--- np2/i386c/ia32/instructions/data_trans.c	2004/03/12 13:34:08	1.13
+++ np2/i386c/ia32/instructions/data_trans.c	2004/05/22 16:35:07	1.16
@@ -1,4 +1,4 @@
-/*	$Id: data_trans.c,v 1.13 2004/03/12 13:34:08 monaka Exp $	*/
+/*	$Id: data_trans.c,v 1.16 2004/05/22 16:35:07 monaka Exp $	*/
 
 /*
  * Copyright (c) 2003 NONAKA Kimihiro
@@ -153,7 +153,6 @@ MOV_EdSw(void)
 		} else {
 			CPU_WORKCLOCK(3);
 			madr = calc_ea_dst(op);
-			// LSB only...
 			cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, (UINT16)src);
 		}
 		return;
@@ -374,7 +373,7 @@ CMOVO_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (CPU_OV) {
+	if (CC_O) {
 		*out = (UINT16)src;
 	}
 }
@@ -386,7 +385,7 @@ CMOVO_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (CPU_OV) {
+	if (CC_O) {
 		*out = src;
 	}
 }
@@ -398,7 +397,7 @@ CMOVNO_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (!CPU_OV) {
+	if (CC_NO) {
 		*out = (UINT16)src;
 	}
 }
@@ -410,7 +409,7 @@ CMOVNO_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (!CPU_OV) {
+	if (CC_NO) {
 		*out = src;
 	}
 }
@@ -422,7 +421,7 @@ CMOVC_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (CPU_FLAGL & C_FLAG) {
+	if (CC_C) {
 		*out = (UINT16)src;
 	}
 }
@@ -434,7 +433,7 @@ CMOVC_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (CPU_FLAGL & C_FLAG) {
+	if (CC_C) {
 		*out = src;
 	}
 }
@@ -446,7 +445,7 @@ CMOVNC_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (!(CPU_FLAGL & C_FLAG)) {
+	if (CC_NC) {
 		*out = (UINT16)src;
 	}
 }
@@ -458,7 +457,7 @@ CMOVNC_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (!(CPU_FLAGL & C_FLAG)) {
+	if (CC_NC) {
 		*out = src;
 	}
 }
@@ -470,7 +469,7 @@ CMOVZ_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (CPU_FLAGL & Z_FLAG) {
+	if (CC_Z) {
 		*out = (UINT16)src;
 	}
 }
@@ -482,7 +481,7 @@ CMOVZ_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (CPU_FLAGL & Z_FLAG) {
+	if (CC_Z) {
 		*out = src;
 	}
 }
@@ -494,7 +493,7 @@ CMOVNZ_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (!(CPU_FLAGL & Z_FLAG)) {
+	if (CC_NZ) {
 		*out = (UINT16)src;
 	}
 }
@@ -506,7 +505,7 @@ CMOVNZ_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (!(CPU_FLAGL & Z_FLAG)) {
+	if (CC_NZ) {
 		*out = src;
 	}
 }
@@ -518,7 +517,7 @@ CMOVA_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (!(CPU_FLAGL & (Z_FLAG | C_FLAG))) {
+	if (CC_A) {
 		*out = (UINT16)src;
 	}
 }
@@ -530,7 +529,7 @@ CMOVA_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (!(CPU_FLAGL & (Z_FLAG | C_FLAG))) {
+	if (CC_A) {
 		*out = src;
 	}
 }
@@ -542,7 +541,7 @@ CMOVNA_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (CPU_FLAGL & (Z_FLAG | C_FLAG)) {
+	if (CC_NA) {
 		*out = (UINT16)src;
 	}
 }
@@ -554,7 +553,7 @@ CMOVNA_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (CPU_FLAGL & (Z_FLAG | C_FLAG)) {
+	if (CC_NA) {
 		*out = src;
 	}
 }
@@ -566,7 +565,7 @@ CMOVS_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (CPU_FLAGL & S_FLAG) {
+	if (CC_S) {
 		*out = (UINT16)src;
 	}
 }
@@ -578,7 +577,7 @@ CMOVS_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (CPU_FLAGL & S_FLAG) {
+	if (CC_S) {
 		*out = src;
 	}
 }
@@ -590,7 +589,7 @@ CMOVNS_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (!(CPU_FLAGL & S_FLAG)) {
+	if (CC_NS) {
 		*out = (UINT16)src;
 	}
 }
@@ -602,7 +601,7 @@ CMOVNS_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (!(CPU_FLAGL & S_FLAG)) {
+	if (CC_NS) {
 		*out = src;
 	}
 }
@@ -614,7 +613,7 @@ CMOVP_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (CPU_FLAGL & P_FLAG) {
+	if (CC_P) {
 		*out = (UINT16)src;
 	}
 }
@@ -626,7 +625,7 @@ CMOVP_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (CPU_FLAGL & P_FLAG) {
+	if (CC_P) {
 		*out = src;
 	}
 }
@@ -638,7 +637,7 @@ CMOVNP_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if (!(CPU_FLAGL & P_FLAG)) {
+	if (CC_NP) {
 		*out = (UINT16)src;
 	}
 }
@@ -650,7 +649,7 @@ CMOVNP_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if (!(CPU_FLAGL & P_FLAG)) {
+	if (CC_NP) {
 		*out = src;
 	}
 }
@@ -662,7 +661,7 @@ CMOVL_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if ((!CPU_OV) != (!(CPU_FLAGL & S_FLAG))) {
+	if (CC_L) {
 		*out = (UINT16)src;
 	}
 }
@@ -674,7 +673,7 @@ CMOVL_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if ((!CPU_OV) != (!(CPU_FLAGL & S_FLAG))) {
+	if (CC_L) {
 		*out = src;
 	}
 }
@@ -686,7 +685,7 @@ CMOVNL_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if ((!CPU_OV) == (!(CPU_FLAGL & S_FLAG))) {
+	if (CC_NL) {
 		*out = (UINT16)src;
 	}
 }
@@ -698,7 +697,7 @@ CMOVNL_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if ((!CPU_OV) == (!(CPU_FLAGL & S_FLAG))) {
+	if (CC_NL) {
 		*out = src;
 	}
 }
@@ -710,7 +709,7 @@ CMOVLE_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if ((CPU_FLAGL & Z_FLAG) || ((!(CPU_FLAGL & S_FLAG)) != (!CPU_OV))) {
+	if (CC_LE) {
 		*out = (UINT16)src;
 	}
 }
@@ -722,7 +721,7 @@ CMOVLE_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if ((CPU_FLAGL & Z_FLAG) || ((!(CPU_FLAGL & S_FLAG)) != (!CPU_OV))) {
+	if (CC_LE) {
 		*out = src;
 	}
 }
@@ -734,7 +733,7 @@ CMOVNLE_GwEw(void)
 	UINT32 op, src;
 
