Arm/RAM vs THUMB/ROM
Written by Tinara and Tony C.

It is recommended that you use 32-bit ARM instructions in RAM and 
16-bit THUMB instruction in ROM, which seems a bit backwards really 
as the longer instructions are better for the smaller amount of memory.

Relatively, ROM access times are butt slow, one reason that the 
"BIOS Shadowing" option is used to speed up PC's.  I'm guessing 
that at decent clock speeds, the ARM cpu has problems keeping 
its instruction pipeline full when fetching 32-bit instructions 
from ROM, THUMB instructions would allow it to fill the pipeline
roughly twice as much, meaning less pipeline stalls that essentially 
waste one or more CPU cycles.

CPUs generally follow a  3-step process when executing an instruction:
Fetch- Decode- Execute.  Now when you hear the term pipelining, it 
refers to the fact that modern CPUs can implement 2 or 3 stages at 
the same time. That is,  a CPU could be Executing  INSTR A,  while 
Decoding INSTR B,  and Fetching INSTR C, all simultaneously.   This 
obviously saves time over: FETCH A.  DECODE A.  EXECUTE A.   
FETCH B. DECODE B. EXECUTE B.  (which is how the "z80" on the GB(C) 
works).   Now from Nintendo warnings that you want to use 32 bit 
instructions internal and 16 bit external,   the bus internally is 
probably 32 bit and only 16 bit externally.   This means if you use 
ARM 32 bit instructions externally,  then the CPU has to hang around
and wait for 2 states while the FETCH gets  1/2 the instruction, 
then the other 1/2.    Obviously this cuts down on how effecient it 
can follow the above pipelining method.   The best it can hope to 
do is Execute A, while FETCHing B.