(op1))
     TRACE_AND_STEP();
     if (M.x86.mode & (SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE)) {
         /* REPE  */
-        /* move them until CX is ZERO. */
-        while (M.x86.R_CX != 0) {
+        /* move them until (E)CX is ZERO. */
+        while (((M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX) != 0) {
             if (M.x86.mode & SYSMODE_PREFIX_DATA) {
                 val1 = fetch_data_long(M.x86.R_SI);
                 val2 = fetch_data_long_abs(M.x86.R_ES, M.x86.R_DI);
@@ -2655,11 +2674,16 @@ void x86emuOp_cmps_word(u8 X86EMU_UNUSED(op1))
                 val2 = fetch_data_word_abs(M.x86.R_ES, M.x86.R_DI);
                 cmp_word((u16)val1, (u16)val2);
             }
-            M.x86.R_CX -= 1;
+            if (M.x86.mode & SYSMODE_32BIT_REP)
+                M.x86.R_ECX -= 1;
+            else
+                M.x86.R_CX -= 1;
             M.x86.R_SI += inc;
             M.x86.R_DI += inc;
             if ( (M.x86.mode & SYSMODE_PREFIX_REPE) && ACCESS_FLAG(F_ZF) == 0 ) break;
             if ( (M.x86.mode & SYSMODE_PREFIX_REPNE) && ACCESS_FLAG(F_ZF) ) break;
+            if (M.x86.intr & INTR_HALTED)
+                break;
         }
         M.x86.mode &= ~(SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE);
     } else {
@@ -2741,11 +2765,16 @@ void x86emuOp_stos_byte(u8 X86EMU_UNUSED(op1))
     TRACE_AND_STEP();
     if (M.x86.mode & (SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE)) {
         /* dont care whether REPE or REPNE */
-        /* move them until CX is ZERO. */
-        while (M.x86.R_CX != 0) {
+        /* move them until (E)CX is ZERO. */
+        while (((M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX) != 0) {
             store_data_byte_abs(M.x86.R_ES, M.x86.R_DI, M.x86.R_AL);
-            M.x86.R_CX -= 1;
+            if (M.x86.mode & SYSMODE_32BIT_REP)
+                M.x86.R_ECX -= 1;
+            else
+                M.x86.R_CX -= 1;
             M.x86.R_DI += inc;
+            if (M.x86.intr & INTR_HALTED)
+                break;
         }
         M.x86.mode &= ~(SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE);
     } else {
@@ -2783,9 +2812,11 @@ void x86emuOp_stos_word(u8 X86EMU_UNUSED(op1))
     count = 1;
     if (M.x86.mode & (SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE)) {
         /* dont care whether REPE or REPNE */
-        /* move them until CX is ZERO. */
-        count = M.x86.R_CX;
+        /* move them until (E)CX is ZERO. */
+        count = (M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX;
         M.x86.R_CX = 0;
+	if (M.x86.mode & SYSMODE_32BIT_REP)
+            M.x86.R_ECX = 0;
         M.x86.mode &= ~(SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE);
     }
     while (count--) {
@@ -2795,6 +2826,8 @@ void x86emuOp_stos_word(u8 X86EMU_UNUSED(op1))
             store_data_word_abs(M.x86.R_ES, M.x86.R_DI, M.x86.R_AX);
         }
         M.x86.R_DI += inc;
+        if (M.x86.intr & INTR_HALTED)
