commit 7571ece6bb1447d2ae8963e9cc4c7c5025440b41
author Devin Jeanpierre <jeanpierreda@google.com> Fri Jan 13 11:39:26 2023 -0800
committer Copybara-Service <copybara-worker@google.com> Fri Jan 13 11:40:02 2023 -0800
tree 97415940db443754473a1dc1abf739a993946a90
parent 60386e948dbeeeb7e39362e2756a18ddd2a686ba

If a header is in multiple targets, take it as part of the first target (lexicographically).

This is not Good Behavior. In particular, this means that adding a dependency can break Rust callers. Consider the following:

```cc
//h1.h
struct S {};
```

```cc
//h2.h
void Foo(S s);
```

with this BUILD file:

```py
cc_library(name="h1_a", hdrs=["h1.h"], ...)
cc_library(name="h1_b", hdrs=["h1.h"], ...)

cc_library(name="h2_a", hdrs=["h2.h"], deps=[":h1_a"])
cc_library(name="h2_b", hdrs=["h2.h"], deps=[":h1_b"])
cc_library(name="h2_ab", hdrs=["h2.h"], deps=[":h1_a", ":h1_b"])
```

Then:

* `h2_a` contains `fn Foo(s: h1_a::S);`
* `h2_b` contains `fn Foo(s: h2_a::S);`
* `h2_ab` contains `fn Foo(s: h1_a::S);`

Meaning that adding a single (TRANSITIVE) dependency can accidentally change the type of your parameters, breaking Rust callers -- a caller who was passing `h2_a::S` would get broken if a transitive dependency on `h1_a` were added.

But note: without this CL, we're even _worse_ off, because that same transitive dependency breaks the build entirely, and there is no way to resolve it. After this CL, there is a quick fix: use `h1_a::S`.

---

Alternatives considered:

* We could fix all headers everywhere to have a single canonical build target, but this is a big task. We do likely want to do it eventually, but it shouldn't block early adopters from trying Crubit.

---

It's worth noting that with the proposed[]

PiperOrigin-RevId: 501900537