To use this:

• Configure it with cross-compilation enabled, then use any ARM target
• I've been using ./arm-linux-gnueabi-tcc -nostdinc -nostdlib -c test.c -o test.o to compile, and arm-none-eabi-objdump -S test.o to dump the output
• I was comparing it with what I got from arm-none-eabi-gcc -mlittle-endian -mthumb -mcpu=cortex-m3 -mfix-cortex-m3-ldrd -mfloat-abi=soft -nostdinc -nostdlib -c test.c -o test.o

In terms of implementation:

• Everything is in arm-gen.c.
• tcc-gen.c has a single mod, which adds one to the symbol address (signifying that the symbol is Thumb, not ARM). There must be a nicer way of doing this, but hey - proof of concept.
• I added ot() which outputs a thumb instruction - the original o() would output a 32 bit ARM instruction - so every call to o() needs replacing with a thumb equivalent
• The function prolog/epilog is hacked up at the moment so only works with very basic stuff (no stack pushing)
• encbranch/decbranch/gsym_addr do seem to work, but encbranch won't work reliably for conditional branches as these have less bits for the address

Would be nice:

• Ideally we'd start out with a new thumb-gen.c containing stubs - it would get rid of the RELICENSING issue
• Find a way to separate ARM and Thumb code gen (ideally separate files), and maybe make TCC respect the -mthumb switch like GCC does (rather than having a separate binary).

#!/bin/bash
arm-none-eabi-gcc -mlittle-endian -mthumb -mcpu=cortex-m3  -mfix-cortex-m3-ldrd -mfloat-abi=soft -nostdinc -nostdlib -c test.c -o test.o 
echo ------------------------------------------------
echo   GCC
echo ------------------------------------------------
arm-none-eabi-objdump -S test.o
mv test.o test-gcc.o


./arm-linux-gnueabi-tcc -nostdinc -nostdlib -c test.c -o test.o 
echo ------------------------------------------------
echo   TCC
echo ------------------------------------------------
arm-none-eabi-objdump -S test.o
mv test.o test-tcc.o

// test.c
int foobar() {
  return 4;
}

is here