+            break;
     }
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
@@ -2817,11 +2850,16 @@ void x86emuOp_lods_byte(u8 X86EMU_UNUSED(op1))
         inc = 1;
     if (M.x86.mode & (SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE)) {
         /* dont care whether REPE or REPNE */
-        /* move them until CX is ZERO. */
-        while (M.x86.R_CX != 0) {
+        /* move them until (E)CX is ZERO. */
+        while (((M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX) != 0) {
             M.x86.R_AL = fetch_data_byte(M.x86.R_SI);
-            M.x86.R_CX -= 1;
+            if (M.x86.mode & SYSMODE_32BIT_REP)
+                M.x86.R_ECX -= 1;
+            else
+                M.x86.R_CX -= 1;
             M.x86.R_SI += inc;
+            if (M.x86.intr & INTR_HALTED)
+                break;
         }
         M.x86.mode &= ~(SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE);
     } else {
@@ -2859,9 +2897,11 @@ void x86emuOp_lods_word(u8 X86EMU_UNUSED(op1))
     count = 1;
     if (M.x86.mode & (SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE)) {
         /* dont care whether REPE or REPNE */
-        /* move them until CX is ZERO. */
-        count = M.x86.R_CX;
+        /* move them until (E)CX is ZERO. */
+        count = (M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX;
         M.x86.R_CX = 0;
+	if (M.x86.mode & SYSMODE_32BIT_REP)
+            M.x86.R_ECX = 0;
         M.x86.mode &= ~(SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE);
     }
     while (count--) {
@@ -2871,6 +2911,8 @@ void x86emuOp_lods_word(u8 X86EMU_UNUSED(op1))
             M.x86.R_AX = fetch_data_word(M.x86.R_SI);
         }
         M.x86.R_SI += inc;
+        if (M.x86.intr & INTR_HALTED)
+            break;
     }
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
@@ -2894,26 +2936,36 @@ void x86emuOp_scas_byte(u8 X86EMU_UNUSED(op1))
         inc = 1;
     if (M.x86.mode & SYSMODE_PREFIX_REPE) {
         /* REPE  */
-        /* move them until CX is ZERO. */
-        while (M.x86.R_CX != 0) {
+        /* move them until (E)CX is ZERO. */
+        while (((M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX) != 0) {
             val2 = fetch_data_byte_abs(M.x86.R_ES, M.x86.R_DI);
             cmp_byte(M.x86.R_AL, val2);
-            M.x86.R_CX -= 1;
+            if (M.x86.mode & SYSMODE_32BIT_REP)
+                M.x86.R_ECX -= 1;
+            else
+                M.x86.R_CX -= 1;
             M.x86.R_DI += inc;
             if (ACCESS_FLAG(F_ZF) == 0)
                 break;
+            if (M.x86.intr & INTR_HALTED)
+                break;
         }
         M.x86.mode &= ~SYSMODE_PREFIX_REPE;
     } else if (M.x86.mode & SYSMODE_PREFIX_REPNE) {
         /* REPNE  */
-        /* move them until CX is ZERO. */
-        while (M.x86.R_CX != 0) {
+        /* move them until (E)CX is ZERO. */
+        while (((M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX) != 0) {
