--- np2/i386c/ia32/instructions/bit_byte.c	2003/12/08 00:55:32	1.1
+++ np2/i386c/ia32/instructions/bit_byte.c	2004/03/12 18:19:57	1.7
@@ -1,4 +1,4 @@
-/*	$Id: bit_byte.c,v 1.1 2003/12/08 00:55:32 yui Exp $	*/
+/*	$Id: bit_byte.c,v 1.7 2004/03/12 18:19:57 yui Exp $	*/
 
 /*
  * Copyright (c) 2002-2003 NONAKA Kimihiro
@@ -33,6 +33,24 @@
 
 #include "bit_byte.h"
 
+#if 1
+#define	BIT_OFFSET16(v)		(2 * (((SINT16)(v)) >> 4))
+#define	BIT_INDEX16(v)		((v) & 0xf)
+#define	BIT_MAKEBIT16(v)	(1 << BIT_INDEX16(v))
+
+#define	BIT_OFFSET32(v)		(4 * (((SINT32)(v)) >> 5))
+#define	BIT_INDEX32(v)		((v) & 0x1f)
+#define	BIT_MAKEBIT32(v)	(1 << BIT_INDEX32(v))
+#else
+#define	BIT_OFFSET16(v)		(2 * (((SINT16)(v)) / 16))
+#define	BIT_INDEX16(v)		((v) & 0xf)
+#define	BIT_MAKEBIT16(v)	(1 << BIT_INDEX16(v))
+
+#define	BIT_OFFSET32(v)		(4 * (((SINT32)(v)) / 32))
+#define	BIT_INDEX32(v)		((v) & 0x1f)
+#define	BIT_MAKEBIT32(v)	(1 << BIT_INDEX32(v))
+#endif
+
 
 /*
  * BT
@@ -40,7 +58,7 @@
 void
 BT_EwGw(void)
 {
-	DWORD op, src, dst, madr, ad;
+	UINT32 op, src, dst, madr;
 
 	PREPART_EA_REG16(op, src);
 	if (op >= 0xc0) {
@@ -49,21 +67,17 @@ BT_EwGw(void)
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		ad += 2 * (src >> 4);
-		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, ad);
-	}
-	if (dst & (1 << (src & 0x0f))) {
-		CPU_FLAGL |= C_FLAG;
-	} else {
-		CPU_FLAGL &= ~C_FLAG;
+		madr += BIT_OFFSET16(src);
+		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
 	}
+	CPU_FLAGL &= ~C_FLAG;
+	CPU_FLAGL |= (dst >> BIT_INDEX16(src)) & 1;
 }
 
 void
 BT_EdGd(void)
 {
-	DWORD op, src, dst, madr, ad;
+	UINT32 op, src, dst, madr;
 
 	PREPART_EA_REG32(op, src);
 	if (op >= 0xc0) {
@@ -72,61 +86,49 @@ BT_EdGd(void)
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		ad += 4 * (src >> 5);
-		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, ad);
-	}
-	if (dst & (1 << (src & 0x1f))) {
-		CPU_FLAGL |= C_FLAG;
-	} else {
-		CPU_FLAGL &= ~C_FLAG;
+		madr += BIT_OFFSET32(src);
+		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
 	}
+	CPU_FLAGL &= ~C_FLAG;
+	CPU_FLAGL |= (dst >> BIT_INDEX32(src)) & 1;
 }
 
 void
-BT_EwIb(DWORD op)
+BT_EwIb(UINT32 op)
 {
-	DWORD src, dst, madr, ad;
+	UINT32 src, dst, madr;
 
-	GET_PCBYTE(src);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
+		GET_PCBYTE(src);
 		dst = *(reg16_b20[op]);
 	} else {
 		CPU_WORKCLOCK(6);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		ad += 2 * (src >> 4);
-		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, ad);
-	}
-	if (dst & (1 << (src & 0x0f))) {
-		CPU_FLAGL |= C_FLAG;
-	} else {
-		CPU_FLAGL &= ~C_FLAG;
+		GET_PCBYTE(src);
+		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
 	}
+	CPU_FLAGL &= ~C_FLAG;
+	CPU_FLAGL |= (dst >> BIT_INDEX16(src)) & 1;
 }
 
 void
-BT_EdIb(DWORD op)
+BT_EdIb(UINT32 op)
 {
-	DWORD src, dst, madr, ad;
+	UINT32 src, dst, madr;
 
