--- a/gen_x86.c	Tue Nov 27 22:54:38 2012 -0800
+++ b/gen_x86.c	Tue Dec 04 19:13:12 2012 -0800
@@ -31,6 +31,8 @@
 #define OP_RETN 0xC3
 #define OP_MOV_IEA 0xC6
 #define OP_CALL 0xE8
+#define OP_JMP 0xE9
+#define OP_JMP_BYTE 0xEB
 #define OP_CALL_EA 0xFF
 
 #define OP2_JCC 0x80
@@ -49,12 +51,6 @@
 #define BIT_DIR 0x2
 #define BIT_SIZE 0x1
 
-#define M68K_N_REG RBX
-#define M68K_V_REG BH
-#define M68K_Z_REG RDX
-#define M68K_C_REG DH
-
-#define M68K_SCRATCH RCX
 
 enum {
 	X86_RAX = 0,
@@ -79,13 +75,6 @@
 	X86_R15
 } x86_regs_enc;
 
-enum {
-	MODE_REG_INDIRECT = 0,
-	MODE_REG_DISPLACE8 = 0x40,
-	MODE_REG_DIPSLACE32 = 0x80,
-	MODE_REG_DIRECT = 0xC0
-} x86_modes;
-
 uint8_t * x86_rr_sizedir(uint8_t * out, uint8_t opcode, uint8_t src, uint8_t dst, uint8_t size)
 {
 	//TODO: Deal with the fact that AH, BH, CH and DH can only be in the R/M param when there's a REX prefix
@@ -160,6 +149,10 @@
 	}
 	*(out++) = opcode | dir;
 	*(out++) = MODE_REG_DISPLACE8 | base | (reg << 3);
+	if (base == RSP) {
+		//add SIB byte, with no index and RSP as base
+		*(out++) = (RSP << 3) | RSP;
+	}
 	*(out++) = disp;
 	return out;
 }
@@ -195,6 +188,10 @@
 	}
 	*(out++) = opcode | dir;
 	*(out++) = MODE_REG_INDIRECT | base | (reg << 3);
+	if (base == RSP) {
+		//add SIB byte, with no index and RSP as base
+		*(out++) = (RSP << 3) | RSP;
+	}
 	return out;
 }
 
@@ -251,6 +248,51 @@
 	return out;
 }
 
+uint8_t * x86_irdisp8(uint8_t * out, uint8_t opcode, uint8_t op_ex, int32_t val, uint8_t dst, int8_t disp, uint8_t size)
+{
+	uint8_t sign_extend = 0;
+	if ((size == SZ_D || size == SZ_Q) && val <= 0x7F && val >= -0x80) {
+		sign_extend = 1;
+		opcode |= BIT_DIR;
+	}
+	if (size == SZ_W) {
+		*(out++) = PRE_SIZE;
+	}
+
+	if (size == SZ_Q || dst >= R8 || (size == SZ_B && dst >= RSP && dst <= RDI)) {
+		*out = PRE_REX;
+		if (size == SZ_Q) {
+			*out |= REX_QUAD;
+		}
+		if (dst >= R8) {
+			*out |= REX_RM_FIELD;
+			dst -= (R8 - X86_R8);
+		}
+		out++;
+	}
+	if (dst >= AH && dst <= BH) {
+		dst -= (AH-X86_AH);
+	}
+	if (size != SZ_B) {
+		opcode |= BIT_SIZE;
+	}
+	*(out++) = opcode;
+	*(out++) = MODE_REG_DISPLACE8 | dst | (op_ex << 3);
+	*(out++) = disp;
+	*(out++) = val;
+	if (size != SZ_B && !sign_extend) {
+		val >>= 8;
+		*(out++) = val;
+		if (size != SZ_W) {
+			val >>= 8;
+			*(out++) = val;
+			val >>= 8;
+			*(out++) = val;
+		}
+	}
+	return out;
+}
+
 
 uint8_t * add_rr(uint8_t * out, uint8_t src, uint8_t dst, uint8_t size)
 {
@@ -262,6 +304,21 @@
 	return x86_ir(out, OP_IMMED_ARITH, OP_EX_ADDI, OP_ADD, val, dst, size);
 }
 
+uint8_t * add_irdisp8(uint8_t * out, int32_t val, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_irdisp8(out, OP_IMMED_ARITH, OP_EX_ADDI, val, dst_base, disp, size);
+}
+
+uint8_t * add_rrdisp8(uint8_t * out, uint8_t src, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_ADD, src, dst_base, disp, size, 0);
+}
+
+uint8_t * add_rdisp8r(uint8_t * out, uint8_t src_base, int8_t disp, uint8_t dst, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_ADD, dst, src_base, disp, size, BIT_DIR);
+}
+
 uint8_t * or_rr(uint8_t * out, uint8_t src, uint8_t dst, uint8_t size)
 {
 	return x86_rr_sizedir(out, OP_OR, src, dst, size);
@@ -271,6 +328,21 @@
 	return x86_ir(out, OP_IMMED_ARITH, OP_EX_ORI, OP_OR, val, dst, size);
 }
 
