On 03-Feb-2014 20:01, Krzysztof Kosiński wrote:
In the long term, I think the consensus is that we should move away from garbage collection and towards explicit memory management with smart pointers. In the short term, I'm afraid there is little we can do about this. The initial pieces of the puzzle are already in place with the C++ification work on the SP tree.
Apparently the current GC does not call destructors when it automatically removes memory. Inkscape has code in gc-finalized to tack a destructor onto an entry, but it isn't easy to see how often it is employed (grep shows "Finalized" in only a handful of source of files), and gc-finalized.h lists several caveats, which say that there are lots of situations where it might never act even if invoked.
GC is mainly used in code related to XML and undo handling. Removing GC there requires modifying every piece of code that creates or deletes XML nodes - at least 277 places according to grep.
Speaking of that section, is there some debugger trick for keeping track of the objects that that code creates and deletes? I've been trying to follow it in gdb but since the memory pointers move around from run to run, and just running gdb seems to change what inkscape does (probably because mouse events are caught in strange ways, or windows overlap), it is really tedious.
I ask because my text decoration mods blow up in style.cpp in sp_style_paint_server_ref_modified() when a certain simple object is selected on the drawing surface and then copied - but not if it is cloned. That object is just this:
<text style="font-size:40px;fill:none;stroke:url(#MyPattern);stroke-width:5;" y="40"><tspan>A<tspan style="text-decoration:underline">B</tspan></tspan></text>
The copy (==^C) triggers an insane amount of activity which passes in and out of style.cpp, where signals are attached and detached, apparently to/from the same objects much of the time, until eventually it drops into the function mentioned above with a PaintServer server that no longer matches the ones in style->stroke and style->text_decoration_stroke (which match each other), which triggers the crash. (After having previously been in that same function maybe a dozen times without issue.) It is my impression that there is vastly more activity for this simple copy then there is in just reading in that section of the file in the first place. (It definitely calls sp_style_write_string() at least 2 times for each of the styles in the section being copied, probably because execution control is through the setting and resetting of modified flags, rather than by doing a single "atomic" 1:1 copy.) Conversely, this object:
<text style="font-size:40px;fill:url(#MyPattern);stroke:none" y="40"><tspan>A<tspan style="text-decoration:underline">B</tspan></tspan></text>
copies without issue. It too triggers a lot of activity, including redundant sp_style_write_string() calls, but somehow it avoids whatever pitfall the first one falls into. As far as I can tell the implementations for style->text_decoration_fill and style->text_decoration_stroke are symmetrical, so why one crashes, and the other doesn't is pretty mysterious at present. It has something to do with the paintServer href, since there are no problems with text_decoration_stroke when it does not use a pattern or a gradient.
Tracing execution in this region is no fun since much of it is driven by signals that trigger calls to stored references to functions. I can watch it creating patterns, objects, and styles, but figuring out which is which solely from their pointers, which change from run to run, is not a lot of fun.
Thanks,
David Mathog mathog@...1176... Manager, Sequence Analysis Facility, Biology Division, Caltech