--- np2/i386c/ia32/instructions/data_trans.c	2003/12/08 00:55:32	1.1
+++ np2/i386c/ia32/instructions/data_trans.c	2004/02/05 16:43:45	1.6
@@ -1,4 +1,4 @@
-/*	$Id: data_trans.c,v 1.1 2003/12/08 00:55:32 yui Exp $	*/
+/*	$Id: data_trans.c,v 1.6 2004/02/05 16:43:45 monaka Exp $	*/
 
 /*
  * Copyright (c) 2003 NONAKA Kimihiro
@@ -121,7 +121,7 @@ MOV_EwSw(void)
 	BYTE idx;
 
 	GET_PCBYTE(op);
-	idx = (op >> 3) & 3;
+	idx = (op >> 3) & 7;
 	if (idx < CPU_SEGREG_NUM) {
 		src = CPU_REGS_SREG(idx);
 		if (op >= 0xc0) {
@@ -144,7 +144,7 @@ MOV_SwEw(void)
 	BYTE idx;
 
 	GET_PCBYTE(op);
-	idx = (op >> 3) & 3;
+	idx = (op >> 3) & 7;
 	if (idx != CPU_CS_INDEX && idx < CPU_SEGREG_NUM) {
 		if (op >= 0xc0) {
 			CPU_WORKCLOCK(2);
@@ -814,6 +814,34 @@ void XCHG_EDIEAX(void) { CPU_WORKCLOCK(3
 /*
  * BSWAP
  */
+#if defined(bswap32) && !defined(IA32_USE_ASM_BSWAP)
+#define	BSWAP_DWORD(v)	bswap32(v)
+#else	/* !bswap32 || IA32_USE_ASM_BSWAP */
+INLINE static DWORD
+BSWAP_DWORD(DWORD val)
+{
+#if defined(__GNUC__) && (defined(i386) || defined(__i386__))
+	__asm__ __volatile__ (
+#if defined(IA32_USE_ASM_BSWAP)
+		"bswap %0"
+#else	/* !IA32_USE_ASM_BSWAP */
+		"rorw $8, %w1\n\t"
+		"rorl $16, %1\n\t"
+		"rorw $8, %w1\n\t"
+#endif	/* IA32_USE_ASM_BSWAP */
+		: "=r" (val) : "0" (val));
+	return val;
+#else	/* !(__GNUC__ && (i386 || __i386__)) */
+	DWORD v;
+	v  = (val & 0x000000ff) << 24;
+	v |= (val & 0x0000ff00) << 8;
+	v |= (val & 0x00ff0000) >> 8;
+	v |= (val & 0xff000000) >> 24;
+	return v;
+#endif	/* __GNUC__ && (i386 || __i386__) */
+}
+#endif	/* bswap32 && !IA32_USE_ASM_BSWAP */
+
 void BSWAP_EAX(void) { CPU_WORKCLOCK(2); CPU_EAX = BSWAP_DWORD(CPU_EAX); }
 void BSWAP_ECX(void) { CPU_WORKCLOCK(2); CPU_ECX = BSWAP_DWORD(CPU_ECX); }
 void BSWAP_EDX(void) { CPU_WORKCLOCK(2); CPU_EDX = BSWAP_DWORD(CPU_EDX); }
@@ -1016,7 +1044,7 @@ void PUSH_AX(void) { CPU_WORKCLOCK(3); P
 void PUSH_CX(void) { CPU_WORKCLOCK(3); PUSH0_16(CPU_CX); }
 void PUSH_DX(void) { CPU_WORKCLOCK(3); PUSH0_16(CPU_DX); }
 void PUSH_BX(void) { CPU_WORKCLOCK(3); PUSH0_16(CPU_BX); }
-void PUSH_SP(void) { CPU_WORKCLOCK(3); SP_PUSH0_16(CPU_SP); }
+void PUSH_SP(void) { CPU_WORKCLOCK(3); SP_PUSH_16(CPU_SP); }
 void PUSH_BP(void) { CPU_WORKCLOCK(3); PUSH0_16(CPU_BP); }
 void PUSH_SI(void) { CPU_WORKCLOCK(3); PUSH0_16(CPU_SI); }
 void PUSH_DI(void) { CPU_WORKCLOCK(3); PUSH0_16(CPU_DI); }
@@ -1025,7 +1053,7 @@ void PUSH_EAX(void) { CPU_WORKCLOCK(3); 
 void PUSH_ECX(void) { CPU_WORKCLOCK(3); PUSH0_32(CPU_ECX); }
 void PUSH_EDX(void) { CPU_WORKCLOCK(3); PUSH0_32(CPU_EDX); }
 void PUSH_EBX(void) { CPU_WORKCLOCK(3); PUSH0_32(CPU_EBX); }
-void PUSH_ESP(void) { CPU_WORKCLOCK(3); ESP_PUSH0_32(CPU_ESP); }
+void PUSH_ESP(void) { CPU_WORKCLOCK(3); ESP_PUSH_32(CPU_ESP); }
 void PUSH_EBP(void) { CPU_WORKCLOCK(3); PUSH0_32(CPU_EBP); }
 void PUSH_ESI(void) { CPU_WORKCLOCK(3); PUSH0_32(CPU_ESI); }
 void PUSH_EDI(void) { CPU_WORKCLOCK(3); PUSH0_32(CPU_EDI); }
@@ -1071,7 +1099,7 @@ PUSH_Ib(void)
 
