RE: [avr-libc-dev] Improved optimizer for mixed 32 / 16 / 8 bitexpressions

[Bernard Fouché]

>I, thus, whish to test my compiler with "real" code in order to avoid
>that I finally tune the instruction patterns for
>artificial test-cases that might not be helpful in practice.

Hi Björn.

What is required to test your patch? head gcc 4.0? How to get it? I can
evaluate it here, on a application where code size is more important than
speed. The application often takes 8 or 16 bits values then perform 32 bits
operations on them. Also there is a bunch of bit expressions because it
packs groups of bits coming from 8 and 16 bits values in groups of uint32_t.
The resulting size is about 37k now (atmega64).

We are also developping a bootloader that will fit in 8k with difficulty,
this bootloader also performing 32 bits calculations from 8 and 16 bits
values. Anyway to get more compact code is welcome.

So I can apply your patch, recompile and test, and give you results for a
medium size application and a big bootloader.

However I must tell that following what I've seen in the current generated
code with gcc 3.4.2, the application and bootloader have been already
modified to reduce, for instance, type casting from 8 to 32 bits or type
casting from 16 to 32 bits.

 Bernard