[Vm-dev] Array new: SmallInteger maxVal
John M McIntosh
johnmci at smalltalkconsulting.com
Fri Oct 9 20:20:17 UTC 2009
On 2009-10-09, at 7:34 AM, David T. Lewis wrote:
> On Tue, Oct 06, 2009 at 10:05:02PM +0200, Nicolas Cellier wrote:
>>> From http://code.google.com/p/pharo/issues/detail?id=1282
>> Is this known?
> A performance / GC related question:
> I found one part of the problem, but it is not a complete solution
> (there must be more bugs).
> Assume 32 bit object memory and assume that we are rigorously using
> usqint rather than sqint for all OOP address calculations.
> Then in Intepreter>>sufficientSpaceToInstantiate:indexableSize: if
> we are
> trying to allocate an Array of size 16r3FFFFFFF (i.e. SmallInteger
> then we get an arithmetic overflow when we add 2500 at the end of
> the method.
> Adding the 2500 is explained in the method comment:
> "Details: For speed, over-estimate space needed for fixed fields
> or literals; the low space threshold is a blurry line."
I think that is a wild guess, in my study of the low space threshold
issues you can
enter a state where this doesn't *fix* anything and endlessly cycle
and CG activity. The solution was the bias towards growth code.
However removing the 2500 might lead to side-effects.
So much much is too big? I'd think asking for anything with the range of
0xFFFFFFFFF - 10MB to 0xFFFFFFFF as unsigned numbers could be viewed
as too big
given it's reasonable to consider a image would be on the order of
10MB so you don't preclude
the ability to actually allocate somewhat close to 4GB of space.
Now obviously most operating system won't allow that, and some can
grind away for a long
time before deciding you are stupid as they attempt to give you the
Still I believe the bug is not in the primitive, it does I believe
signal primitive failure?
Lastly is this value signed? since I've seen things like Array new:
where deep down someone in foobarfigureoutoptionsize does ^-1 to
indicate a fatal error.
> This addition was not present in the original Dan Ingalls version of
> the method, so the overflow bug is a side effect of the optimization.
> The fix would be to either a) add a check to fail the primitive if
> size is too great, or b) remove the "add 2500" speed optimization.
> Can anyone suggest which solution would be the least harmful for
> performance? I suspect that adding the size check would be best, but I
> have no data.
John M. McIntosh <johnmci at smalltalkconsulting.com> Twitter:
Corporate Smalltalk Consulting Ltd. http://www.smalltalkconsulting.com
More information about the Vm-dev