* 64bit integer mult with overflow checks + test bench
* correctly align value types on the stack and in arrays
* impl. marshalling attributes for pinvoke
* raise exceptions everywhere
* exceptions: save/restore floating point state
* implement all floating point instruction in x86.brg, we also need to check
  the floating branch instruction - some of them seems to be wrong, we need to
  write better tests for that.
* document the functions and variables in the JIT 
* implement a register set for FP registers (just count register usage on x86)

* inline all calls in System.Math::*, we should use coprocessor instructions
  directly whenever possible. 

* Calling conventions:

	Implement fast call

* Overflow conversions:

	Implement CONV_OVF_XXX (constant) variants that emit either
	loads or exceptions (as we can always tell)
	
	
* use short jump opcodes where possible (for backward jumps), will give better
  performance. 

* Memory allocation of dynamic code.

	The current interface to generate code and allocate the memory
	for it is not very nice.

	Use large buffers to allocate temporary code, rather than allocating
	256 bytes every time, and hoping that it will be enough.