#define UNLOCKED 0x0
#define LOCKED	0xffffffff

__declspec(naked) void HookFunction()
{
	__asm
	{
		pushf //4 bytes
		pusha //8 * 4 bytes = 32 bytes
		
		//Inlined TlsGetValue:
		mov		ebx,	fs:0x18 //Get linear address of ThreadInformationBlock (file:///N:/reference.html#ThreadInformationBlock)
		mov		ecx,	[tlsIndex]
		mov		eax,	[ebx + ecx*4 + 0xe10]

checkSemaphore:
		cmp		eax,	LOCKED
		jz		execute

		//set the semaphore:
		mov		eax,					LOCKED 
		mov		[ebx + ecx*4 + 0xe10],	eax

log:
		//Do actual logging via lookup table
		//Don't forget to retrieve 'stack-size' of hooked function for use in stack-shift.

shiftstack:		
		add		esp,	0x4 //add some more space for a copy of return address of the target.

		mov		ecx,	0xa //flags + registers + the a return address TODO: Add number of args.
		mov		edi,	esp
		sub		edi,	0x2c
		mov		esi,	edi
		mov		eax,	esi //Keep a copy of the original esp in eax.
		sub		esi,	0x4
		rep		movs	edi,	esi 
		
		//We've now shifted the entire 'state block' down on the stack, creating a void.
		/* In addition, in this instance, the data we want in the 'void' 
		 * (our target's return address)
		 * is already there ;)
		 */

		//now overwrite the return address to come back to returnHandler so we can unlock the hook.
		mov		[eax],		returnHandler
	
execute:
		//restore state:
		popa
		popf
		//Execute the overwritten prolog:
		push	ebp
		mov		ebp,	esp
		//return to the function:
		ret

returnHandler:
		pushf
		pusha

		//Inlined TlsGetValue:
		mov		ebx,	fs:0x18 //Get linear address of ThreadInformationBlock (file:///N:/reference.html#ThreadInformationBlock)
		mov		ecx,	[tlsIndex]
		//set the semaphore:
		mov		eax,					UNLOCKED 
		mov		[ebx + ecx*4 + 0xe10],	eax

		popf
		popa
		ret //to the original caller!
	}
}