On 1/2/23 22:30, Kevin Kofler via devel wrote:
Fabio Valentini wrote:
> - incompatible compile-time options (i.e. resulting in conditional
> compilation): different packages depend on crates with different sets
> of features enabled, sometimes with conflicting options. Even with a
> stable ABI, you'd need to build crates for all necessary combinations
> of configurations, and that matrix quickly explodes (i.e. usually
> exponentially - 2^n builds for for n independent flags). This is a
> deal-breaker for shared libraries in most cases, and also can't be
> solved by using a different compiler. (Unless you want to figure out
> *which* combinations to build, and *only* build these.)
Let me try formulating my criticism more constructively (since my previous
reply failed both at being polite and at getting my point through, sorry
again for that):
I am really surprised to read above that Rust apparently allows applications
to pick the flag with which the libraries they depend on are compiled. I
really have to wonder why anyone would think that allowing that would be a
good idea, but then again I guess I know the answer: Whoever added this
feature was so set in a mindset where everything is compiled on demand and
statically linked that they figured: why not?
One of the major uses is to allow code that requires a particular
dependency to be disabled when that dependency is not available.
In particular, Rust targets platforms (such as OS kernels and embedded
systems) where the standard library is not available. This would be
extremely difficult without Cargo features.
And if you are in that mindset, that actually sounds like a
reasonable call
to make. Source-based software distributions do have the advantage of
offering this kind of flexibility on demand, see also the USE flags in
Gentoo. Those are in fact one of the main reasons some people decide to
compile an entire GNU/Linux distribution from source (and hence pick a
distribution such as Gentoo) to begin with. Likewise, the Rust way of
compiling dependencies on demand allows applications to make this kind of
settings for them.
Still, I can see several issues with that approach, e.g., what if an
application depends on two libraries A and B that both depend on library C,
but with conflicting flags?
Last I checked, Cargo features are additive, so the answer is that
C will be compiled with the union of all flags used by A and B.
But the main issue is that, as you point out, it
makes binary distribution of shared libraries highly impractical. That is
why I think this was a short-sighted design decision.
Cargo features are supposed to be additive, so one can sometimes ship
a single package with the union of all features used by its reverse
dependencies. This must be handled on a case-by-case basis, though.
--
Sincerely,
Demi Marie Obenour (she/her/hers)