Thumb2 (16bit/32bit instructions) port
================================================================================

Building (cross-compiling on x86 linux):
#export CXX=arm-linux-gnueabihf-g++-4.7
#export LINK=arm-linux-gnueabihf-g++-4.7
#make dependencies
#make arm snapshot=off hardfp=on armthumb=on jitthumb=on

Additional options:
--enable-size-opt

Optimizes code for size. Currently this allows the use of IT blocks to suppress
the flag-setting behavior of short instructions. However this typically results
in worse performance.

Everything is implemented and works except:
- disassembler
- simulator

Big problems (missed opportunities/future work):
 - choice of r11 as fp is a bad one on thumb since most of the short instructions
   support only r[0-7].
 - choice of tagging scheme makes all fields offsets odd. Odd immediate offsets
   are not encodable in 16-bit load/stores (4-byte alignment is required). So
   all field accesses are 32-bit instructions.
 - most assembly code assumes the ALU instruction don't set flags, but most 16-bit
   instructions do. An IT block emission is required to suppress that. The code
   needs to be refactored to use SetCC by default.