 	PREPART_REG16_EA(op, src, out, 2, 5);
-	if ((!(CPU_FLAGL & Z_FLAG)) && ((!(CPU_FLAGL & S_FLAG)) == (!CPU_OV))) {
+	if (CC_NLE) {
 		*out = (UINT16)src;
 	}
 }
@@ -746,7 +745,7 @@ CMOVNLE_GdEd(void)
 	UINT32 op, src;
 
 	PREPART_REG32_EA(op, src, out, 2, 5);
-	if ((!(CPU_FLAGL & Z_FLAG)) && ((!(CPU_FLAGL & S_FLAG)) == (!CPU_OV))) {
+	if (CC_NLE) {
 		*out = src;
 	}
 }
@@ -754,12 +753,19 @@ CMOVNLE_GdEd(void)
 /*
  * XCHG
  */
+static UINT32
+XCHG(UINT32 dst, void *arg)
+{
+	UINT32 src = PTR_TO_UINT32(arg);
+	(void)dst;
+	return src;
+}
+
 void
 XCHG_EbGb(void)
 {
 	UINT8 *out, *src;
 	UINT32 op, madr;
-	UINT8 tmp;
 
 	PREPART_EA_REG8P(op, src);
 	if (op >= 0xc0) {
@@ -769,9 +775,7 @@ XCHG_EbGb(void)
 	} else {
 		CPU_WORKCLOCK(5);
 		madr = calc_ea_dst(op);
-		tmp = cpu_vmemoryread(CPU_INST_SEGREG_INDEX, madr);
-		cpu_vmemorywrite(CPU_INST_SEGREG_INDEX, madr, *src);
-		*src = tmp;
+		*src = (UINT8)cpu_memory_access_va_RMW(CPU_INST_SEGREG_INDEX, madr, XCHG, UINT32_TO_PTR(*src));
 	}
 }
 