             val2 = fetch_data_byte_abs(M.x86.R_ES, M.x86.R_DI);
             cmp_byte(M.x86.R_AL, val2);
-            M.x86.R_CX -= 1;
+            if (M.x86.mode & SYSMODE_32BIT_REP)
+                M.x86.R_ECX -= 1;
+            else
+                M.x86.R_CX -= 1;
             M.x86.R_DI += inc;
             if (ACCESS_FLAG(F_ZF))
                 break;          /* zero flag set means equal */
+            if (M.x86.intr & INTR_HALTED)
+                break;
         }
         M.x86.mode &= ~SYSMODE_PREFIX_REPNE;
     } else {
@@ -2951,8 +3003,8 @@ void x86emuOp_scas_word(u8 X86EMU_UNUSED(op1))
     TRACE_AND_STEP();
     if (M.x86.mode & SYSMODE_PREFIX_REPE) {
         /* REPE  */
-        /* move them until CX is ZERO. */
-        while (M.x86.R_CX != 0) {
+        /* move them until (E)CX is ZERO. */
+        while (((M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX) != 0) {
             if (M.x86.mode & SYSMODE_PREFIX_DATA) {
                 val = fetch_data_long_abs(M.x86.R_ES, M.x86.R_DI);
                 cmp_long(M.x86.R_EAX, val);
@@ -2960,16 +3012,21 @@ void x86emuOp_scas_word(u8 X86EMU_UNUSED(op1))
                 val = fetch_data_word_abs(M.x86.R_ES, M.x86.R_DI);
                 cmp_word(M.x86.R_AX, (u16)val);
             }
-            M.x86.R_CX -= 1;
+            if (M.x86.mode & SYSMODE_32BIT_REP)
+                M.x86.R_ECX -= 1;
+            else
+                M.x86.R_CX -= 1;
             M.x86.R_DI += inc;
             if (ACCESS_FLAG(F_ZF) == 0)
                 break;
+            if (M.x86.intr & INTR_HALTED)
+                break;
         }
         M.x86.mode &= ~SYSMODE_PREFIX_REPE;
     } else if (M.x86.mode & SYSMODE_PREFIX_REPNE) {
         /* REPNE  */
-        /* move them until CX is ZERO. */
-        while (M.x86.R_CX != 0) {
+        /* move them until (E)CX is ZERO. */
+        while (((M.x86.mode & SYSMODE_32BIT_REP) ? M.x86.R_ECX : M.x86.R_CX) != 0) {
             if (M.x86.mode & SYSMODE_PREFIX_DATA) {
                 val = fetch_data_long_abs(M.x86.R_ES, M.x86.R_DI);
                 cmp_long(M.x86.R_EAX, val);
@@ -2977,10 +3034,15 @@ void x86emuOp_scas_word(u8 X86EMU_UNUSED(op1))
                 val = fetch_data_word_abs(M.x86.R_ES, M.x86.R_DI);
                 cmp_word(M.x86.R_AX, (u16)val);
             }
-            M.x86.R_CX -= 1;
+            if (M.x86.mode & SYSMODE_32BIT_REP)
+                M.x86.R_ECX -= 1;
+            else
+                M.x86.R_CX -= 1;
             M.x86.R_DI += inc;
             if (ACCESS_FLAG(F_ZF))
                 break;          /* zero flag set means equal */
+            if (M.x86.intr & INTR_HALTED)
+                break;
         }
         M.x86.mode &= ~SYSMODE_PREFIX_REPNE;
     } else {
@@ -3238,9 +3300,9 @@ void x86emuOp_ret_near_IMM(u8 X86EMU_UNUSED(op1))
     DECODE_PRINTF("RET\t");
     imm = fetch_word_imm();
     DECODE_PRINTF2("%x\n", imm);
-	RETURN_TRACE("RET",M.x86.saved_cs,M.x86.saved_ip);
 	TRACE_AND_STEP();
