Lpc2000

I did a simple comparison using the test program for our fast floating
point maths library and found that GCC generates THUMB code 30% smaller
than ARM code but takes 40% longer to execute.

Regards
Michael

ARM vs Thumb for Intel Strong ARM
This paper says that mixing the ARM and Thumb minimizes the code size 
and performance penalty.

http://www.cs.arizona.edu/people/gupta/research/Publications/Comp/L14-
krishnaswamy.pdf

The average performance hit in Thumb mode is about 20%.
One of the slides in this presentation compares various EEMBC 
benchmarks.

http://www.jp.arm.com/kk/arm_forum2003/ppt/architecture%
20_extensions.ppt

--- In lpc2100@yahoogroups.com, "Michael Johnson" <mpj@r...> wrote:
> I did a simple comparison using the test program for our fast 
floating
> point maths library and found that GCC generates THUMB code 30% 
smaller

> than ARM code but takes 40% longer to execute.
> 
> Regards
> Michael

--- In lpc2100@yahoogroups.com, "lpc2100" <lpc2100@y...> wrote:
> ARM vs Thumb for Intel Strong ARM
> This paper says that mixing the ARM and Thumb minimizes the code 
size and performance penalty.

Be sure to compare like with like; using Thumb code on a cached 
processor can change the cache behaviour quite considerably - you 
may have less cache-line evictions, or you may have adverse effects 
because of interworking or long-jump veneers taking up cache lines - 
now that isn't an issue with LPC, but any documents you find *may* 
be measuring a system with memory speeds quite unlike your own.

I'm bemused by the "ARM vs Thumb for Intel Strong ARM" tag - 
StrongARM doesn't have Thumb capability.

To get back on-topic, I strongly advise profiling *your* application 
to see which functions can win you back code space if compiled as 
Thumb code without adversely affecting performance, Michael's 
figures seem to be a good first-pass estimate for general use for 
people using the gnu compiler on LPC.

Peter.