<MrGwen> bonzini: I've removed the ObjectTable
<bonzini> ??!?!?!? :)
* zecke (~ich@91-64-83-241-dynip.superkabel.de) ha entrat a #gnu-smalltalk
<MrGwen> yep :D
<bonzini> how is performance?
<bonzini> you're batshit crazy :)
<zecke> hi
<MrGwen> :)
<bonzini> examples/Bench.st will do :)
<MrGwen> :)
<MrGwen> I've removed the compiler/bootstrapper and only generate images from a st bootstrap
<MrGwen> also multiple kernels in the VM
<bonzini> ahah
<bonzini> gimme those numbers for the OT come on!
<bonzini> :)
* bonzini only 13 to go
<bonzini> also, how do you handle objects that move during GC?
<bonzini> kind of like the incubator does with the OT?
<MrGwen> I've no GC right now
<bonzini> aaaah :) :)
<MrGwen> I've do to one but don't know what will be the design
<MrGwen> I've time
<bonzini> you can reuse 99% of the code in mine, i think
<MrGwen> yes but I'm not sure if I won't use a kind of // gc
<bonzini> the current GC can be made parallel
<bonzini> just replace the marking stack with a lock-free stack
<bonzini> the copying phase is not parallel, true, but it is very short
<bonzini> and update the REACHABLE bits atomically
<MrGwen> true
<MrGwen> i've broke also readline stuff and I plan to do it in the image too, by the way I generate small images (1 mo for 64 bits)
<bonzini> :)
<MrGwen> removing all the objects known (and the compiler, parser, ...) by the vm makes the vm easier to hack
<bonzini> yep, how do you bootstrap?
<bonzini> that'd be the first part i'd like to merge
<MrGwen> simple i load everything in a namespace and I've made a kind of objectdumper too  but I control if I don't escape the world
<bonzini> so how do you build your loading format from kernel/*
<MrGwen> I load files from kernel (updated a bit) and inject them in a namespace
<MrGwen> I follow the graph and save the objects
<bonzini> you distribute a binary then
<MrGwen> yep
<MrGwen> an image
<bonzini> and you have a C loader for the objectdumper
<bonzini> ?
<MrGwen> yes
<MrGwen> I load the image and fix the pointers
<MrGwen> I use your infrastructure
<bonzini> do you save the image from Smalltalk code (SystemTracer), or is it a different format?
<MrGwen> the format is a bit diff since I take care of object alignment (like your allocator)
<MrGwen> correct me If I am wrong
<bonzini> yes, i think i understand
<bonzini> it's ugly though to distribute a binary...
<bonzini> you cannot build from ./configure && make
<MrGwen> I think the model could be C world (parser/compiler) generate a simple image => from the image generate a full image
<MrGwen> the cworld would send no messages but only see objects as structures
<bonzini> or, move the parser/compiler to a separate user of libgst
<bonzini> and out of libgst
<bonzini> it would break the ABI though
<MrGwen> good idea
<MrGwen> Paolo I've break too much stuff :D
<bonzini> and move the REPL to a smalltalk script like gst-load etc.
<bonzini> now you have to fix it :
<bonzini> )
<MrGwen> Also the context allocation is refactored in a way allowing the vm to have multiple smalltalk stacks
* MrGwen has a lot of work
<MrGwen> for become there is the solution described in the "Third generational Smalltalk" paper they put a bit on the object header if it is set to one this is "proxy" I 've to investigate on it