     M.x86.R_IP = pop_word();
+	RETURN_TRACE(M.x86.saved_cs,M.x86.saved_ip, M.x86.R_CS, M.x86.R_IP, "NEAR");
     M.x86.R_SP += imm;
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
@@ -3254,9 +3316,9 @@ void x86emuOp_ret_near(u8 X86EMU_UNUSED(op1))
 {
     START_OF_INSTR();
     DECODE_PRINTF("RET\n");
-	RETURN_TRACE("RET",M.x86.saved_cs,M.x86.saved_ip);
 	TRACE_AND_STEP();
     M.x86.R_IP = pop_word();
+	RETURN_TRACE(M.x86.saved_cs,M.x86.saved_ip, M.x86.R_CS, M.x86.R_IP, "NEAR");
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
 }
@@ -3471,10 +3533,10 @@ void x86emuOp_ret_far_IMM(u8 X86EMU_UNUSED(op1))
     DECODE_PRINTF("RETF\t");
     imm = fetch_word_imm();
     DECODE_PRINTF2("%x\n", imm);
-	RETURN_TRACE("RETF",M.x86.saved_cs,M.x86.saved_ip);
 	TRACE_AND_STEP();
     M.x86.R_IP = pop_word();
     M.x86.R_CS = pop_word();
+	RETURN_TRACE(M.x86.saved_cs,M.x86.saved_ip, M.x86.R_CS, M.x86.R_IP, "FAR");
     M.x86.R_SP += imm;
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
@@ -3488,10 +3550,10 @@ void x86emuOp_ret_far(u8 X86EMU_UNUSED(op1))
 {
     START_OF_INSTR();
     DECODE_PRINTF("RETF\n");
-	RETURN_TRACE("RETF",M.x86.saved_cs,M.x86.saved_ip);
 	TRACE_AND_STEP();
     M.x86.R_IP = pop_word();
     M.x86.R_CS = pop_word();
+	RETURN_TRACE(M.x86.saved_cs,M.x86.saved_ip, M.x86.R_CS, M.x86.R_IP, "FAR");
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
 }
@@ -4020,8 +4082,11 @@ void x86emuOp_loopne(u8 X86EMU_UNUSED(op1))
     ip += (s16) M.x86.R_IP;
     DECODE_PRINTF2("%04x\n", ip);
     TRACE_AND_STEP();
-    M.x86.R_CX -= 1;
-    if (M.x86.R_CX != 0 && !ACCESS_FLAG(F_ZF))      /* CX != 0 and !ZF */
+    if (M.x86.mode & SYSMODE_PREFIX_ADDR)
+        M.x86.R_ECX -= 1;
+    else
+        M.x86.R_CX -= 1;
+    if (((M.x86.mode & SYSMODE_PREFIX_ADDR) ? M.x86.R_ECX : M.x86.R_CX) != 0 && !ACCESS_FLAG(F_ZF))      /* (E)CX != 0 and !ZF */
         M.x86.R_IP = ip;
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
@@ -4041,8 +4106,11 @@ void x86emuOp_loope(u8 X86EMU_UNUSED(op1))
     ip += (s16) M.x86.R_IP;
     DECODE_PRINTF2("%04x\n", ip);
     TRACE_AND_STEP();
-    M.x86.R_CX -= 1;
-    if (M.x86.R_CX != 0 && ACCESS_FLAG(F_ZF))       /* CX != 0 and ZF */
+    if (M.x86.mode & SYSMODE_PREFIX_ADDR)
+        M.x86.R_ECX -= 1;
+    else
+        M.x86.R_CX -= 1;
+    if (((M.x86.mode & SYSMODE_PREFIX_ADDR) ? M.x86.R_ECX : M.x86.R_CX) != 0 && ACCESS_FLAG(F_ZF))      /* (E)CX != 0 and ZF */
         M.x86.R_IP = ip;
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
@@ -4062,8 +4130,11 @@ void x86emuOp_loop(u8 X86EMU_UNUSED(op1))
     ip += (s16) M.x86.R_IP;
     DECODE_PRINTF2("%04x\n", ip);
     TRACE_AND_STEP();
-    M.x86.R_CX -= 1;
-    if (M.x86.R_CX != 0)
+    if (M.x86.mode & SYSMODE_PREFIX_ADDR)
+        M.x86.R_ECX -= 1;
+    else
+        M.x86.R_CX -= 1;
+    if (((M.x86.mode & SYSMODE_PREFIX_ADDR) ? M.x86.R_ECX : M.x86.R_CX) != 0)      /* (E)CX != 0 */
         M.x86.R_IP = ip;
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
