Thanks Esteban. This is interesting to tune in on whats happening in vm and FFI. cheers -ben
On Mon, Jan 23, 2017 at 4:00 PM, <esteba...@gmail.com> wrote: > Hello! > > This is my weekly ChangeLog, from 16 January 2017 to 22 January 2017. > You can see it in a better format by going here: http://log.smallworks.eu > > ChangeLog > ========= > > 20 January 2017: > ---------------- > > * I'm still fixing a couple of final failing tests on iceberg, > multi-remotes branch. Problem is that there are still somethings to > understand > on the "command line > backend" (we need to find a better name for this), and the #pushTo: > method: if I > do it as I > think is correct, a couple of failing tests appear... but then, I need > to > reflect if this is > a problem on the functionality or on the tests. > > 19 January 2017: > ---------------- > > * I think I finally got the problem with FT2Plugin! > Yesterday night talking in slack I realised that there was actually > duplicated > instances of > FT2Handle hierarchy with exactly same handle (pointer address). This, > of > course, causes a double free when those instances are released... and > that's the VM crash :) > So I quickly sketched a workaround, that you can see here: > FT2Handle>>pvtDestroyHandle > "This should only be sent from the finalizer. > It runs inside a mutex because in strange cases it can happen that > this is > executed twice, > causing a primitiveFailed to be raised." > self class destroyMutex critical: [ | handleToRelease | > self isValid ifFalse: [ ^self ]. > handleToRelease := self handle copy. > self primDestroyHandle. > "This is super-ugly, but it will prevent duplicated handles to be > released" > FT2Handle allSubInstancesDo: [ :each | > (handleToRelease = each handle) ifTrue: [ each beNull ] ] ] > and so far nobody who is testing it crashed! > This can be important, but of course... now we need to realise why the > duplicated instances are happening. > * I spent some time preparing a new way to annoy people :) > I prepared a way to send a digest mail about my activity (this stream) > to > pharo-dev list. > I hope it will not be too much. > > 18 January 2017: > ---------------- > > * ... and still some more work on iceberg, this time making some test > pass on the command line backend (who is just a fallback, but well... > we still > need it around). > * Still working on UFFI 64bits. Now I introduced a new hierarchy, > FFIArchitecture > for now with two children: > * FFI_i386 > * FFI_x86_64 > the purpose is obvious: to allow UFFI to handle differences between > different > architectures (for example, > long has to be parsed to FFIInt32 in i386 and to FFIInt64 in x86_64). > > 17 January 2017: > ---------------- > > * I'm working on the UFFI support for 64bits. > In general, it works fine, but there is a huge problem with > structures. What's > this problem? Well, let's take > this simple structure as example: > struct point { > unsigned long x; > unsigned long y; > } > Well, thing is in 32bits sizeof(struct point) = 8 while in 64bits > sizeof(struct > point) = 16, and of > course this breaks our current implementation which would be something > like this > StructPoint>>x > ^ handle unsignedLongAt: 1 > StructPoint>>y > ^ handle unsignedLongAt: 5 > and of course, this is not correct for 64bits. > My solution is quite simple, instead hardcoding the field offsets, I > store those > values into class variables > that can be calculated each session (but just once). > So, this structure would end looking like this: > StructPoint>>x > ^ handle unsignedLongAt: OFFSET_X > StructPoint>>y > ^ handle unsignedLongAt: OFFSET_Y > this would solve the problem for most of the cases, but there is still > the issue > that emerges when complete > implementation of the struct changes between platforms. I'm planning to > implement a strategy pattern to solve > this, but not yet (probably I will wait until having a concrete case). > This seems to be working, but now I have other problems not related to > UFFI: > many people has implemented > void * parameters of funtions as unsigned long (including me, time > ago) which > was correct in 32bits but > it is not in 64bits. So now I need to review this implementations to > see where a > fix is needed. > But this is not directly a problem of UFFI, I think. > * I worked with Nico on fixing libgit2 backend tests for iceberg. > Now (more thanks to Nico than me, heh), test for multi-remotes branch > are > passing for linux but > still not for macOS (and forget it about Pharo 5.0). > I suspect a problem with VM, I need to review if I can configure > smalltalkci > to work with latest vm instead stable. > > 15 January 2017: > ---------------- > > * ... and now I restored the posix permissions to windows platform. Next > latest VM > will have this change, I > hope that will end the problems with FilePlugin, but we'll see. > * I just fixed a problem with FilePlugin on macOS: the problem was when > determining if a symlink is also a > directory: > ‘/tmp’ asFileReference isDirectory > was answering false, because it is a symlink and when trying to > resolve the > symlink to obtain the > attributes, there was a cyclic condition that was transforming ‘/tmp’ > into > ‘/private/tmp’ and then > back to ‘/tmp’, and because of that the test for directory was failing. > > cheers! > Esteban > >
