Re: [avr-libc-dev] Quick test of Björn Haase's relax patch

[Björn Haase]

Dmitry K. wrote on Sonntag, 5. März 2006 06:22 :
> On Saturday 04 March 2006 20:23, you wrote:
> > Björn and Dmitry,
> > have you tried "program-at-once" compilation (avr-gcc -combine
> > -fwhole-program *.c ..)? This may save some more bytes with recent gccs.
>
> In result:
>
> Options:                 rlx     cmb  cmb+rlx
> ---------------------------------------------
> flash, bytes   11412   11056   11412   11104
> call             198      24     198      46
> rcall            143     316     143     294
> jmp               34      10      34      14
> rjmp             442     467     442     463
>
The issue probably is that "cmb" reorders the sequence of the functions in 
memory in a less favorable fashion. The functions itself seem to be the same. 
I'd like to suggest to use -ffunction-sections and -Wl,--gc-sections. Maybe 
this helps improving the ordering of the functions.
> And another question about replacing '[r]call/ret' to
> '[r]jmp/ret'. Whether this optimization is safe?
>    For example, avr-libc: fplib.  Split function pops
> two bytes from stack and returns into more high level
> in error case. Similar tricks make such optimization
> to be erroneous. 
I agree that in these cases one could run into trouble. However, as long as 
gcc generated assembly is used, I don't see where problems could arise.
> On the other hand, the GCC x86 port 
> for a long time is able to optimize such tails. It also
> does it more effectively as deletes not necessary
> more 'ret'.
I know that there are ways to implement this in gcc itself. IIRC the method is 
rather complicated. Since the inclusion of this optimization in the linker 
was not very difficult, I had simply added it. In the long run, of course, it 
would be better to make gcc do it itself.
>    Can be add an additional option for obvious inclusion
> of this optimization?
Yes, of course. The question is wether to activate or deactivate the 
optimization by default. However, in order to add such a switch, I would have 
to learn beforehand how to add switches to ld. Probably one would have to 
write a avr emulation template for ld.? What I am planning also is to add  
"use-16k-wrap-around" and "use-32k-wrap-around" flags. Adding this 
optimization is already prepared in the patch but currently not activated.

Bjoern.