-	GET_PCBYTE(src);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
+		GET_PCBYTE(src);
 		dst = *(reg32_b20[op]);
 	} else {
 		CPU_WORKCLOCK(6);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		ad += 4 * (src >> 5);
-		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, ad);
-	}
-	if (dst & (1 << (src & 0x1f))) {
-		CPU_FLAGL |= C_FLAG;
-	} else {
-		CPU_FLAGL &= ~C_FLAG;
+		GET_PCBYTE(src);
+		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
 	}
+	CPU_FLAGL &= ~C_FLAG;
+	CPU_FLAGL |= (dst >> BIT_INDEX32(src)) & 1;
 }
 
 /*
@@ -135,136 +137,142 @@ BT_EdIb(DWORD op)
 void
 BTS_EwGw(void)
 {
-	WORD *out;
-	DWORD op, src, dst, madr, ad;
-	WORD bit;
+	UINT16 *out;
+	UINT32 op, src, dst, res, madr;
+	UINT16 bit;
 
 	PREPART_EA_REG16(op, src);
-	bit = 1 << (src & 0x0f);
+	bit = BIT_MAKEBIT16(src);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		out = reg16_b20[op];
-		if (*out & bit) {
+		dst = *out;
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
-			*out |= bit;
 			CPU_FLAGL &= ~C_FLAG;
+			res = dst | bit;
+			*out = (UINT16)res;
 		}
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		ad += 2 * (src >> 4);
-		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, ad);
+		madr += BIT_OFFSET16(src);
+		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
 		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
-			dst |= bit;
-			cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, ad, dst);
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst | bit;
+		cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, (UINT16)res);
 	}
 }
 
 void
 BTS_EdGd(void)
 {
-	DWORD *out;
-	DWORD op, src, dst, madr, ad;
-	DWORD bit;
+	UINT32 *out;
+	UINT32 op, src, dst, res, madr;
+	UINT32 bit;
 
 	PREPART_EA_REG32(op, src);
-	bit = 1 << (src & 0x1f);
+	bit = BIT_MAKEBIT32(src);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		out = reg32_b20[op];
-		if (*out & bit) {
+		dst = *out;
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
-			*out |= bit;
 			CPU_FLAGL &= ~C_FLAG;
+			res = dst | bit;
+			*out = res;
 		}
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		ad += 4 * (src >> 5);
-		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, ad);
+		madr += BIT_OFFSET32(src);
+		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
 		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
-			dst |= bit;
-			cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, ad, dst);
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst | bit;
+		cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, madr, res);
 	}
 }
 
 void
-BTS_EwIb(DWORD op)
+BTS_EwIb(UINT32 op)
 {
-	WORD *out;
-	DWORD src, dst, madr, ad;
-	WORD bit;
+	UINT16 *out;
+	UINT32 src, dst, res, madr;
+	UINT16 bit;
 
-	GET_PCBYTE(src);
-	bit = 1 << (src & 0x0f);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
+		GET_PCBYTE(src);
 		out = reg16_b20[op];
-		if (*out & bit) {
+		dst = *out;
+		bit = BIT_MAKEBIT16(src);
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
-			*out |= bit;
 			CPU_FLAGL &= ~C_FLAG;
+			res = dst | bit;
+			*out = (UINT16)res;
 		}
 	} else {
 		CPU_WORKCLOCK(6);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		ad += 2 * (src >> 4);
-		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, ad);
+		GET_PCBYTE(src);
+		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
+		bit = BIT_MAKEBIT16(src);
 		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
-			dst |= bit;
-			cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, ad, dst);
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst | bit;
+		cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, (UINT16)res);
 	}
 }
 
 void
-BTS_EdIb(DWORD op)
+BTS_EdIb(UINT32 op)
 {
-	DWORD *out;
-	DWORD src, dst, madr, ad;
-	DWORD bit;
+	UINT32 *out;
+	UINT32 src, dst, res, madr;
+	UINT32 bit;
 