@@ -780,7 +784,6 @@ XCHG_EwGw(void)
 {
 	UINT16 *out, *src;
 	UINT32 op, madr;
-	UINT16 tmp;
 
 	PREPART_EA_REG16P(op, src);
 	if (op >= 0xc0) {
@@ -790,9 +793,7 @@ XCHG_EwGw(void)
 	} else {
 		CPU_WORKCLOCK(5);
 		madr = calc_ea_dst(op);
-		tmp = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, *src);
-		*src = tmp;
+		*src = (UINT16)cpu_memory_access_va_RMW_w(CPU_INST_SEGREG_INDEX, madr, XCHG, UINT32_TO_PTR(*src));
 	}
 }
 
@@ -801,7 +802,6 @@ XCHG_EdGd(void)
 {
 	UINT32 *out, *src;
 	UINT32 op, madr;
-	UINT32 tmp;
 
 	PREPART_EA_REG32P(op, src);
 	if (op >= 0xc0) {
@@ -811,9 +811,7 @@ XCHG_EdGd(void)
 	} else {
 		CPU_WORKCLOCK(5);
 		madr = calc_ea_dst(op);
-		tmp = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, madr, *src);
-		*src = tmp;
+		*src = cpu_memory_access_va_RMW_d(CPU_INST_SEGREG_INDEX, madr, XCHG, UINT32_TO_PTR(*src));
 	}
 }
 
@@ -844,25 +842,25 @@ void XCHG_EDIEAX(void) { CPU_WORKCLOCK(3
 INLINE static UINT32
 BSWAP_DWORD(UINT32 val)
 {
-#if defined(__GNUC__) && (defined(i386) || defined(__i386__))
+#if defined(GCC_CPU_ARCH_IA32)
 	__asm__ __volatile__ (
-#if defined(IA32_USE_ASM_BSWAP)
+#if defined(IA32_USE_ASM_BSWAP) || defined(GCC_CPU_ARCH_AMD64)
 		"bswap %0"
-#else	/* !IA32_USE_ASM_BSWAP */
+#else	/* !(IA32_USE_ASM_BSWAP || GCC_CPU_ARCH_AMD64) */
 		"rorw $8, %w1\n\t"
 		"rorl $16, %1\n\t"
 		"rorw $8, %w1\n\t"
-#endif	/* IA32_USE_ASM_BSWAP */
+#endif	/* IA32_USE_ASM_BSWAP || GCC_CPU_ARCH_AMD64 */
 		: "=r" (val) : "0" (val));
 	return val;
-#else	/* !(__GNUC__ && (i386 || __i386__)) */
+#else	/* !(__GNUC__ && GCC_CPU_ARCH_IA32) */
 	UINT32 v;
 	v  = (val & 0x000000ff) << 24;
 	v |= (val & 0x0000ff00) << 8;
 	v |= (val & 0x00ff0000) >> 8;
 	v |= (val & 0xff000000) >> 24;
 	return v;
-#endif	/* __GNUC__ && (i386 || __i386__) */
+#endif	/* __GNUC__ && GCC_CPU_ARCH_IA32 */
 }
 #endif	/* bswap32 && !IA32_USE_ASM_BSWAP */
 