@@ -4085,8 +4156,10 @@ void x86emuOp_jcxz(u8 X86EMU_UNUSED(op1))
     target = (u16)(M.x86.R_IP + offset);
     DECODE_PRINTF2("%x\n", target);
     TRACE_AND_STEP();
-    if (M.x86.R_CX == 0)
+    if (M.x86.R_CX == 0) {
         M.x86.R_IP = target;
+	JMP_TRACE(M.x86.saved_cs, M.x86.saved_ip, M.x86.R_CS, M.x86.R_IP, " CXZ ");
+    }
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
 }
@@ -4213,6 +4286,7 @@ void x86emuOp_jump_near_IMM(u8 X86EMU_UNUSED(op1))
     ip = (s16)fetch_word_imm();
     ip += (s16)M.x86.R_IP;
     DECODE_PRINTF2("%04x\n", ip);
+    JMP_TRACE(M.x86.saved_cs, M.x86.saved_ip, M.x86.R_CS, ip, " NEAR ");
     TRACE_AND_STEP();
     M.x86.R_IP = (u16)ip;
     DECODE_CLEAR_SEGOVR();
@@ -4233,6 +4307,7 @@ void x86emuOp_jump_far_IMM(u8 X86EMU_UNUSED(op1))
     cs = fetch_word_imm();
     DECODE_PRINTF2("%04x:", cs);
     DECODE_PRINTF2("%04x\n", ip);
+    JMP_TRACE(M.x86.saved_cs, M.x86.saved_ip, cs, ip, " FAR ");
     TRACE_AND_STEP();
     M.x86.R_IP = ip;
     M.x86.R_CS = cs;
@@ -4254,6 +4329,7 @@ void x86emuOp_jump_byte_IMM(u8 X86EMU_UNUSED(op1))
     offset = (s8)fetch_byte_imm();
     target = (u16)(M.x86.R_IP + offset);
     DECODE_PRINTF2("%x\n", target);
+    JMP_TRACE(M.x86.saved_cs, M.x86.saved_ip, M.x86.R_CS, target, " BYTE ");
     TRACE_AND_STEP();
     M.x86.R_IP = target;
     DECODE_CLEAR_SEGOVR();
@@ -4357,6 +4433,8 @@ void x86emuOp_repne(u8 X86EMU_UNUSED(op1))
     DECODE_PRINTF("REPNE\n");
     TRACE_AND_STEP();
     M.x86.mode |= SYSMODE_PREFIX_REPNE;
+    if (M.x86.mode & SYSMODE_PREFIX_ADDR)
+        M.x86.mode |= SYSMODE_32BIT_REP;
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
 }
@@ -4371,6 +4449,8 @@ void x86emuOp_repe(u8 X86EMU_UNUSED(op1))
     DECODE_PRINTF("REPE\n");
     TRACE_AND_STEP();
     M.x86.mode |= SYSMODE_PREFIX_REPE;
+    if (M.x86.mode & SYSMODE_PREFIX_ADDR)
+        M.x86.mode |= SYSMODE_32BIT_REP;
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
 }
@@ -5013,12 +5093,14 @@ void x86emuOp_opcFF_word_RM(u8 X86EMU_UNUSED(op1))
             break;
         case 4:         /* jmp word ptr ... */
             destval = fetch_data_word(destoffset);
+            JMP_TRACE(M.x86.saved_cs, M.x86.saved_ip, M.x86.R_CS, destval, " WORD ");
             TRACE_AND_STEP();
             M.x86.R_IP = destval;
             break;
         case 5:         /* jmp far ptr ... */
             destval = fetch_data_word(destoffset);
             destval2 = fetch_data_word(destoffset + 2);
+            JMP_TRACE(M.x86.saved_cs, M.x86.saved_ip, destval2, destval, " FAR ");
             TRACE_AND_STEP();
             M.x86.R_IP = destval;
             M.x86.R_CS = destval2;
diff --git a/util/x86emu/x86emu/ops2.c b/util/x86emu/x86emu/ops2.c
index 448d968570..53a777697e 100644
--- a/util/x86emu/x86emu/ops2.c
+++ b/util/x86emu/x86emu/ops2.c
@@ -149,8 +149,69 @@ void x86emuOp2_long_jump(u8 op2)
     target += (s16) M.x86.R_IP;
     DECODE_PRINTF2("%04x\n", target);
