Fix #622

[bgregoir]

This patch as the advantage to fixes the current problem, but I think it is not the write way to do it.
I think, the best will be to introduce speudo instruction in arm:
x = #lea e
this one can be decomposed in a sequence of addition with shift.
ie x = #lea (y + z << i + c)
---->
x = #ADD(y, z << i);
x = #ADD(x, c);

x = #lea (y + c)
---->
x = #ADD(y, c); // if c is to large we can use x to compute it.

and
x = #GMOV e
this
can become either a LDR or a STR or a MOV depending of the source e and destination x.

The current patch poorly do that.

[bgregoir]

It seem that I did not start from the good branch...