@@ -878,6 +876,33 @@ void BSWAP_EDI(void) { CPU_WORKCLOCK(2);
 /*
  * XADD
  */
+static UINT32
+XADD1(UINT32 dst, void *arg)
+{
+	UINT32 src = PTR_TO_UINT32(arg);
+	UINT32 res;
+	BYTE_ADD(res, dst, src);
+	return res;
+}
+
+static UINT32
+XADD2(UINT32 dst, void *arg)
+{
+	UINT32 src = PTR_TO_UINT32(arg);
+	UINT32 res;
+	WORD_ADD(res, dst, src);
+	return res;
+}
+
+static UINT32
+XADD4(UINT32 dst, void *arg)
+{
+	UINT32 src = PTR_TO_UINT32(arg);
+	UINT32 res;
+	DWORD_ADD(res, dst, src);
+	return res;
+}
+
 void
 XADD_EbGb(void)
 {
@@ -889,16 +914,13 @@ XADD_EbGb(void)
 		CPU_WORKCLOCK(2);
 		out = reg8_b20[op];
 		dst = *out;
-		ADD_BYTE(res, dst, *src);
+		BYTE_ADD(res, dst, *src);
 		*src = (UINT8)dst;
 		*out = (UINT8)res;
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		dst = cpu_vmemoryread(CPU_INST_SEGREG_INDEX, madr);
-		ADD_BYTE(res, dst, *src);
-		*src = (UINT8)dst;
-		cpu_vmemorywrite(CPU_INST_SEGREG_INDEX, madr, (UINT8)res);
+		*src = (UINT8)cpu_memory_access_va_RMW(CPU_INST_SEGREG_INDEX, madr, XADD1, UINT32_TO_PTR(*src));
 	}
 }
 
@@ -913,16 +935,13 @@ XADD_EwGw(void)
 		CPU_WORKCLOCK(2);
 		out = reg16_b20[op];
 		dst = *out;
-		ADD_WORD(res, dst, *src);
+		WORD_ADD(res, dst, *src);
 		*src = (UINT16)dst;
 		*out = (UINT16)res;
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		ADD_WORD(res, dst, *src);
-		*src = (UINT16)dst;
-		cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, (UINT16)res);
+		*src = (UINT16)cpu_memory_access_va_RMW_w(CPU_INST_SEGREG_INDEX, madr, XADD2, UINT32_TO_PTR(*src));
 	}
 }
 
@@ -937,135 +956,46 @@ XADD_EdGd(void)
 		CPU_WORKCLOCK(2);
 		out = reg32_b20[op];
 		dst = *out;
-		ADD_DWORD(res, dst, *src);
+		DWORD_ADD(res, dst, *src);
 		*src = dst;
 		*out = res;
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		ADD_DWORD(res, dst, *src);
-		*src = dst;
-		cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, madr, res);
+		*src = cpu_memory_access_va_RMW_d(CPU_INST_SEGREG_INDEX, madr, XADD4, UINT32_TO_PTR(*src));
 	}
 }
 
 /*
  * CMPXCHG
  */
-#if 1
-void
-CMPXCHG_EbGb(void)
-{
-	UINT8 *out;
-	UINT32 op, src, dst, madr, tmp;
-
-	PREPART_EA_REG8(op, src);
-	if (op >= 0xc0) {
-		out = reg8_b20[op];
-		dst = *out;
-		BYTE_SUB(tmp, CPU_AL, dst);
-		if (CPU_FLAGL & Z_FLAG) {
-			*out = (UINT8)src;
-		} else {
-			CPU_AL = (UINT8)dst;
-		}
-	} else {
-		madr = calc_ea_dst(op);
-		dst = cpu_vmemoryread(CPU_INST_SEGREG_INDEX, madr);
-		BYTE_SUB(tmp, CPU_AL, dst);
-		if (CPU_FLAGL & Z_FLAG) {
-			cpu_vmemorywrite(CPU_INST_SEGREG_INDEX, madr, (UINT8)src);
-		} else {
-			CPU_AL = (UINT8)dst;
-		}
-	}
-}
-
-void
-CMPXCHG_EwGw(void)
-{
-	UINT16 *out;
-	UINT32 op, src, dst, madr, tmp;
-
-	PREPART_EA_REG16(op, src);
-	if (op >= 0xc0) {
-		out = reg16_b20[op];
-		dst = *out;
-		WORD_SUB(tmp, CPU_AX, dst);
-		if (CPU_FLAGL & Z_FLAG) {
-			*out = (UINT16)src;
-		} else {
-			CPU_AX = (UINT16)dst;
-		}
-	} else {
-		madr = calc_ea_dst(op);
-		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		WORD_SUB(tmp, CPU_AX, dst);
-		if (CPU_FLAGL & Z_FLAG) {
-			cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, (UINT16)src);
-		} else {
-			CPU_AX = (UINT16)dst;
-		}
-	}
-}
-
-void
-CMPXCHG_EdGd(void)
-{
-	UINT32 *out;
-	UINT32 op, src, dst, madr, tmp;
-
-	PREPART_EA_REG32(op, src);
-	if (op >= 0xc0) {
-		out = reg32_b20[op];
-		dst = *out;
-		DWORD_SUB(tmp, CPU_EAX, dst);
-		if (CPU_FLAGL & Z_FLAG) {
-			*out = src;
-		} else {
-			CPU_EAX = dst;
-		}
-	} else {
-		madr = calc_ea_dst(op);
-		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		DWORD_SUB(tmp, CPU_EAX, dst);
-		if (CPU_FLAGL & Z_FLAG) {
-			cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, madr, src);
-		} else {
-			CPU_EAX = dst;
-		}
-	}
-}
-#else
 void
 CMPXCHG_EbGb(void)
 {
 	UINT8 *out;
 	UINT32 op, src, dst, madr, tmp;
+	UINT8 al;
 
