17:31 <agentzh> fche: thanks for the tip on debug_print()

17:33 <agentzh> one sad thing is that the expression nodes do not have a pointer to point to their parent nodes.

17:33 <agentzh> it makes debugging harder.

17:34 <agentzh> i already tried dumping the current expression nodes but the output do not make much sense, seems to be some placeholder nodes. so i need to traverse to its ancestor nodes to get a bigger picture.

17:34 <fche> backtrace

17:34 <fche> and print from a higher level visitor

17:34 <fche> frame

17:34 <agentzh> okay, i'll try it. thanks.

17:35 <agentzh> the visitor mode does not use recursive descent, so i guess i'll only get the root node instead of near ancestors?

17:36 <fche> hm, recursive descent is almost all we do though

17:36 <fche> double-indirect dispatch makes the backtrace a little depeer

17:36 <fche> but it should all be there

17:36 <agentzh> ah, my mistake then :)

17:38 <agentzh> fche: btw, can we allow the various stack/uwnind tapset functions take a variant that accepts user-specified IP and SP reg values?

17:38 <agentzh> the reason i'm asking is that some times we cannot use the current IP and SP regs, for example, when programmatically skipping frames for JIT compiled machine code.

17:40 <agentzh> i hacked in a sprint_ubacktrace2:string(pc:long,sp:long) tapset function last night and it seems to work well to recover the full C backtrace from machine code frames generated by LuaJIT.

17:40 <agentzh> it's not a very good tapset name though. wondering if you have any better suggestions here.

17:41 <agentzh> we can do the same trick with other JIT'd code like PCRE JIT and etc.

17:41 <fche> there was a plan to support multi-language tracebacks by splitting up the backtrace functions into per-frame/step ones

17:42 <fche> and the tapsets would provide different-possible-langage implementations for the step ones

17:42 <fche> and use the 'function unwind_one() { if ($not_my_language) next; } type runtime overloading facility to let it try different languages

17:42 <fche> we haven't developed the idea farther though

17:43 <agentzh> it reminds me of dwarf's personality handler design.

17:44 <agentzh> i think it's for solving different problems but i agree the iterative unwinding API could be tweaked to accept custom IP/SP values.

17:45 <agentzh> i could already unwind LuaJIT/Python/Perl-land backtraces with the current stap impl without any assistent from the target side (unlike stap's current python addon, for example).

17:45 <agentzh> *i can

17:45 <agentzh> my current hurdle is to recover C-land backtraces below the JIT mcode frame.

17:46 <agentzh> hence the sprint_ubacktrace2() hack mentioned above.