Re: [avr-gcc-list] patch posted

[Paulo Marques]

Hi, Andrew,

Andrew Hutchinson wrote:
> I posted patch to cure spill bugs. It passed all testing.

I tested the patch too, and it seems to fix a few failures. Nice work.

before your patch (complete gcc.c-torture):

# of expected passes            21244
# of unexpected failures        162
# of unresolved testcases       52
# of unsupported tests          85
/home/pmarques/Desktop/gcc-4.2.2/svn/obj/gcc/xgcc  version 4.3.0 
20080129 (experimental) (GCC)

after your patch (complete gcc.c-torture):

# of expected passes            21247
# of unexpected failures        152
# of unresolved testcases       53
# of unsupported tests          86
/home/pmarques/Desktop/gcc-4.2.2/svn/obj/gcc/xgcc  version 4.3.0 
20080129 (experimental) (GCC)

I had to apply the patch manually, though, as I couldn't get the patch 
out of the mailing list archive without corruption :(

That extra unresolved and extra unsupported test case is probably a 
".text segment full" error. If this is the case, then probably the 
generated code is slightly larger for at least one test case. I still 
have to dig some more to find out exactly what happened.

> I would be interested in seeing how this affects size,speed of avr-gcc.

Me too. I was wondering what would be the best way to use the testsuite 
and avrtest to get a quantitative estimate of code size / execution time 
variations.

I thought that the first rough estimate would be to make avrtest 
accumulate all the bytes loaded and cycles executed during a testsuite 
run. Then I could apply a patch and redo the tests and see how that 
affected the values.

But this doesn't really work, because:

 - if a test fails in one run and not on the other, it can exit early 
and execute in less cycles. Not using failed tests for the statistics 
has similar problems.

 - we probably want code size with -Os and execution speed with -O2 / 
-O3 and not both mixed.

I know that the testsuite probably isn't the best speed / code size / 
RAM used benchmark available, but since I'll be running it on a regular 
basis and I'll be using it to bisect regressions, I could get these 
results basically for free...

--
Paulo Marques
Software Development Department - Grupo PIE, S.A.
Phone: +351 252 290600, Fax: +351 252 290601
Web: www.grupopie.com

"There cannot be a crisis today; my schedule is already full."