+uint8_t * or_irdisp8(uint8_t * out, int32_t val, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_irdisp8(out, OP_IMMED_ARITH, OP_EX_ORI, val, dst_base, disp, size);
+}
+
+uint8_t * or_rrdisp8(uint8_t * out, uint8_t src, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_OR, src, dst_base, disp, size, 0);
+}
+
+uint8_t * or_rdisp8r(uint8_t * out, uint8_t src_base, int8_t disp, uint8_t dst, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_OR, dst, src_base, disp, size, BIT_DIR);
+}
+
 uint8_t * and_rr(uint8_t * out, uint8_t src, uint8_t dst, uint8_t size)
 {
 	return x86_rr_sizedir(out, OP_AND, src, dst, size);
@@ -281,6 +353,21 @@
 	return x86_ir(out, OP_IMMED_ARITH, OP_EX_ANDI, OP_AND, val, dst, size);
 }
 
+uint8_t * and_irdisp8(uint8_t * out, int32_t val, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_irdisp8(out, OP_IMMED_ARITH, OP_EX_ANDI, val, dst_base, disp, size);
+}
+
+uint8_t * and_rrdisp8(uint8_t * out, uint8_t src, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_AND, src, dst_base, disp, size, 0);
+}
+
+uint8_t * and_rdisp8r(uint8_t * out, uint8_t src_base, int8_t disp, uint8_t dst, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_AND, dst, src_base, disp, size, BIT_DIR);
+}
+
 uint8_t * xor_rr(uint8_t * out, uint8_t src, uint8_t dst, uint8_t size)
 {
 	return x86_rr_sizedir(out, OP_XOR, src, dst, size);
@@ -291,6 +378,21 @@
 	return x86_ir(out, OP_IMMED_ARITH, OP_EX_XORI, OP_XOR, val, dst, size);
 }
 
+uint8_t * xor_irdisp8(uint8_t * out, int32_t val, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_irdisp8(out, OP_IMMED_ARITH, OP_EX_XORI, val, dst_base, disp, size);
+}
+
+uint8_t * xor_rrdisp8(uint8_t * out, uint8_t src, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_XOR, src, dst_base, disp, size, 0);
+}
+
+uint8_t * xor_rdisp8r(uint8_t * out, uint8_t src_base, int8_t disp, uint8_t dst, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_XOR, dst, src_base, disp, size, BIT_DIR);
+}
+
 uint8_t * sub_rr(uint8_t * out, uint8_t src, uint8_t dst, uint8_t size)
 {
 	return x86_rr_sizedir(out, OP_SUB, src, dst, size);
@@ -301,6 +403,21 @@
 	return x86_ir(out, OP_IMMED_ARITH, OP_EX_SUBI, OP_SUB, val, dst, size);
 }
 
+uint8_t * sub_irdisp8(uint8_t * out, int32_t val, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_irdisp8(out, OP_IMMED_ARITH, OP_EX_SUBI, val, dst_base, disp, size);
+}
+
+uint8_t * sub_rrdisp8(uint8_t * out, uint8_t src, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_SUB, src, dst_base, disp, size, 0);
+}
+
+uint8_t * sub_rdisp8r(uint8_t * out, uint8_t src_base, int8_t disp, uint8_t dst, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_SUB, dst, src_base, disp, size, BIT_DIR);
+}
+
 uint8_t * cmp_rr(uint8_t * out, uint8_t src, uint8_t dst, uint8_t size)
 {
 	return x86_rr_sizedir(out, OP_CMP, src, dst, size);
@@ -311,6 +428,21 @@
 	return x86_ir(out, OP_IMMED_ARITH, OP_EX_CMPI, OP_CMP, val, dst, size);
 }
 
