Re: [hatari-devel] Data cache Issues?

[Nicolas Pomarède <npomarede@xxxxxxxxxxxx>]

Le 06/06/2015 22:07, Nicolas Pomarède a écrit :
> If the depacker doesn't flush the instr cache, you can be sure a crash
> can happen ; depending on the packers, some of them are moving the
> depacked code at the same place that the depacker code was (to reduce
> memory footprint), so chances of getting a cache mismatch are even higher.

Hi

just to be sure the problem was not related to the depacker, I extracted 
"Whirlpool" demo by manually depacking the Atomic 3.5 data.
The resulting program was still crashing, so it was another problem than 
clearing the cache after depacking.
Adding some checks showed that it was a data cache mismatch.

After quite some low level debugging and diving into the cache's inner 
work :) , I found it was related to a case where a write to memory 
didn't correctly invalidate an old entry in the data cache (on a 
'miss'), causing subsequent reads to return an old value instead of the 
most recent one, hence the crash.

Whirlpool is now running correctly, even when packed (but the fact that 
the depacker doesn't clear cache could still lead to crash sometimes, 
although it should be very rare)

Eero, could you test again those demos that didn't work and for which 
you added a note ?

Nicolas