-	GET_PCBYTE(src);
-	bit = 1 << (src & 0x1f);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
+		GET_PCBYTE(src);
 		out = reg32_b20[op];
-		if (*out & bit) {
+		dst = *out;
+		bit = BIT_MAKEBIT32(src);
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
-			*out |= bit;
 			CPU_FLAGL &= ~C_FLAG;
+			res = dst | bit;
+			*out = res;
 		}
 	} else {
 		CPU_WORKCLOCK(6);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		ad += 4 * (src >> 5);
-		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, ad);
+		GET_PCBYTE(src);
+		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
+		bit = BIT_MAKEBIT32(src);
 		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
-			dst |= bit;
-			cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, ad, dst);
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst | bit;
+		cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, madr, res);
 	}
 }
 
@@ -274,273 +282,287 @@ BTS_EdIb(DWORD op)
 void
 BTR_EwGw(void)
 {
-	WORD *out;
-	DWORD op, src, dst, madr, ad;
-	WORD bit;
+	UINT16 *out;
+	UINT32 op, src, dst, res, madr;
+	UINT16 bit;
 
 	PREPART_EA_REG16(op, src);
-	bit = 1 << (src & 0x0f);
+	bit = BIT_MAKEBIT16(src);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		out = reg16_b20[op];
-		if (*out & bit) {
-			*out &= ~bit;
+		dst = *out;
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
+			res = dst & ~bit;
+			*out = (UINT16)res;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		ad += 2 * (src >> 4);
-		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, ad);
+		madr += BIT_OFFSET16(src);
+		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
 		if (dst & bit) {
-			dst &= ~bit;
-			cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, ad, dst);
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst & ~bit;
+		cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, (UINT16)res);
 	}
 }
 
 void
 BTR_EdGd(void)
 {
-	DWORD *out;
-	DWORD op, src, dst, madr, ad;
-	DWORD bit;
+	UINT32 *out;
+	UINT32 op, src, dst, res, madr;
+	UINT32 bit;
 
 	PREPART_EA_REG32(op, src);
-	bit = 1 << (src & 0x1f);
+	bit = BIT_MAKEBIT32(src);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		out = reg32_b20[op];
 		dst = *out;
-		if (*out & bit) {
-			*out &= ~bit;
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
+			res = dst & ~bit;
+			*out = res;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		ad += 4 * (src >> 5);
-		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, ad);
+		madr += BIT_OFFSET32(src);
+		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
 		if (dst & bit) {
-			dst &= ~bit;
-			cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, ad, dst);
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst & ~bit;
+		cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, madr, res);
 	}
 }
 
 void
-BTR_EwIb(DWORD op)
+BTR_EwIb(UINT32 op)
 {
-	WORD *out;
-	DWORD src, dst, madr, ad;
-	WORD bit;
+	UINT16 *out;
+	UINT32 src, dst, res, madr;
+	UINT16 bit;
 
-	GET_PCBYTE(src);
-	bit = 1 << (src & 0x0f);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
+		GET_PCBYTE(src);
 		out = reg16_b20[op];
-		if (*out & bit) {
-			*out &= ~bit;
+		dst = *out;
+		bit = BIT_MAKEBIT16(src);
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
+			res = dst & ~bit;
+			*out = (UINT16)res;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
 	} else {
 		CPU_WORKCLOCK(6);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		ad += 2 * (src >> 4);
-		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, ad);
+		GET_PCBYTE(src);
+		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
+		bit = BIT_MAKEBIT16(src);
 		if (dst & bit) {
-			dst &= ~bit;
-			cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, ad, dst);
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst & ~bit;
+		cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, (UINT16)res);
 	}
 }
 
 void
-BTR_EdIb(DWORD op)
+BTR_EdIb(UINT32 op)
 {
-	DWORD *out;
-	DWORD src, dst, madr, ad;
-	DWORD bit;
+	UINT32 *out;
+	UINT32 src, dst, res, madr;
+	UINT32 bit;
 
-	GET_PCBYTE(src);
-	bit = 1 << (src & 0x1f);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
+		GET_PCBYTE(src);
 		out = reg32_b20[op];
-		if (*out & bit) {
-			*out &= ~bit;
+		dst = *out;
+		bit = BIT_MAKEBIT32(src);
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
+			res = dst & ~bit;
+			*out = res;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
 	} else {
 		CPU_WORKCLOCK(6);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		ad += 4 * (src >> 5);
-		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, ad);
+		GET_PCBYTE(src);
+		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
+		bit = BIT_MAKEBIT32(src);
 		if (dst & bit) {
-			dst &= ~bit;
-			cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, ad, dst);
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst & ~bit;
+		cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, madr, res);
 	}
 }
 
