build-override results in the same dependency being compiled twice

[link2xt]

Problem

I am compiling tag v1.121.0 of https://github.com/deltachat/deltachat-core-rust with command cargo build. It compiles aho-corasick v1.0.2 dependency twice, once as a dependency and once as a build dependency, resulting in two .rlib field being produced and actually slowing down the build process rather than speeding it up as intended by the build-override feature.

Steps

1. Clone https://github.com/deltachat/deltachat-core-rust, tag v1.121.0, commit a8551510cdb5be7a35ebd3574df79d15a4fa471a
2. Run cargo build -v.
3. Observe aho-corasick v1.0.2 dependency being compiled twice. Once as a proper dependency with -C panic=abort as specified in the manifest, and once as a build dependency without -C panic=abort.
4. Try to specify panic='abort' in the [profile.dev.build-override] section of Cargo.toml as a workaround, it results in an error "panic may not be specified in a build-override profile".

Possible Solution(s)

Do not apply build-override to dependencies that are actual dependencies of the project? Or allow to disable build-override completely.

Notes

I wrote a longer thread with my observations of the build times and duplicate builds of dependencies at https://users.rust-lang.org/t/same-dependency-has-four-different-debug-fingerprints/99580

Version

$ cargo version --verbose
cargo 1.72.0
release: 1.72.0
host: x86_64-unknown-linux-gnu
libgit2: 1.6.4 (sys:0.17.2 vendored)
libcurl: 8.2.1 (sys:0.4.63+curl-8.1.2 system ssl:OpenSSL/3.1.2)
os: Arch Linux [64-bit]

[weihanglo]

That's a consequence of this setting, not build-override. Guess you just had a typo 🤔.

By default, the panic strategy applies to build scripts and build dependencies is panic = "unwind". This cannot be overridden. Here are some prior discussions.

Compiler settings of build scripts are usually different from normal dependencies. This is kind of a tradeoff and somebody may run into the bad part. However, I don't really know why we cannot force change panic for build scripts. Let's nominate this to discuss.

[link2xt]

Setting panic = 'abort' is on purpose to reduce the binary size and is not a typo. unwind is not really useful for us as deltachat-core-rust runs either on mobile or in a GUI application and user will not see the stacktrace on panic anyway. We have got rid of all panicking code and have not seen panic for a long time, so supporting stacktraces is just a waste of space.

Prior discussions are about tests which probably makes sense because it is common to use assertions that panic in the tests instead of returning Result, but for build scripts I would be totally fine if they just abort on panic.

[weihanglo]

Sorry not being clear. I meant the title is a bit misleading. It's about the panic setting always sticking to unwind. If that is the case, could you update to something like "support panic for build-override profile"?

[link2xt]

Sorry not being clear. I meant the title is a bit misleading. It's about the panic setting always sticking to unwind. If that is the case, could you update to something like "support panic for build-override profile"?

Fourth step to reproduce (trying to set panic to abort) is more like a workaround, I was trying to unify the profiles manually. But ideally this should not be necessary, cargo could detect automatically that dependency is used as a normal dependency and build dependency and use normal dependency built will all optimizations without the build override. So the issue is not just about the ability to build a build dependency with panic='abort', but cargo not applying the override at all when it does not reduce the build time.

[link2xt]

I have created a separate proposal #12647 to track panic validation for build-override.