First of all search for missing volatiles. Most but not all problems 
seen when higher optimization levels are enabled are caused by missing 
volatiles (in empty loops or when accessing data from ISRs and the 
"main-thread").

Maybe the LCD does not work since delay-loops are used which get 
"optimized away". Try to add volatile to the loop-counter declaration if 
not already there.

Do you used stdio? IRC the Keil 3.3.1 GNU package uses a different libc 
and different lowlevel interface than most other GNU packages which 
usually use newlib (like Yagarto).

The PLL-configuration should be independent from any optimization 
settings.

How are ISRs declared? Do you use the GCC IRQ attribute, nacked 
attribute with inline-assembler entry/exit code or an assembler-wrapper 
for IRQs?

Regarding 'volatile' read this: 
http://www.embedded.com/columns/beginerscorner/9900209

GCC 4.x performs stronger optimisation than 3.x. I would be surprised if 
it optimised away a variable access when no optimisation was set. That 
said, I would not be surprised if some latent bug or flaw in your code 
manifested itself differently when built with a diferent compiler - 
you'll just have to debug it, because the problem is most likly with 
your code rather then the tool.

Are you using the same CRT start-up code? If not, does the new code set 
the PLL up differently?

I have seen vastly differing performance between newlib builds for code 
using floating-point. You should avoid using floating-point in any case.

Clifford

Hi,

@mthomas
All the behaviours mentioned above are for no optimization level.
stdio is being used.

ISR attributes are declared like this:

1

__attribute__ ((interrupt));

@clifford
I'm using the same start-up code (*.s) and the same file for memory and 
sections definitions (*.ld) as before.


Thanks you both, I've now some topics to explore.

Cheers.