+uint8_t * cmp_irdisp8(uint8_t * out, int32_t val, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_irdisp8(out, OP_IMMED_ARITH, OP_EX_CMPI, val, dst_base, disp, size);
+}
+
+uint8_t * cmp_rrdisp8(uint8_t * out, uint8_t src, uint8_t dst_base, int8_t disp, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_CMP, src, dst_base, disp, size, 0);
+}
+
+uint8_t * cmp_rdisp8r(uint8_t * out, uint8_t src_base, int8_t disp, uint8_t dst, uint8_t size)
+{
+	return x86_rrdisp8_sizedir(out, OP_CMP, dst, src_base, disp, size, BIT_DIR);
+}
+
 uint8_t * mov_rr(uint8_t * out, uint8_t src, uint8_t dst, uint8_t size)
 {
 	return x86_rr_sizedir(out, OP_MOV, src, dst, size);
@@ -360,12 +492,12 @@
 		dst -= (AH-X86_AH);
 	}
 	if (size == SZ_B) {
-		*(out++) = OP_MOV_I8R;
+		*(out++) = OP_MOV_I8R | dst;
 	} else if (size == SZ_Q && sign_extend) {
 		*(out++) = OP_MOV_IEA | BIT_SIZE;
 		*(out++) = MODE_REG_DIRECT | dst;
 	} else {
-		*(out++) = OP_MOV_IR;
+		*(out++) = OP_MOV_IR | dst;
 	}
 	*(out++) = val;
 	if (size != SZ_B) {
@@ -391,6 +523,43 @@
 	return out;
 }
 
+uint8_t * mov_irdisp8(uint8_t * out, int32_t val, uint8_t dst, int8_t disp, uint8_t size)
+{
+	if (size == SZ_W) {
+		*(out++) = PRE_SIZE;
+	}
+	if (size == SZ_Q || dst >= R8 || (size == SZ_B && dst >= RSP && dst <= RDI)) {
+		*out = PRE_REX;
+		if (size == SZ_Q) {
+			*out |= REX_QUAD;
+		}
+		if (dst >= R8) {
+			*out |= REX_RM_FIELD;
+			dst -= (R8 - X86_R8);
+		}
+		out++;
+	}
+	if (dst >= AH && dst <= BH) {
+		dst -= (AH-X86_AH);
+	}
+	*(out++) = OP_MOV_IEA | (size == SZ_B ? 0 : BIT_SIZE);
+	*(out++) = MODE_REG_DISPLACE8 | dst;
+	*(out++) = disp;
+
+	*(out++) = val;
+	if (size != SZ_B) {
+		val >>= 8;
+		*(out++) = val;
+		if (size != SZ_W) {
+			val >>= 8;
+			*(out++) = val;
+			val >>= 8;
+			*(out++) = val;
+		}
+	}
+	return out;
+}
+
 uint8_t * pushf(uint8_t * out)
 {
 	*(out++) = OP_PUSHF;
@@ -451,21 +620,53 @@
 	return out;
 }
 
-uint8_t * jcc(uint8_t * out, uint8_t cc, int32_t disp)
+uint8_t * jcc(uint8_t * out, uint8_t cc, uint8_t * dest)
 {
+	ptrdiff_t disp = dest-(out+2);
 	if (disp <= 0x7F && disp >= -0x80) {
 		*(out++) = OP_JCC | cc;
 		*(out++) = disp;
 	} else {
-		*(out++) = PRE_2BYTE;
-		*(out++) = OP2_JCC | cc;
-		*(out++) = disp;
-		disp >>= 8;
+		disp = dest-(out+6);
+		if (disp <= 0x7FFFFFFF && disp >= -2147483648) {
+			*(out++) = PRE_2BYTE;
+			*(out++) = OP2_JCC | cc;
+			*(out++) = disp;
+			disp >>= 8;
+			*(out++) = disp;
+			disp >>= 8;
+			*(out++) = disp;
+			disp >>= 8;
+			*(out++) = disp;
+		} else {
+			printf("%p - %p = %lX\n", dest, out + 6, disp);
+			return NULL;
+		}
+	}
+	return out;
+}
+
+uint8_t * jmp(uint8_t * out, uint8_t * dest)
+{
+	ptrdiff_t disp = dest-(out+2);
+	if (disp <= 0x7F && disp >= -0x80) {
+		*(out++) = OP_JMP_BYTE;
 		*(out++) = disp;
-		disp >>= 8;
-		*(out++) = disp;
-		disp >>= 8;
-		*(out++) = disp;
+	} else {
+		disp = dest-(out+5);
+		if (disp <= 0x7FFFFFFF && disp >= -2147483648) {
+			*(out++) = OP_JMP;
+			*(out++) = disp;
+			disp >>= 8;
+			*(out++) = disp;
+			disp >>= 8;
+			*(out++) = disp;
+			disp >>= 8;
+			*(out++) = disp;
+		} else {
+			printf("%p - %p = %lX\n", dest, out + 6, disp);
+			return NULL;
+		}
 	}
 	return out;
 }
@@ -483,8 +684,8 @@
 		disp >>= 8;
 		*(out++) = disp;
 	} else {
-		//TODO: Implement far call
-		printf("%p - %p = %ld, %d, %d, %d\n", fun, out + 5, disp, disp <= 0x7FFFFFFF, disp >= (-2147483648), -2147483648);
+		//TODO: Implement far call???
+		printf("%p - %p = %lX\n", fun, out + 5, disp);
 		return NULL;
 	}
 	return out;