ARM/THUMB long calls stubs

[Richard Earnshaw]

> Hi Andy,
> 
> > By way of a suggestion, how about the _call_via.. functions be made 
> > inline?
> 
> Ah no, you see the entire point of the __call_via functions is that
> they are reached via the BL instruction so that the link register
> (r14) contains the return address.  Then the code can just use the BX 
> instruction to jump to the address held in the relevant register.
> (The BLX instruction for the Thumb only exists in v5 based
> architectures).  If you inline the function you will eliminate the
> initialisation of the link register and that would be bad.
> 
> Of course what you could do is to copy the PC into r14, adjust it as
> necessary and then do the jump, as in:
> 
>         BL  0
>         BX  <register_rX>

This won't work, because lr will point to the BX instruction.  The 
following would, but it still doesn't solve the overall problem of 
unreachable libcalls.

	bl	1f
1:
	add	lr, lr, #4
	bx	Rn

I really don't like this, though.  It makes every call-via site 8 bytes 
long.  Thumb code is about reducing code size and if that means calling a 
helper function, then so be it -- at least those can be shared.

R.