Re: [hatari-devel] Hatari profiler updates and CPU cycle questions

[ Thread Index | Date Index | More Archives ]

v4.4.5 i.e. it probably does less aggressive optimizations than
your newer GCC version.  I think GCC 4.5 started using extra aliasing
optimizations at higher optimization levels.

strict aliasing optimisations are supposed to be restricted to -O2 upwards, although I have been through the GCC options handling code in the past and 'inconsistent' is a good word for it all :-)

Do you see any compiler warnings from code under src/debug/ directory
if you rebuild those objects?

I saw no warnings to stdout/stderr while compiling... however I don't know if CMAKE is shuffling them to a logfile somewhere. Do you know if I should look elsewhere for warning output?
On x86 encountering GCC bugs is are really rare/corner case
due to better testing.  On ARM they're not so rare...

Yes, I found similar experience. :-) 
CMake / Hatari defaults to -O3.
When you have extra time, you might try whether the bug goes away
with -O1:
- if yes, please try also with -O2,
- if not, try with -O0.

I had tried this already, and the bug surfaces with any level above -O0 (that is: -O/-O1, -O2, -O3)

That should give a fairly good pointer on whether and what what
kind of optimizations trigger the issue.

I had started subdividing the -fopts for -O but I didn't get very far before trying the patch (divided them in roughly half - and the bug still occurred). I can continue with that but GCC's optimization groups and overrides are very inconsistent and I would not be surprised to find it still happens even with all of the opts manually switched off. :-)

Mail converted by MHonArc 2.6.19+