-
 /*
  * BTC
  */
 void
 BTC_EwGw(void)
 {
-	WORD *out;
-	DWORD op, src, dst, madr, ad;
-	WORD bit;
+	UINT16 *out;
+	UINT32 op, src, dst, res, madr;
+	UINT16 bit;
 
 	PREPART_EA_REG16(op, src);
-	bit = 1 << (src & 0x0f);
+	bit = BIT_MAKEBIT16(src);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		out = reg16_b20[op];
-		if (*out & bit) {
+		dst = *out;
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
-		*out ^= bit;
+		res = dst ^ bit;
+		*out = (UINT16)res;
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		ad += 2 * (src >> 4);
-		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, ad);
-		cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, ad, dst ^ bit);
+		madr += BIT_OFFSET16(src);
+		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
 		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst ^ bit;
+		cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, (UINT16)res);
 	}
 }
 
 void
 BTC_EdGd(void)
 {
-	DWORD *out;
-	DWORD op, src, dst, madr, ad;
-	DWORD bit;
+	UINT32 *out;
+	UINT32 op, src, dst, res, madr;
+	UINT32 bit;
 
 	PREPART_EA_REG32(op, src);
-	bit = 1 << (src & 0x1f);
+	bit = BIT_MAKEBIT32(src);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		out = reg32_b20[op];
-		if (*out & bit) {
+		dst = *out;
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
-		*out ^= bit;
+		res = dst ^ bit;
+		*out = res;
 	} else {
 		CPU_WORKCLOCK(7);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		ad += 4 * (src >> 5);
-		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, ad);
-		cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, ad, dst ^ bit);
+		madr += BIT_OFFSET32(src);
+		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
 		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst ^ bit;
+		cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, madr, res);
 	}
 }
 
 void
-BTC_EwIb(DWORD op)
+BTC_EwIb(UINT32 op)
 {
-	WORD *out;
-	DWORD src, dst, madr, ad;
-	WORD bit;
+	UINT16 *out;
+	UINT32 src, dst, res, madr;
+	UINT16 bit;
 
-	GET_PCBYTE(src);
-	bit = 1 << (src & 0x0f);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
+		GET_PCBYTE(src);
 		out = reg16_b20[op];
-		if (*out & bit) {
+		dst = *out;
+		bit = BIT_MAKEBIT16(src);
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
-		*out ^= bit;
+		res = dst ^ bit;
+		*out = (UINT16)res;
 	} else {
 		CPU_WORKCLOCK(6);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
-		ad += 2 * (src >> 4);
-		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, ad);
-		cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, ad, dst ^ bit);
+		GET_PCBYTE(src);
+		dst = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
+		bit = BIT_MAKEBIT16(src);
 		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst ^ bit;
+		cpu_vmemorywrite_w(CPU_INST_SEGREG_INDEX, madr, (UINT16)res);
 	}
 }
 
 void
-BTC_EdIb(DWORD op)
+BTC_EdIb(UINT32 op)
 {
-	DWORD *out;
-	DWORD src, dst, madr, ad;
-	DWORD bit;
+	UINT32 *out;
+	UINT32 src, dst, res, madr;
+	UINT32 bit;
 
-	GET_PCBYTE(src);
-	bit = 1 << (src & 0x1f);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
+		GET_PCBYTE(src);
 		out = reg32_b20[op];
-		if (*out & bit) {
+		dst = *out;
+		bit = BIT_MAKEBIT32(src);
+		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
-		*out ^= bit;
+		res = dst ^ bit;
+		*out = res;
 	} else {
 		CPU_WORKCLOCK(6);
 		madr = calc_ea_dst(op);
-		ad = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
-		ad += 4 * (src >> 5);
-		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, ad);
-		cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, ad, dst ^ bit);
+		GET_PCBYTE(src);
+		dst = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
+		bit = BIT_MAKEBIT32(src);
 		if (dst & bit) {
 			CPU_FLAGL |= C_FLAG;
 		} else {
 			CPU_FLAGL &= ~C_FLAG;
 		}
+		res = dst ^ bit;
+		cpu_vmemorywrite_d(CPU_INST_SEGREG_INDEX, madr, res);
 	}
 }
 