 	PREPART_EA_REG8(op, src);
+	al = CPU_AL;
 	if (op >= 0xc0) {
 		out = reg8_b20[op];
 		dst = *out;
-		if (CPU_AL == dst) {
-			CPU_FLAGL |= Z_FLAG;
+		if (al == dst) {
 			*out = (UINT8)src;
 		} else {
-			CPU_FLAGL &= ~Z_FLAG;
 			CPU_AL = (UINT8)dst;
 		}
 	} else {
 		madr = calc_ea_dst(op);
 		dst = cpu_vmemoryread(CPU_INST_SEGREG_INDEX, madr);
-		if (CPU_AL == dst) {
-			CPU_FLAGL |= Z_FLAG;
+		if (al == dst) {
 			cpu_vmemorywrite(CPU_INST_SEGREG_INDEX, madr, (UINT8)src);
 		} else {
-			CPU_FLAGL &= ~Z_FLAG;
 			CPU_AL = (UINT8)dst;
 		}
 	}
+	BYTE_SUB(tmp, al, dst);
 }
 
 void
@@ -1073,29 +1003,28 @@ CMPXCHG_EwGw(void)
 {
 	UINT16 *out;
 	UINT32 op, src, dst, madr, tmp;
+	UINT16 ax;
 
 	PREPART_EA_REG16(op, src);
+	ax = CPU_AX;
 	if (op >= 0xc0) {
 		out = reg16_b20[op];
 		dst = *out;
-		if (CPU_AX == dst) {
-			CPU_FLAGL |= Z_FLAG;
+		if (ax == dst) {
 			*out = (UINT16)src;
 		} else {
-			CPU_FLAGL &= ~Z_FLAG;
 			CPU_AX = (UINT16)dst;
 		}
 	} else {
 		madr = calc_ea_dst(op);
 		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		if (CPU_AX == dst) {
-			CPU_FLAGL |= Z_FLAG;
+		if (ax == dst) {
 			cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, (UINT16)src);
 		} else {
-			CPU_FLAGL &= ~Z_FLAG;
 			CPU_AX = (UINT16)dst;
 		}
 	}
+	WORD_SUB(tmp, ax, dst);
 }
 
 void
@@ -1103,31 +1032,29 @@ CMPXCHG_EdGd(void)
 {
 	UINT32 *out;
 	UINT32 op, src, dst, madr, tmp;
+	UINT32 eax;
 
 	PREPART_EA_REG32(op, src);
+	eax = CPU_EAX;
 	if (op >= 0xc0) {
 		out = reg32_b20[op];
 		dst = *out;
-		if (CPU_EAX == dst) {
-			CPU_FLAGL |= Z_FLAG;
+		if (eax == dst) {
 			*out = src;
 		} else {
-			CPU_FLAGL &= ~Z_FLAG;
 			CPU_EAX = dst;
 		}
 	} else {
 		madr = calc_ea_dst(op);
 		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		if (CPU_EAX == dst) {
-			CPU_FLAGL |= Z_FLAG;
+		if (eax == dst) {
 			cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, madr, src);
 		} else {
-			CPU_FLAGL &= ~Z_FLAG;
 			CPU_EAX = dst;
 		}
 	}
+	DWORD_SUB(tmp, eax, dst);
 }
-#endif
 
 void
 CMPXCHG8B(UINT32 op)