some simple optimizations for speed. this improves performance about 5%

SStream.c

@@ -24,8 +24,11 @@ void SStream_Init(SStream *ss)
 void SStream_concat0(SStream *ss, char *s)
 {
 #ifndef CAPSTONE_DIET
-	strcpy(ss->buffer + ss->index, s);
-	ss->index += (int) strlen(s);
+	unsigned int len = strlen(s);
+
+	memcpy(ss->buffer + ss->index, s, len);
+	ss->index += len;
+	ss->buffer[ss->index] = '\0';
 #endif
 }
 

arch/X86/X86ATTInstPrinter.c

@@ -485,9 +485,9 @@ static void printOperand(MCInst *MI, unsigned OpNo, SStream *O)
 {
 	MCOperand *Op  = MCInst_getOperand(MI, OpNo);
 	if (MCOperand_isReg(Op)) {
-		printRegName(O, MCOperand_getReg(Op));
+		unsigned int reg = MCOperand_getReg(Op);
+		printRegName(O, reg);
 		if (MI->csh->detail) {
-			unsigned int reg = MCOperand_getReg(Op);
 			if (MI->csh->doing_mem) {
 				MI->flat_insn->detail->x86.operands[MI->flat_insn->detail->x86.op_count].mem.base = reg;
 			} else {

arch/X86/X86GenAsmWriter.inc

@@ -14113,8 +14113,9 @@ static void printInstruction(MCInst *MI, SStream *O, MCRegisterInfo *MRI)
 #endif
 
   // Emit the opcode for the instruction.
-  uint64_t Bits1 = OpInfo[MCInst_getOpcode(MI)];
-  uint64_t Bits2 = OpInfo2[MCInst_getOpcode(MI)];
+  unsigned int opcode = MCInst_getOpcode(MI);
+  uint64_t Bits1 = OpInfo[opcode];
+  uint64_t Bits2 = OpInfo2[opcode];
   uint64_t Bits = (Bits2 << 32) | Bits1;
   // assert(Bits != 0 && "Cannot print this instruction.");
   if (!X86_lockrep(MI, O))

arch/X86/X86GenAsmWriter1.inc

@@ -13817,8 +13817,9 @@ static void printInstruction(MCInst *MI, SStream *O, MCRegisterInfo *MRI)
 #endif
 
   // Emit the opcode for the instruction.
-  uint64_t Bits1 = OpInfo[MCInst_getOpcode(MI)];
-  uint64_t Bits2 = OpInfo2[MCInst_getOpcode(MI)];
+  unsigned int opcode = MCInst_getOpcode(MI);
+  uint64_t Bits1 = OpInfo[opcode];
+  uint64_t Bits2 = OpInfo2[opcode];
   uint64_t Bits = (Bits2 << 32) | Bits1;
   // assert(Bits != 0 && "Cannot print this instruction.");
   if (!X86_lockrep(MI, O))

arch/X86/X86IntelInstPrinter.c

@@ -547,9 +547,10 @@ static void printOperand(MCInst *MI, unsigned OpNo, SStream *O)
 {
 	MCOperand *Op  = MCInst_getOperand(MI, OpNo);
 	if (MCOperand_isReg(Op)) {
-		printRegName(O, MCOperand_getReg(Op));
+		unsigned int reg = MCOperand_getReg(Op);
+
+		printRegName(O, reg);
 		if (MI->csh->detail) {
-			unsigned int reg = MCOperand_getReg(Op);
 			if (MI->csh->doing_mem) {
 				MI->flat_insn->detail->x86.operands[MI->flat_insn->detail->x86.op_count].mem.base = reg;
 			} else {

arch/X86/X86Mapping.c

@@ -47163,11 +47163,6 @@ static insn_map insns[] = {	// reduce x86 instructions
 };
 #endif
 
-// post printer for X86. put all the hacky stuff here
-void X86_post_printer(csh handle, cs_insn *insn, char *insn_asm, MCInst *mci)
-{
-}
-
 // given internal insn id, return public instruction info
 void X86_get_insn_id(cs_struct *h, cs_insn *insn, unsigned int id)
 {

arch/X86/X86Mapping.h

@@ -28,9 +28,6 @@ const char *X86_insn_name(csh handle, unsigned int id);
 // return group name, given group id
 const char *X86_group_name(csh handle, unsigned int id);
 
-// post printer for X86.
-void X86_post_printer(csh handle, cs_insn *pub_insn, char *insn_asm, MCInst *mci);
-
 // return register of given instruction id
 // return 0 if not found
 // this is to handle instructions embedding accumulate registers into AsmStrs[]

arch/X86/X86Module.c

@@ -30,7 +30,7 @@ static cs_err init(cs_struct *ud)
 	ud->insn_id = X86_get_insn_id;
 	ud->insn_name = X86_insn_name;
 	ud->group_name = X86_group_name;
-	ud->post_printer = X86_post_printer;
+	ud->post_printer = NULL;;
 
 	if (ud->mode == CS_MODE_64)
 		ud->regsize_map = regsize_map_64;

cs.c

@@ -260,7 +260,7 @@ static void fill_insn(struct cs_struct *handle, cs_insn *insn, char *buffer, MCI
 #endif
 
 	// fill the instruction bytes
-	memcpy(insn->bytes, code, MIN(sizeof(insn->bytes), insn->size));
+	memcpy(insn->bytes, code, insn->size);
 
 	// map internal instruction opcode to public insn ID
 	handle->insn_id(handle, insn, MCInst_getOpcode(mci));
@@ -292,7 +292,6 @@ static void fill_insn(struct cs_struct *handle, cs_insn *insn, char *buffer, MCI
 
 	// copy @op_str
 	if (*sp) {
-		*sp = '\0';
 		// find the next non-space char
 		sp++;
 		for (; ((*sp == ' ') || (*sp == '\t')); sp++);