Expose all OS-specific modules in libstd doc. #43348
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
(rust_highfive has picked a reviewer for you, use r? to override) |
alexcrichton
added
the
S-waiting-on-review
|
friendly ping for review @alexcrichton! also pinging you on irc! |
|
Ah sorry for the delay, but thanks for the PR @kennytm! This has defnitely long since been one of the major issues with the documentation of the standard library, and it's exciting to see movement towards fixing it! This is a pretty big PR though is sort of difficult to review, and it's unclear to me at least how easy it'll be to maintain this into the future. I wonder if we could pursue a slightly smaller implementation in the meantime which may have practically the same impact? It looks like a lot of the complication here comes from handling the The "biggies" to handle here I think are One other worry I'd have is that I think we'll want to do an audit of the documentation to ensure that everything is clearly documented as either Unix or Windows specific. Otherwise it may be confusing to see those windows methods in the Unix docs! Also cc @rust-lang/libs, others may be interested in this as well! |
kennytm
force-pushed
the
fix-24658-doc-every-platform
branch
from
0923acd
to
5e89b23
Compare
|
@alexcrichton Thanks for the comments :) I've reverted most of the (Additionally, the This PR only labels OsStrExt/OsStringExt with the "Windows only" and "Unix only" labels, since I believe these will confuse most people. There could be a special badge for these non-portable function, like the stability badge. Ideally this should be handled by rustdoc by checking the
|
|
Oh right yes, Reading over this again though I'm still pretty wary personally. This seems very meticulously crafted in the sense that it's going to be pretty difficult to maintain over time. For example I could see that perhaps impls are added which involve adding more and more cases to the I wonder if perhaps we pursue solutions in rustdoc other than the standard library? For example there's no real need for rustdoc to actually typecheck these function bodies. Could rustdoc unconditionally always replace all function bodies with I think we may want to also pursue a solution in rustdoc for the code examples here. For example we could have something like Overall I'm just wondering if we could reduce the maintenance burden on the standard library as we add more impls/examples here over time! |
|
@alexcrichton Sounds good to me! That could get rid of (cc @GuillaumeGomez @steveklabnik about the potential rustdoc changes) |
|
Yeah we'll definitely still need |
|
Wo, that's quite the improvement! Consider I'm all 👍 |
kennytm
changed the title
kennytm
force-pushed
the
fix-24658-doc-every-platform
branch
2 times, most recently
from
c74278b
to
768f6da
Compare
kennytm
changed the title
|
@alexcrichton Updated to include I did not feature-gate Current output. cc @GuillaumeGomez some CSS changed.
Fullpage screenshot of `Metadata` structure
|
kennytm
force-pushed
the
fix-24658-doc-every-platform
branch
from
768f6da
to
a59eb52
Compare
|
Seems good to me (for the CSS part). |
|
@alexcrichton - are you looking at reviewing this? Friendly ping to make sure this isn't getting lost. |
|
This is awesome! |
|
This also looks fantastic to me, thanks so much for taking on the refactoring @kennytm! Only two thoughts now at this point:
|
|
I went ahead and checked out this PR to render an example std doc bundle with it: https://tonberry.quietmisdreavus.net/std-43348/std/os/index.html (This is rendered on a Linux server, in case that's useful information.) |
4 tasks
kennytm
force-pushed
the
fix-24658-doc-every-platform
branch
from
a59eb52
to
7228c96
Compare
|
@alexcrichton Done :). Feature gate tracked in #43781. |
|
☔ The latest upstream changes (presumably #43730) made this pull request unmergeable. Please resolve the merge conflicts. |
This prevents compilation failure we want to document a platform-specific
module. Every function is replaced by `loop {}` using the same construct
as `--unpretty everybody_loops`.
Note also a workaround to rust-lang#43636 is included: `const fn` will retain their
bodies, since the standard library has quite a number of them.