     TRACE_AND_STEP();
-    if (cond)
+    if (cond) {
         M.x86.R_IP = (u16)target;
+	JMP_TRACE(M.x86.saved_cs, M.x86.saved_ip, M.x86.R_CS, M.x86.R_IP, " LONG COND ");
+    }
+    DECODE_CLEAR_SEGOVR();
+    END_OF_INSTR();
+}
+
+/****************************************************************************
+REMARKS:
+Handles opcode 0x0f,0xC8-0xCF
+****************************************************************************/
+s32 x86emu_bswap(s32 reg)
+{
+   // perform the byte swap
+   s32 temp = reg;
+   reg = (temp & 0xFF000000) >> 24;
+   reg |= (temp & 0xFF0000) >> 8;
+   reg |= (temp & 0xFF00) << 8;
+   reg |= (temp & 0xFF) << 24;
+   return reg;
+}
+
+void x86emuOp2_bswap(u8 op2)
+{
+    /* byte swap 32 bit register */
+    START_OF_INSTR();
+    DECODE_PRINTF("BSWAP\t");
+    switch (op2) {
+      case 0xc8:
+        DECODE_PRINTF("EAX\n");
+        M.x86.R_EAX = x86emu_bswap(M.x86.R_EAX);
+        break;
+      case 0xc9:
+        DECODE_PRINTF("ECX\n");
+        M.x86.R_ECX = x86emu_bswap(M.x86.R_ECX);
+        break;
+      case 0xca:
+        DECODE_PRINTF("EDX\n");
+        M.x86.R_EDX = x86emu_bswap(M.x86.R_EDX);
+        break;
+      case 0xcb:
+        DECODE_PRINTF("EBX\n");
+        M.x86.R_EBX = x86emu_bswap(M.x86.R_EBX);
+        break;
+      case 0xcc:
+        DECODE_PRINTF("ESP\n");
+        M.x86.R_ESP = x86emu_bswap(M.x86.R_ESP);
+        break;
+      case 0xcd:
+        DECODE_PRINTF("EBP\n");
+        M.x86.R_EBP = x86emu_bswap(M.x86.R_EBP);
+        break;
+      case 0xce:
+        DECODE_PRINTF("ESI\n");
+        M.x86.R_ESI = x86emu_bswap(M.x86.R_ESI);
+        break;
+      case 0xcf:
+        DECODE_PRINTF("EDI\n");
+        M.x86.R_EDI = x86emu_bswap(M.x86.R_EDI);
+        break;
+    }
+    TRACE_AND_STEP();
     DECODE_CLEAR_SEGOVR();
     END_OF_INSTR();
 }
@@ -1702,14 +1763,14 @@ void (*x86emu_optab2[256])(u8) =
 /*  0xc5 */ x86emuOp2_illegal_op,
 /*  0xc6 */ x86emuOp2_illegal_op,
 /*  0xc7 */ x86emuOp2_illegal_op,
-/*  0xc8 */ x86emuOp2_illegal_op,  /* TODO: bswap */
-/*  0xc9 */ x86emuOp2_illegal_op,  /* TODO: bswap */
-/*  0xca */ x86emuOp2_illegal_op,  /* TODO: bswap */
-/*  0xcb */ x86emuOp2_illegal_op,  /* TODO: bswap */
-/*  0xcc */ x86emuOp2_illegal_op,  /* TODO: bswap */
-/*  0xcd */ x86emuOp2_illegal_op,  /* TODO: bswap */
-/*  0xce */ x86emuOp2_illegal_op,  /* TODO: bswap */
-/*  0xcf */ x86emuOp2_illegal_op,  /* TODO: bswap */
+/*  0xc8 */ x86emuOp2_bswap,
+/*  0xc9 */ x86emuOp2_bswap,
+/*  0xca */ x86emuOp2_bswap,
+/*  0xcb */ x86emuOp2_bswap,
+/*  0xcc */ x86emuOp2_bswap,
+/*  0xcd */ x86emuOp2_bswap,
+/*  0xce */ x86emuOp2_bswap,
+/*  0xcf */ x86emuOp2_bswap,
 
 /*  0xd0 */ x86emuOp2_illegal_op,
 /*  0xd1 */ x86emuOp2_illegal_op,
diff --git a/util/x86emu/x86emu/prim_ops.c b/util/x86emu/x86emu/prim_ops.c
index fa19cb4844..a4a46a20ee 100644
--- a/util/x86emu/x86emu/prim_ops.c
+++ b/util/x86emu/x86emu/prim_ops.c
@@ -1921,7 +1921,7 @@ Implements the IMUL instruction and side effects.