 	CPU_WORKCLOCK(3);
 	GET_PCBYTESD(val);
-	XPUSH0((WORD)val);
+	XPUSH0(val);
 }
 
 void
@@ -1115,7 +1143,7 @@ void POP_AX(void) { CPU_WORKCLOCK(5); PO
 void POP_CX(void) { CPU_WORKCLOCK(5); POP0_16(CPU_CX); }
 void POP_DX(void) { CPU_WORKCLOCK(5); POP0_16(CPU_DX); }
 void POP_BX(void) { CPU_WORKCLOCK(5); POP0_16(CPU_BX); }
-void POP_SP(void) { CPU_WORKCLOCK(5); SP_POP0_16(CPU_SP); }
+void POP_SP(void) { CPU_WORKCLOCK(5); SP_POP_16(CPU_SP); }
 void POP_BP(void) { CPU_WORKCLOCK(5); POP0_16(CPU_BP); }
 void POP_SI(void) { CPU_WORKCLOCK(5); POP0_16(CPU_SI); }
 void POP_DI(void) { CPU_WORKCLOCK(5); POP0_16(CPU_DI); }
@@ -1124,7 +1152,7 @@ void POP_EAX(void) { CPU_WORKCLOCK(5); P
 void POP_ECX(void) { CPU_WORKCLOCK(5); POP0_32(CPU_ECX); }
 void POP_EDX(void) { CPU_WORKCLOCK(5); POP0_32(CPU_EDX); }
 void POP_EBX(void) { CPU_WORKCLOCK(5); POP0_32(CPU_EBX); }
-void POP_ESP(void) { CPU_WORKCLOCK(5); ESP_POP0_32(CPU_ESP); }
+void POP_ESP(void) { CPU_WORKCLOCK(5); ESP_POP_32(CPU_ESP); }
 void POP_EBP(void) { CPU_WORKCLOCK(5); POP0_32(CPU_EBP); }
 void POP_ESI(void) { CPU_WORKCLOCK(5); POP0_32(CPU_ESI); }
 void POP_EDI(void) { CPU_WORKCLOCK(5); POP0_32(CPU_EDI); }
@@ -1316,6 +1344,7 @@ POP32_GS(void)
 void
 PUSHA(void)
 {
+	WORD sp = CPU_SP;
 
 	CPU_WORKCLOCK(17);
 	if (!CPU_STAT_SS32) {
@@ -1326,19 +1355,19 @@ PUSHA(void)
 		REGPUSH0(CPU_CX);
 		REGPUSH0(CPU_DX);
 		REGPUSH0(CPU_BX);
-		SP_PUSH0(CPU_SP);
+		REGPUSH0(sp);
 		REGPUSH0(CPU_BP);
 		REGPUSH0(CPU_SI);
 		REGPUSH0(CPU_DI);
 	} else {
 		if (CPU_STAT_PM) {
-			CHECK_STACK_PUSH(&CPU_STAT_SREG(CPU_SS_INDEX), CPU_ESP, 32);
+			CHECK_STACK_PUSH(&CPU_STAT_SREG(CPU_SS_INDEX), CPU_ESP, 16);
 		}
 		REGPUSH0_16_32(CPU_AX);
 		REGPUSH0_16_32(CPU_CX);
 		REGPUSH0_16_32(CPU_DX);
 		REGPUSH0_16_32(CPU_BX);
-		SP_PUSH0_16_32(CPU_SP);
+		REGPUSH0_16_32(sp);
 		REGPUSH0_16_32(CPU_BP);
 		REGPUSH0_16_32(CPU_SI);
 		REGPUSH0_16_32(CPU_DI);
@@ -1348,17 +1377,18 @@ PUSHA(void)
 void
 PUSHAD(void)
 {
+	DWORD esp = CPU_ESP;
 