This attribute has two effects: 1. Items with this attribute and their children will have the "This is supported on **** only" message attached in the documentation. 2. The items' doc tests will be skipped if the configuration does not match.
kennytm
force-pushed
the
fix-24658-doc-every-platform
branch
from
7228c96
to
b4114eb
Compare
This was referenced Jun 18, 2020
Manishearth
added a commit
to Manishearth/rust
that referenced
this pull request
Jul 16, 2020
Don't run `everybody_loops` for rustdoc; instead ignore resolution errors r? @eddyb cc @petrochenkov, @GuillaumeGomez, @Manishearth, @ecstatic-morse, @marmeladema ~~Blocked on rust-lang#73743 Merged. ~~Blocked on crater run.~~ Crater popped up some ICEs ([now fixed](rust-lang#73566 (comment))). See [crater run](https://crater-reports.s3.amazonaws.com/pr-73566/index.html), [ICEs](rust-lang#73566 (comment)). ~~Blocked on rust-lang#74070 so that we don't make typeck_tables_of public when it shouldn't be.~~ Merged. Closes rust-lang#71820, closes rust-lang#71104, closes rust-lang#65863. ## What is the motivation for this change? As seen from a lengthy trail of PRs and issues (rust-lang#73532, rust-lang#73103, rust-lang#71820, rust-lang#71104), `everybody_loops` is causing bugs in rustdoc. The main issue is that it does not preserve the validity of the `DefId` tree, meaning that operations on DefIds may unexpectedly fail when called later. This is blocking intra-doc links (see rust-lang#73101). This PR starts by removing `everybody_loops`, fixing rust-lang#71104 and rust-lang#71820. However, that brings back the bugs seen originally in rust-lang#43348: Since libstd documents items for all platforms, the function bodies sometimes do not type check. Here are the errors from documenting `libstd` with `everybody_loops` disabled and no other changes: ```rust error[E0433]: failed to resolve: could not find `handle` in `sys` --> src/libstd/sys/windows/ext/process.rs:13:27 | 13 | let handle = sys::handle::Handle::new(handle as *mut _); | ^^^^^^ could not find `handle` in `sys` error[E0425]: cannot find function `symlink_inner` in module `sys::fs` --> src/libstd/sys/windows/ext/fs.rs:544:14 | 544 | sys::fs::symlink_inner(src.as_ref(), dst.as_ref(), false) | ^^^^^^^^^^^^^ not found in `sys::fs` error[E0425]: cannot find function `symlink_inner` in module `sys::fs` --> src/libstd/sys/windows/ext/fs.rs:564:14 | 564 | sys::fs::symlink_inner(src.as_ref(), dst.as_ref(), true) | ^^^^^^^^^^^^^ not found in `sys::fs` ``` ## Why does this need changes to `rustc_resolve`? Normally, this could be avoided by simply not calling the `typeck_item_bodies` pass. However, the errors above happen before type checking, in name resolution itself. Since name resolution is intermingled with macro expansion, and rustdoc needs expansion to happen before it knows all items to be documented, there needs to be someway to ignore _resolution_ errors in function bodies. An alternative solution suggested by @petrochenkov was to not run `everybody_loops` on anything containing a nested `DefId`. This would solve some of the immediate issues, but isn't bullet-proof: the following functions still could not be documented if the items in the body failed to resolve: - Functions containing a nested `DefId` (rust-lang#71104) - ~~Functions returning `impl Trait` (rust-lang#43878 These ended up not resolving anyway with this PR. - ~~`const fn`, because `loop {}` in `const fn` is unstable (rust-lang#43636 `const_loop` was just stabilized. This also isn't exactly what rustdoc wants, which is to avoid looking at function bodies in the first place. ## What changes were made? The hack implemented