antioxidant update: librsvg builds, and other things (core-updates)

[Maxime Devos]

Some updates:

• core-updates is now targeted instead of master (because the librsvg on core-updates has a less complicated build system and hence easier to support with antioxidant)
• librsvg now builds. It's about 3 times larger than the cargo-build-system librsvg (at least, the librsvg from master, didn't compare against core-updates yet because there was no substitute available), so perhaps some compilation flags need to be changed.
• According to ci.guix.gnu.org, antioxidation is at 90%, but some of the 'failures' are because of cancelled builds in dependencies, the real number should be higher. Maybe 'Restart all builds' would be correct that, though I don't know if that wouldn't restart too much builds (including builds that succeeded).
  
• Running tests is supported, #:tests? and #:parallel-tests? is respected, there is an option to skip some tests (using #:test-options).  As antioxidant doesn't do #:skip-build? #t, I expect more packages to be tested with antioxidant than with cargo-build-system.
  
  There are some test failure, things like "tries accessing the network" are simply skipped. Sometimes it is assuming incorrect things about the build environment such assuming that stderr is a terminal, those are simply skipped. Sometimes test files are removed from the tarballs uploaded at crates.io, this appears to be intentional, there is a project to remove all such 'bloat', disregarding the value of tests -- for those packages we will have to switch to git checkouts, but for now I simply disabled tests there.
  
  I've also found some time bombs (e.g. rust-rustls appears to have a certificate expiration problem), there tests are disabled too.
  
  There were also some genuine test failures, some of them have been reported upstream but not all of them yet.
• antioxidant now has some defenses against bundling -- it detects if the feature "bundle", "vendor" or "vendored" is implicitly enabled and if so, bails out unless they were explicitly added to the package definition.
• I don't know if I mentioned it previously, but antioxidant now supports some 'unstable' rust code even when using 'stable' Rust compilers, by setting RUSTC_BOOTSTRAP=1. It's 'not supposed to be used', but if it's good enough for the Rust compiler, then it should be good enough for other software too, I'd think.
  

Some questions:

• Some Rust crates have 'examples' and 'benchmarks' that can be compiled and installed. I could teach antioxidant to compile and install them, though it appears to provide very little value at cost of more compile time and a greater closure size (more binaries, more dependencies and wasn't there some bug with grafts whose fix causes outputs to be substituted that ended up unused? Though hopefully someone could figure out a better fix ...)
• Due to how regularised the Rust build system is, it's feasible to compile tests even when cross-compiling (*), so cross-compiled could run the cross-compiled tests on the system they are cross-compiling for after the cross-compilation to verify their cross-compiled software.
  
  Currently, tests are only compiled and installed when #:tests? #true which is only the case when compiling natively(*), but with some work I could separate building tests from running tests and build tests by default when cross-compiling. Downside: tests have to be installed in an output, which increases the closure size.
  
  (Currently tests are installed in an output, but I could change that to the working directory).
  

(*) antioxidant doesn't support cross-compilation yet, but in the past I've tried out cross-compiling the standard library and it seemed to work, though it wasn't completed.

OpenPGP_0x49E3EE22191725EE.asc
Description: OpenPGP public key

OpenPGP_signature
Description: OpenPGP digital signature