Guile: fix {set,long}jmp-related crashes on Windows

Patch Guile until upstream fix is available. For details, see the new patch file 0010-Fix-setjmp-longjmp-related-crashes-on-Windows.patch

Fixes #6430 (closed) Fixes #6737 (closed)

added Windows label

assigned to @michaelkaeppler

added Patchnew label

Upstream bug report: https://debbugs.gnu.org/cgi/bugreport.cgi?bug=73167

LGTM, thanks for working on this!

added Patchreview label and removed Patchnew label

I'm not sure this is the correct fix since no unwinding means that we might leak objects on the stack. You write that JIT-generated code violates the Windows x64 calling conventions - is there a way to fix this instead?

But on Linux longjmp does no stack unwinding either, no?

https://blog.reverberate.org/2013/05/deep-wizardry-stack-unwinding.html

What concerns your question: I don't know, I would have to work through it. In any case it would be a way more invasive fix, I assume.

This is just a keyword-triggered response, I have not taken a look at the actual code. In Lisp/Scheme implementations, setjmp/longjmp may actually be used as a hook into stack winding, namely for deconstructing a stack and setting it back up in the context of continuations. This kind of abuse of those primitives in order to route around having to code assembly for stack-related operations does not really have an equivalent in C++. I don't think that LilyPond makes significant use of continuations (if at all) particularly for winding purposes (taking up a stored continuation not just going upstack) but I don't know how Guile itself behaves in that respect.

Leaving at Review for open discussions.

approved this merge request

Let's merge this (before next weekend, no hurry) if it fixes at least one investigated problem on Windows

added Patchcountdown label and removed Patchreview label

added Patchpush label and removed Patchcountdown label

resolved all threads

added 27 commits

• ad9b8d60...ca29747e - 26 commits from branch master
• b4f51394 - Guile: fix {set,long}jmp-related crashes on Windows

Compare with previous version

Equal diff for commits ad9b8d60 and b4f51394, not resetting patch status.

enabled an automatic merge when all merge checks for b4f51394 pass

merged

removed Patchpush label

I didn't test this before merging, but this requires the installation of automake because it touches Makefile.am and therefore doesn't work in the minimal installation given by Ansible. I'm working around this for 2.25.20, but do we need that modification in the patch? I guess it's mostly for make dist of Guile itself, right?

changed milestone to %2.25.20

mentioned in merge request !2567 (merged)