in this PR is to add an option to ignore all resolution errors in function bodies. This is enabled only for rustdoc. Since resolution errors are ignored, the MIR generated will be invalid, as can be seen in the following ICE: ```rust error: internal compiler error: broken MIR in DefId(0:11 ~ doc_cfg[8787]::uses_target_feature[0]) ("return type"): bad type [type error] --> /home/joshua/src/rust/src/test/rustdoc/doc-cfg.rs:51:1 | 51 | / pub unsafe fn uses_target_feature() { 52 | | content::should::be::irrelevant(); 53 | | } | |_^ ``` Fortunately, rustdoc does not need to access MIR in order to generate documentation. Therefore this also removes the call to `analyze()` in `rustdoc::run_core`. This has the side effect of not generating all lints by default. Most lints are safe to ignore (does rustdoc really need to run liveness analysis?) but `missing_docs` in particular is disabled when it should not be. Re-running `missing_docs` specifically does not help, because it causes the typechecking pass to be run, bringing back the errors from rust-lang#24658: ``` error[E0599]: no method named `into_handle` found for struct `sys::unix::pipe::AnonPipe` in the current scope --> src/libstd/sys/windows/ext/process.rs:71:27 | 71 | self.into_inner().into_handle().into_raw() as *mut _ | ^^^^^^^^^^^ method not found in `sys::unix::pipe::AnonPipe` | ``` Because of rust-lang#73743, we only run typeck on demand. So this only causes an issue for functions returning `impl Trait`, which were already special cased by `ReplaceFunctionWithBody`. However, it now considers `async fn f() -> T` to be considered `impl Future<Output = T>`, where before it was considered to have a concrete `T` type. ## How will this affect future changes to rustdoc? - Any new changes to rustdoc will not be able to perform type checking without bringing back resolution errors in function bodies. + As a corollary, any new lints cannot require or perform type checking. In some cases this may require refactoring other parts of the compiler to perform type-checking only on-demand, see for example rust-lang#73743. + As a corollary, rustdoc can never again call `tcx.analysis()` unless this PR is reverted altogether. ## Current status - ~~I am not yet sure how to bring back `missing_docs` without running typeck. @eddyb suggested allowing lints to opt-out of type-checking, which would probably be another rabbit hole.~~ The opt-out was implemented in rust-lang#73743. However, of the rustc lints, now _only_ missing_docs is run and no other lints: rust-lang#73566 (comment). We need a team decision on whether that's an acceptable tradeoff. Note that all rustdoc lints are still run (`intra_doc_link_resolution_failure`, etc). **UPDATE**: This was deemed acceptable in rust-lang#73566 (comment) - ~~The implementation of optional errors in `rustc_resolve` is very brute force, it should probably be moved from `LateResolver` to `Resolver` to avoid duplicating the logic in many places.~~ I'm mostly happy with it now. - This no longer allows errors in `async fn f() -> T`. This caused breakage in 50 crates out of a full crater run, all of which (that I looked at) didn't compile when run with rustc directly. In other words, it used to be that they could not be compiled but could still be documented; now they can't be documented either. This needs a decision from the rustdoc team on whether this is acceptable breakage. **UPDATE**: This was deemed acceptable in rust-lang#73566 (comment) - ~~This makes `fn typeck_tables_of` in `rustc_typeck` public. This is not desired behavior, but needs the changes from rust-lang#74070 in order to be fixed.~~ Reverted.
