GCC ARM assembly questions

[Phil Endecott]

Hi Priya,

Priya Narasimhan wrote:
> I am compiling relatively simple C programs using "gcc -S
> -fomit-frame-pointer -mcpu=xscale" on a gumstix (ARM Xscale) processor. I
> noticed that the generated assembly code has additional stuff in, beyond
> what I would typically see in a standard ARM assembly program (e.g., one
> that was generated instead through the ARM Development Suite toolchain). 
>
> For instance, I see the following comments in the GNU-generated assembly
> code. What do these signify? 
>
> @ zero_extendqisi2
> @ args =0, pretend =0 (what does pretend mean?)
> @ frame_needed = 0, uses_anonymous_args = 0 (what does uses_anonymous_args
> mean?)

I don't know, and I suspect that the answer is not very interesting.  
The best way to find answers to these questions is probably to study 
the gcc source code.

> Secondly, I notice that without the "-O" option, saving registers to the
> stack happens almost on every alternate line of the generated assembly code.
> What is the purpose of doing this?

Without -O, gcc runs faster.  It can run faster by not spending time 
working out what can be stored in registers, and just using the stack 
for everything.

> Is there a way of turning this feature off, without resorting to the "-O" option?

Why do you say "resorting"?  If you care about the efficiency of the 
generated code then you should be using -Os or -O3.  The only situation 
in which you might want to not specify any -O option is if you want to 
compile quickly, for some reason.  (Or possibly to make debugging 
easier to follow.)

> Finally, is there any documentation on what "-O" actually accomplishes for
> an ARM/Xscale processor? For instance, does it use the barrel shifter or
> conditional execution wherever possible? Does it optimize for space, for
> time, or for both?

Have you seen the page about "Options That Control Optimization" in the 
gcc docs?

To optimise for speed, use -O3.  To optimise for space, use -Os.

gcc will use the barrel shifter (i.e. use instructions that combine a 
shift and an arithmetic operation), though you might find that the ARM 
compiler uses more of these.  It will also use conditional exection, 
and there are specific options that enable/disable that; again, the ARM 
compiler may make more use of this feature.  If you find any real 
examples of code where the ARM compiler does significantly better 
overall, I would be interested to see them.


Regards,

Phil.