 	CPU_WORKCLOCK(17);
 	if (!CPU_STAT_SS32) {
 		if (CPU_STAT_PM) {
-			CHECK_STACK_PUSH(&CPU_STAT_SREG(CPU_SS_INDEX), CPU_SP, 16);
+			CHECK_STACK_PUSH(&CPU_STAT_SREG(CPU_SS_INDEX), CPU_SP, 32);
 		}
 		REGPUSH0_32_16(CPU_EAX);
 		REGPUSH0_32_16(CPU_ECX);
 		REGPUSH0_32_16(CPU_EDX);
 		REGPUSH0_32_16(CPU_EBX);
-		ESP_PUSH0_32_16(CPU_ESP);
+		REGPUSH0_32_16(esp);
 		REGPUSH0_32_16(CPU_EBP);
 		REGPUSH0_32_16(CPU_ESI);
 		REGPUSH0_32_16(CPU_EDI);
@@ -1370,7 +1400,7 @@ PUSHAD(void)
 		REGPUSH0_32(CPU_ECX);
 		REGPUSH0_32(CPU_EDX);
 		REGPUSH0_32(CPU_EBX);
-		ESP_PUSH0_32(CPU_ESP);
+		REGPUSH0_32(esp);
 		REGPUSH0_32(CPU_EBP);
 		REGPUSH0_32(CPU_ESI);
 		REGPUSH0_32(CPU_EDI);
@@ -1396,7 +1426,7 @@ POPA(void)
 		REGPOP0(CPU_AX);
 	} else {
 		if (CPU_STAT_PM) {
-			CHECK_STACK_POP(&CPU_STAT_SREG(CPU_SS_INDEX), CPU_ESP, 32);
+			CHECK_STACK_POP(&CPU_STAT_SREG(CPU_SS_INDEX), CPU_ESP, 16);
 		}
 		REGPOP0_16_32(CPU_DI);
 		REGPOP0_16_32(CPU_SI);
@@ -1416,7 +1446,7 @@ POPAD(void)
 	CPU_WORKCLOCK(19);
 	if (!CPU_STAT_SS32) {
 		if (CPU_STAT_PM) {
-			CHECK_STACK_POP(&CPU_STAT_SREG(CPU_SS_INDEX), CPU_SP, 16);
+			CHECK_STACK_POP(&CPU_STAT_SREG(CPU_SS_INDEX), CPU_SP, 32);
 		}
 		REGPOP0_32_16(CPU_EDI);
 		REGPOP0_32_16(CPU_ESI);
@@ -1471,7 +1501,7 @@ IN_EAXDX(void)
 void
 IN_ALIb(void)
 {
-	BYTE port;
+	WORD port;
 
 	CPU_WORKCLOCK(5);
 	GET_PCBYTE(port);
@@ -1481,7 +1511,7 @@ IN_ALIb(void)
 void
 IN_AXIb(void)
 {
-	BYTE port;
+	WORD port;
 
 	CPU_WORKCLOCK(5);
 	GET_PCBYTE(port);
@@ -1491,7 +1521,7 @@ IN_AXIb(void)
 void
 IN_EAXIb(void)
 {
-	BYTE port;
+	WORD port;
 
 	CPU_WORKCLOCK(5);
 	GET_PCBYTE(port);
@@ -1528,7 +1558,7 @@ OUT_DXEAX(void)
 void
 OUT_IbAL(void)
 {
-	BYTE port;
+	WORD port;
 
 	CPU_WORKCLOCK(3);
 	GET_PCBYTE(port);
@@ -1538,7 +1568,7 @@ OUT_IbAL(void)
 void
 OUT_IbAX(void)
 {
-	BYTE port;
+	WORD port;
 
 	CPU_WORKCLOCK(3);
 	GET_PCBYTE(port);
@@ -1548,11 +1578,11 @@ OUT_IbAX(void)
 void
 OUT_IbEAX(void)
 {
-	BYTE port;
+	WORD port;
 
 	CPU_WORKCLOCK(3);
 	GET_PCBYTE(port);
-	cpu_out_w(port, CPU_EAX);
+	cpu_out_d(port, CPU_EAX);
 }
 
 /*
@@ -1616,6 +1646,16 @@ MOVSX_GwEb(void)
 }
 
 void
+MOVSX_GwEw(void)
+{
+	WORD *out;
+	DWORD op, src;
+
+	PREPART_REG16_EA(op, src, out, 2, 5);
+	*out = src;
+}
+
+void
 MOVSX_GdEb(void)
 {
 	DWORD *out;
@@ -1649,6 +1689,16 @@ MOVZX_GwEb(void)
 }
 
 void
+MOVZX_GwEw(void)
+{
+	WORD *out;
+	DWORD op, src;
+
+	PREPART_REG16_EA(op, src, out, 2, 5);
+	*out = src;
+}
+
+void
 MOVZX_GdEb(void)
 {
 	DWORD *out;