@@ -550,8 +572,8 @@ BTC_EdIb(DWORD op)
 void
 BSF_GwEw(void)
 {
-	WORD *out;
-	DWORD op, src;
+	UINT16 *out;
+	UINT32 op, src;
 	int bit;
 
 	PREPART_REG16_EA(op, src, out, 2, 7);
@@ -559,18 +581,20 @@ BSF_GwEw(void)
 		CPU_FLAGL |= Z_FLAG;
 		/* dest reg is undefined */
 	} else {
-		for (bit = 0; (bit < 16) && !(src & (1 << bit)); bit++)
-			continue;
-		*out = (WORD)bit;
 		CPU_FLAGL &= ~Z_FLAG;
+		for (bit = 0; bit < 15; bit++) {
+			if (src & (1 << bit))
+				break;
+		}
+		*out = (UINT16)bit;
 	}
 }
 
 void
 BSF_GdEd(void)
 {
-	DWORD *out;
-	DWORD op, src;
+	UINT32 *out;
+	UINT32 op, src;
 	int bit;
 
 	PREPART_REG32_EA(op, src, out, 2, 7);
@@ -578,10 +602,12 @@ BSF_GdEd(void)
 		CPU_FLAGL |= Z_FLAG;
 		/* dest reg is undefined */
 	} else {
-		for (bit = 0; (bit < 32) && !(src & (1 << bit)); bit++)
-			continue;
-		*out = (DWORD)bit;
 		CPU_FLAGL &= ~Z_FLAG;
+		for (bit = 0; bit < 31; bit++) {
+			if (src & (1 << bit))
+				break;
+		}
+		*out = (UINT32)bit;
 	}
 }
 
@@ -591,8 +617,8 @@ BSF_GdEd(void)
 void
 BSR_GwEw(void)
 {
-	WORD *out;
-	DWORD op, src;
+	UINT16 *out;
+	UINT32 op, src;
 	int bit;
 
 	PREPART_REG16_EA(op, src, out, 2, 7);
@@ -600,18 +626,20 @@ BSR_GwEw(void)
 		CPU_FLAGL |= Z_FLAG;
 		/* dest reg is undefined */
 	} else {
-		for (bit = 15; (bit >= 0) && !(src & (1 << bit)); bit--)
-			continue;
-		*out = (WORD)bit;
 		CPU_FLAGL &= ~Z_FLAG;
+		for (bit = 15; bit > 0; bit--) {
+			if (src & (1 << bit))
+				break;
+		}
+		*out = (UINT16)bit;
 	}
 }
 
 void
 BSR_GdEd(void)
 {
-	DWORD *out;
-	DWORD op, src;
+	UINT32 *out;
+	UINT32 op, src;
 	int bit;
 
 	PREPART_REG32_EA(op, src, out, 2, 7);
@@ -619,10 +647,12 @@ BSR_GdEd(void)
 		CPU_FLAGL |= Z_FLAG;
 		/* dest reg is undefined */
 	} else {
-		for (bit = 31; (bit >= 0) && !(src & (1 << bit)); bit--)
-			continue;
-		*out = (DWORD)bit;
 		CPU_FLAGL &= ~Z_FLAG;
+		for (bit = 31; bit > 0; bit--) {
+			if (src & (1 << bit))
+				break;
+		}
+		*out = (UINT32)bit;
 	}
 }
 
