[Vm-dev] Integer overflow with BitBlt rule 20 and depth 32
Juan Vuletich
juan at jvuletich.org
Sat Oct 24 13:40:17 UTC 2009
Hi Folks,
David T. Lewis wrote:
> It should be OK to add the type declarations only where they are actually
> needed for comparison and multiply.
>
> Just FYI it is worth noting that the default sqInt is an 64 bit long
> when generating code for a 64 bit image VM, so it is generally good
> practice to explicitly declare the int and unsigned types when doing
> 32 bit arithmetic. Unfortunately this looks like it would be a huge
> amount of tedious work for the bitblt plugin, so I would not worry
> about it for now (but it would be good to make some unit tests for
> these problems so it will be possible to validate the fixes on a 64
> bit image later).
>
> I'll try building a VM this weekend if nobody else has gotten to it by
> then. Note, I have zero expertise with bitblt (but I may still be able
> to help with the debugging).
>
> Dave
>
Thanks for the offer Dave. I found a little more time last evening and
this morning. I built a new VM on Windows with the patch I sent
yesterday. The behavior is now correct, both on my scripts and on
Henrik's. Using Henrik's script to measure performance, I see a 2% loss
in rgbAdd, most likely because of the comparison I added. We could
unroll the loop and do the new comparison only for the most significant
partition (i.e. alpha). That would enhance the performance a little bit.
I will do a couple of tests to check that this correct behavior is
maintained, including in 64bit images and VMs.
What worries me a bit is the other changes I needed to do to be able to
run the Smalltalk BitBlt simulation and to do the translation. These are:
BitBltSimulator >> #oopForPointer: "May be harmless"
CArrayAccessor >> #long32At: "Why is this needed?"
CArrayAccessor >> #long32At:put: "Why is this needed?"
CCodeGenerator >> #emitCConstantsOn: "Consequence of recent changes to
Dictionary >> #keys. Most likely harmless. May be other senders aroud!
(yes, in the very same method there is another sender that would be
optimized by #asSet!"
I've not been following the development of VMMaker closely enough to
advise on them, so please everybody, check and comment on them.
Cheers,
Juan Vuletich
More information about the Vm-dev
mailing list