 void imul_long_direct(u32 *res_lo, u32* res_hi,u32 d, u32 s)
 {
 #ifdef  __HAS_LONG_LONG__
-    s64 res = (s64)d * (s64)s;
+    s64 res = (s64)(s32)d * (s64)(s32)s;
 
     *res_lo = (u32)res;
     *res_hi = (u32)(res >> 32);
@@ -2013,7 +2013,7 @@ Implements the MUL instruction and side effects.
 void mul_long(u32 s)
 {
 #ifdef  __HAS_LONG_LONG__
-    u64 res = (u32)M.x86.R_EAX * (u32)s;
+    u64 res = (u64)M.x86.R_EAX * s;
 
     M.x86.R_EAX = (u32)res;
     M.x86.R_EDX = (u32)(res >> 32);
@@ -2312,16 +2312,15 @@ void ins(int size)
     }
     if (M.x86.mode & (SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE)) {
         /* dont care whether REPE or REPNE */
-        /* in until CX is ZERO. */
-        u32 count = ((M.x86.mode & SYSMODE_PREFIX_DATA) ?
+        /* in until (E)CX is ZERO. */
+        u32 count = ((M.x86.mode & SYSMODE_32BIT_REP) ?
                      M.x86.R_ECX : M.x86.R_CX);
-
         while (count--) {
           single_in(size);
           M.x86.R_DI += inc;
           }
         M.x86.R_CX = 0;
-        if (M.x86.mode & SYSMODE_PREFIX_DATA) {
+        if (M.x86.mode & SYSMODE_32BIT_REP) {
             M.x86.R_ECX = 0;
         }
         M.x86.mode &= ~(SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE);
@@ -2355,15 +2354,15 @@ void outs(int size)
     }
     if (M.x86.mode & (SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE)) {
         /* dont care whether REPE or REPNE */
-        /* out until CX is ZERO. */
-        u32 count = ((M.x86.mode & SYSMODE_PREFIX_DATA) ?
+        /* out until (E)CX is ZERO. */
+        u32 count = ((M.x86.mode & SYSMODE_32BIT_REP) ?
                      M.x86.R_ECX : M.x86.R_CX);
         while (count--) {
           single_out(size);
           M.x86.R_SI += inc;
           }
         M.x86.R_CX = 0;
-        if (M.x86.mode & SYSMODE_PREFIX_DATA) {
+        if (M.x86.mode & SYSMODE_32BIT_REP) {
             M.x86.R_ECX = 0;
         }
         M.x86.mode &= ~(SYSMODE_PREFIX_REPE | SYSMODE_PREFIX_REPNE);
diff --git a/util/x86emu/x86emu/sys.c b/util/x86emu/x86emu/sys.c
index 0eac6a3a10..11247b85bf 100644
--- a/util/x86emu/x86emu/sys.c
+++ b/util/x86emu/x86emu/sys.c
@@ -405,6 +405,6 @@ void X86EMU_prepareForInt(int num)
 
 void X86EMU_setMemBase(void *base, size_t size)
 {
-	M.mem_base = (int) base;
+	M.mem_base = (unsigned long) base;
 	M.mem_size = size;
 }