@@ -632,10 +662,10 @@ BSR_GdEd(void)
 void
 SETO_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = CPU_OV ? 1 : 0;
+	UINT32 op, madr;
+	UINT8 v = CPU_OV ? 1 : 0;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -649,10 +679,10 @@ SETO_Eb(void)
 void
 SETNO_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = CPU_OV ? 0 : 1;
+	UINT32 op, madr;
+	UINT8 v = CPU_OV ? 0 : 1;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -666,10 +696,10 @@ SETNO_Eb(void)
 void
 SETC_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = CPU_FLAGL & C_FLAG;
+	UINT32 op, madr;
+	UINT8 v = (UINT8)(CPU_FLAGL & C_FLAG);
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -683,10 +713,10 @@ SETC_Eb(void)
 void
 SETNC_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (CPU_FLAGL & C_FLAG) ? 0 : 1;
+	UINT32 op, madr;
+	UINT8 v = (CPU_FLAGL & C_FLAG) ? 0 : 1;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -700,10 +730,10 @@ SETNC_Eb(void)
 void
 SETZ_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (CPU_FLAGL & Z_FLAG) ? 1 : 0;
+	UINT32 op, madr;
+	UINT8 v = (CPU_FLAGL & Z_FLAG) ? 1 : 0;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -717,10 +747,10 @@ SETZ_Eb(void)
 void
 SETNZ_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (CPU_FLAGL & Z_FLAG) ? 0 : 1;
+	UINT32 op, madr;
+	UINT8 v = (CPU_FLAGL & Z_FLAG) ? 0 : 1;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -734,10 +764,10 @@ SETNZ_Eb(void)
 void
 SETA_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (CPU_FLAGL & (Z_FLAG|C_FLAG)) ? 0 : 1;
+	UINT32 op, madr;
+	UINT8 v = (CPU_FLAGL & (Z_FLAG|C_FLAG)) ? 0 : 1;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -751,10 +781,10 @@ SETA_Eb(void)
 void
 SETNA_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (CPU_FLAGL & (Z_FLAG|C_FLAG)) ? 1 : 0;
+	UINT32 op, madr;
+	UINT8 v = (CPU_FLAGL & (Z_FLAG|C_FLAG)) ? 1 : 0;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -768,10 +798,10 @@ SETNA_Eb(void)
 void
 SETS_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (CPU_FLAGL & S_FLAG) ? 1 : 0;
+	UINT32 op, madr;
+	UINT8 v = (CPU_FLAGL & S_FLAG) ? 1 : 0;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -785,10 +815,10 @@ SETS_Eb(void)
 void
 SETNS_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (CPU_FLAGL & S_FLAG) ? 0 : 1;
+	UINT32 op, madr;
+	UINT8 v = (CPU_FLAGL & S_FLAG) ? 0 : 1;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -802,10 +832,10 @@ SETNS_Eb(void)
 void
 SETP_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (CPU_FLAGL & P_FLAG) ? 1 : 0;
+	UINT32 op, madr;
+	UINT8 v = (CPU_FLAGL & P_FLAG) ? 1 : 0;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -819,10 +849,10 @@ SETP_Eb(void)
 void
 SETNP_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (CPU_FLAGL & P_FLAG) ? 0 : 1;
+	UINT32 op, madr;
+	UINT8 v = (CPU_FLAGL & P_FLAG) ? 0 : 1;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -836,10 +866,10 @@ SETNP_Eb(void)
 void
 SETL_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (!CPU_OV == !(CPU_FLAGL & S_FLAG)) ? 1 : 0;
+	UINT32 op, madr;
+	UINT8 v = (!CPU_OV == !(CPU_FLAGL & S_FLAG)) ? 1 : 0;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -853,10 +883,10 @@ SETL_Eb(void)
 void
 SETNL_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = (!CPU_OV == !(CPU_FLAGL & S_FLAG)) ? 0 : 1;
+	UINT32 op, madr;
+	UINT8 v = (!CPU_OV == !(CPU_FLAGL & S_FLAG)) ? 0 : 1;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -870,10 +900,10 @@ SETNL_Eb(void)
 void
 SETLE_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = ((CPU_FLAGL & Z_FLAG) || (!CPU_OV == !(CPU_FLAGL & S_FLAG))) ? 1 : 0;
+	UINT32 op, madr;
+	UINT8 v = ((CPU_FLAGL & Z_FLAG) || (!CPU_OV == !(CPU_FLAGL & S_FLAG))) ? 1 : 0;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -887,10 +917,10 @@ SETLE_Eb(void)
 void
 SETNLE_Eb(void)
 {
-	DWORD op, src, madr;
-	BYTE v = ((CPU_FLAGL & Z_FLAG) || (!CPU_OV == !(CPU_FLAGL & S_FLAG))) ? 0 : 1;
+	UINT32 op, madr;
+	UINT8 v = ((CPU_FLAGL & Z_FLAG) || (!CPU_OV == !(CPU_FLAGL & S_FLAG))) ? 0 : 1;
 
