[openrisc] gcc, data and bss segments

[Chris Ziomkowski<chris@asics.ws>]

In my attempts to port RTEMS to the or1k architecture, I have
been appalled at the size of the binaries that are resulting
from the compiled gcc code.

The RTEMS kernel that I have been working on, even completely
stripped of all symbols, is approximately 2 MB in size! In doing
some research into why this is so large, it has come to my
attention that about 10% of this figure is the result of large
buffers which have been inadvertently placed into the .data
segment instead of the .bss segment. This causes large blocks
of zeroes to appear in the binary image, when the fact is they
should be created when the program starts up.

There are several advantages to placing all variables that are
initialized to zero in the .bss segment. These include but are
not limited to:

Faster boot times. It is alot faster to simply create a RAM block
of zeroes than it is to copy a block of zeroes from flash to RAM.

Lower power consumption. Power can be saved by creating this block
internally, rather than driving signals offchip to get at a block
of zeroes stored in a flash every time a new process is loaded. This
is addative with the power saved by faster process startup times.

Smaller binaries. On the RTEMS image, the resulting binary will
be about 10% smaller. Other processes may yield different results.
This means smaller flash images (and less expensive chips) for the
same functionality and increased performance.

Since I didn't work on gcc, I am not sure what is involved in
making this change, or why it wasn't done originally. Can someone
comment on what it will take to make this change? While this is
not a crtical bug (it doesn't prevent meaningful work from being
accomplished) it is a serious issue which needs to be fixed.

I would really like to shrink my RTEMS image by about 50%, and
this seems to be a good way to recover alot of wasted space.

Thanks for any advice/asssitance,

Chris
chris@asics.ws





_______________________________________________________
www.asics.ws - Solutions for your ASIC needs
--
To unsubscribe from openrisc mailing list please visit http://www.opencores.org/mailinglists.shtml