Rust Bug Minimization Patterns

[fanf2]

http://blog.pnkfx.org/blog/2019/11/18/rust-bug-minimization-patterns/

[wyldfire]

I think this article probably deserves a passing mention to the classic tool for reducing compiler bug cases, creduce [1]. I know it's used for rust or at least it can be used for rust.

It uses delta debugging [2] to minimize the reproducer, and AFAIK doesn't know about Rust. But it could be taught these rust-specific transformations.

[1] https://embed.cs.utah.edu/creduce/

[2] https://en.wikipedia.org/wiki/Delta_debugging

[comex → wyldfire]

c-reduce has been a godsend for me when I’ve had to figure out the cause of C++ compiler crashes. Manual minimization may be semi-mindless, but for that very reason it’s really boring, and it takes forever. So much nicer to sic c-reduce at it, take a snack break, and most of the time come back to a nice minimal reproducer. (Though at least once it ended up reproducing a different compiler crash than the one in the original code… Oh well.)

I’d love if there was a Rust version. In addition to Rust-specific transformations, you’d probably want to teach it how to simplify an entire crate dependency tree, involving multiple rustc invocations, rather than just one file. C/C++ header files may be crude, but they have an advantage here: after running a source file through the preprocessor, you end up with a completely self-contained file that (probably) behaves the same way. Thus you only need to reduce that one file, and AFAIK that’s all c-reduce supports. Also, for anything that is declared in a header file but defined in a different source file, c-reduce doesn’t need to know or care about the definition; if the compiler crash occurred on a source file not containing the definition, then the definition can’t be responsible. Rust doesn’t have that kind of natural separation, but maybe aggressively emptying out function bodies would be a good enough substitute.

[wyldfire → comex]

> Though at least once it ended up reproducing a different compiler crash than the one in the original code

I use creduce a lot and this experience is not at all uncommon. The fix is to make your interestingness test very specific.

[shoo → wyldfire]

the title for one of the creduce papers starts with "test-case reduction", which i think is clearer terminology than "bug minimisation". "bug minimisation" sounds like it is focusing on reducing the number of defects or something, which this isn't (directly) about.

[ghusbands]

This is called "Test-case reduction". Bug minimisation would be about reducing the number/severity of bugs in code.

[a1369209993]

> But it not always easy to toggle such comments on and off

Try:

  /*-* /
  non_pub_locally_unused_item(----) { ---- }
  //-*/

remove the space on the first line to re-enable it.

[klyrs → a1369209993]

For A/B testing, I use this in C++; it looks like it works in rust (according to my syntax highlighter):

  //* //add/delete the first slash to toggle these blocks
  fooA();
  /*/
  fooB();
  /**/

[FartyMcFarter → klyrs]

I like this for C/C++:

  #if 1 // or 0
  A
  #else
  B
  #endif

This can even be nested.

[ncmncm → FartyMcFarter]

This is so overwhelmingly better in every conceivable way that there is no excuse for ever using the other thing. (Unless you are using Rust and don't have it?)

[klyrs → FartyMcFarter]

Yeah, that's good too. When I don't need nesting, I like toggleable comments because the inactive block gets grayed out.

[FartyMcFarter → klyrs]

Some editors like vim also grey out the text with "#if".

[Digg_mov]

Good job