-	PREPART_EA_REG8(op, src);
+	GET_PCBYTE(op);
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
 		*(reg8_b20[op]) = v;
@@ -907,7 +937,7 @@ SETNLE_Eb(void)
 void
 TEST_EbGb(void)
 {
-	DWORD op, src, tmp, madr;
+	UINT32 op, src, tmp, madr;
 
 	PREPART_EA_REG8(op, src);
 	if (op >= 0xc0) {
@@ -918,13 +948,13 @@ TEST_EbGb(void)
 		madr = calc_ea_dst(op);
 		tmp = cpu_vmemoryread(CPU_INST_SEGREG_INDEX, madr);
 	}
-	ANDBYTE(tmp, src);
+	AND_BYTE(tmp, src);
 }
 
 void
 TEST_EwGw(void)
 {
-	DWORD op, src, tmp, madr;
+	UINT32 op, src, tmp, madr;
 
 	PREPART_EA_REG16(op, src);
 	if (op >= 0xc0) {
@@ -935,13 +965,13 @@ TEST_EwGw(void)
 		madr = calc_ea_dst(op);
 		tmp = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
 	}
-	ANDWORD(tmp, src);
+	AND_WORD(tmp, src);
 }
 
 void
 TEST_EdGd(void)
 {
-	DWORD op, src, tmp, madr;
+	UINT32 op, src, tmp, madr;
 
 	PREPART_EA_REG32(op, src);
 	if (op >= 0xc0) {
@@ -952,46 +982,46 @@ TEST_EdGd(void)
 		madr = calc_ea_dst(op);
 		tmp = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
 	}
-	ANDDWORD(tmp, src);
+	AND_DWORD(tmp, src);
 }
 
 void
 TEST_ALIb(void)
 {
-	DWORD src, tmp;
+	UINT32 src, tmp;
 
 	CPU_WORKCLOCK(3);
 	tmp = CPU_AL;
 	GET_PCBYTE(src);
-	ANDBYTE(tmp, src);
+	AND_BYTE(tmp, src);
 }
 
 void
 TEST_AXIw(void)
 {
-	DWORD src, tmp;
+	UINT32 src, tmp;
 
 	CPU_WORKCLOCK(3);
 	tmp = CPU_AX;
 	GET_PCWORD(src);
-	ANDWORD(tmp, src);
+	AND_WORD(tmp, src);
 }
 
 void
 TEST_EAXId(void)
 {
-	DWORD src, tmp;
+	UINT32 src, tmp;
 
 	CPU_WORKCLOCK(3);
 	tmp = CPU_EAX;
 	GET_PCDWORD(src);
-	ANDDWORD(tmp, src);
+	AND_DWORD(tmp, src);
 }
 
 void
-TEST_EbIb(DWORD op)
+TEST_EbIb(UINT32 op)
 {
-	DWORD src, tmp, madr;
+	UINT32 src, tmp, madr;
 
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
@@ -1002,13 +1032,13 @@ TEST_EbIb(DWORD op)
 		tmp = cpu_vmemoryread(CPU_INST_SEGREG_INDEX, madr);
 	}
 	GET_PCBYTE(src);
-	ANDBYTE(tmp, src);
+	AND_BYTE(tmp, src);
 }
 
 void
-TEST_EwIw(DWORD op)
+TEST_EwIw(UINT32 op)
 {
-	DWORD src, tmp, madr;
+	UINT32 src, tmp, madr;
 
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
@@ -1019,13 +1049,13 @@ TEST_EwIw(DWORD op)
 		tmp = cpu_vmemoryread_w(CPU_INST_SEGREG_INDEX, madr);
 	}
 	GET_PCWORD(src);
-	ANDWORD(tmp, src);
+	AND_WORD(tmp, src);
 }
 
 void
-TEST_EdId(DWORD op)
+TEST_EdId(UINT32 op)
 {
-	DWORD src, tmp, madr;
+	UINT32 src, tmp, madr;
 
 	if (op >= 0xc0) {
 		CPU_WORKCLOCK(2);
@@ -1036,5 +1066,5 @@ TEST_EdId(DWORD op)
 		tmp = cpu_vmemoryread_d(CPU_INST_SEGREG_INDEX, madr);
 	}
 	GET_PCDWORD(src);
-	ANDDWORD(tmp, src);
+	AND_DWORD(tmp, src);
 }