Re: [avr-libc-dev] Testsuite results for gcc-3.4.3 seem to exhibit unknown compiler bugs.

[Björn Haase]

Am Dienstag, 4. Januar 2005 00:28 schrieb E. Weddington:
> Björn Haase wrote:
> >1.) one testreport exhibiting IMHO at least 2 bugs in the present compiler
> >release (don't  panic nothing severe).
> >
> >P.P.S.:
> >I have cross-posted the contents of 1.) also to
> > address@hidden
>
> I took a look at your post in the archives of gcc-testresults (I'm not
> currently subscribed).
>
> Will you also be submitting GCC bug reports for those valid bugs that
> have turned up from running the testsuite?
>
> Thanks
> Eric


Hi Eric,

so far I've not submitted bug reports for the errors. I've been working on 
removing the spurious errors that show up due to tests in the suite that do 
not fit for our tiny targets. There is a rather big mail on the way that 
discusses the different issues a bit more complete. It's on "administrator 
approval" state since the message body is rather large. There you will also 
find a new   run-avr  executable that now should work also for the cygwin 
environment.

I'd suggest to first file bug reports for the 9 files where the debugging 
makes problems and then the problem with the negative shift counts. In fact, 
I didn't even know that something like a >> -13 is supposed to be a legal C 
expression!

Fixing these bugs will already remove the majority of the remaining errors.

There's another hand full of errors that has to do with built-in library 
functions gcc assumes to find that are in fact not available.  And finally, 
we have a number of more complicated issues where compilation fails for 
selected optimizer options. Maybe that these are are related to mid-end code.

BTW when looking at the test suite results for 4.0 there is a huge number of 
regressions. The origin, however, seems to be one and the same. It seems to 
have to do with long long variables that reload does not succeed in placing 
in registers. For the same reason building the compiler is actually failing 
right now. It might be, that the problem is just, that gcc 4.0 assumes long 
long to be 128 bit and avr has, thus, problems with placing two long longs in 
registers at the same time.? Seems that Paul Schlie has prepared some patch, 
but so far, I have not had the time to look after it.

Yours,

Björn