Manishearth
added a commit
to Manishearth/rust
that referenced
this pull request
Jul 16, 2020
Don't run `everybody_loops` for rustdoc; instead ignore resolution errors r? @eddyb cc @petrochenkov, @GuillaumeGomez, @Manishearth, @ecstatic-morse, @marmeladema ~~Blocked on rust-lang#73743 Merged. ~~Blocked on crater run.~~ Crater popped up some ICEs ([now fixed](rust-lang#73566 (comment))). See [crater run](https://crater-reports.s3.amazonaws.com/pr-73566/index.html), [ICEs](rust-lang#73566 (comment)). ~~Blocked on rust-lang#74070 so that we don't make typeck_tables_of public when it shouldn't be.~~ Merged. Closes rust-lang#71820, closes rust-lang#71104, closes rust-lang#65863. ## What is the motivation for this change? As seen from a lengthy trail of PRs and issues (rust-lang#73532, rust-lang#73103, rust-lang#71820, rust-lang#71104), `everybody_loops` is causing bugs in rustdoc. The main issue is that it does not preserve the validity of the `DefId` tree, meaning that operations on DefIds may unexpectedly fail when called later. This is blocking intra-doc links (see rust-lang#73101). This PR starts by removing `everybody_loops`, fixing rust-lang#71104 and rust-lang#71820. However, that brings back the bugs seen originally in rust-lang#43348: Since libstd documents items for all platforms, the function bodies sometimes do not type check. Here are the errors from documenting `libstd` with `everybody_loops` disabled and no other changes: ```rust error[E0433]: failed to resolve: could not find `handle` in `sys` --> src/libstd/sys/windows/ext/process.rs:13:27 | 13 | let handle = sys::handle::Handle::new(handle as *mut _); | ^^^^^^ could not find `handle` in `sys` error[E0425]: cannot find function `symlink_inner` in module `sys::fs` --> src/libstd/sys/windows/ext/fs.rs:544:14 | 544 | sys::fs::symlink_inner(src.as_ref(), dst.as_ref(), false) | ^^^^^^^^^^^^^ not found in `sys::fs` error[E0425]: cannot find function `symlink_inner` in module `sys::fs` --> src/libstd/sys/windows/ext/fs.rs:564:14 | 564 | sys::fs::symlink_inner(src.as_ref(), dst.as_ref(), true) | ^^^^^^^^^^^^^ not found in `sys::fs` ``` ## Why does this need changes to `rustc_resolve`? Normally, this could be avoided by simply not calling the `typeck_item_bodies` pass. However, the errors above happen before type checking, in name resolution itself. Since name resolution is intermingled with macro expansion, and rustdoc needs expansion to happen before it knows all items to be documented, there needs to be someway to ignore _resolution_ errors in function bodies. An alternative solution suggested by @petrochenkov was to not run `everybody_loops` on anything containing a nested `DefId`. This would solve some of the immediate issues, but isn't bullet-proof: the following functions still could not be documented if the items in the body failed to resolve: - Functions containing a nested `DefId` (rust-lang#71104) - ~~Functions returning `impl Trait` (rust-lang#43878 These ended up not resolving anyway with this PR. - ~~`const fn`, because `loop {}` in `const fn` is unstable (rust-lang#43636 `const_loop` was just stabilized. This also isn't exactly what rustdoc wants, which is to avoid looking at function bodies in the first place. ## What changes were made? The hack implemented in this PR is to add an option to ignore all resolution errors in function bodies. This is enabled only for rustdoc. Since resolution errors are ignored, the MIR generated will be invalid, as can be seen in the following ICE: ```rust error: internal compiler error: broken MIR in DefId(0:11 ~ doc_cfg[8787]::uses_target_feature[0]) ("return type"): bad type [type error] --> /home/joshua/src/rust/src/test/rustdoc/doc-cfg.rs:51:1 | 51 | / pub unsafe fn uses_target_feature() { 52 | | content::should::be::irrelevant(); 53 | | } | |_^ ``` Fortunately, rustdoc does not need to access MIR in order to generate documentation. Therefore this also removes the call to `analyze()` in `rustdoc::run_core`. This has the side effect of not generating all lints by default. Most lints are safe to ignore (does rustdoc really need to run liveness analysis?) but `missing_docs` in particular is disabled when it should not be. Re-running `missing_docs` specifically does not help, because it causes the typechecking pass to be run, bringing back the errors from rust-lang#24658: ``` error[E0599]: no method named `into_handle` found for struct `sys::unix::pipe::AnonPipe` in the current scope --> src/libstd/sys/windows/ext/process.rs:71:27 | 71 | self.into_inner().into_handle().into_raw() as *mut _ | ^^^^^^^^^^^ method not found in `sys::unix::pipe::AnonPipe` | ``` Because of rust-lang#73743, we only run typeck on demand. So this only causes an issue for functions returning `impl Trait`, which were already special cased by `ReplaceFunctionWithBody`. However, it now considers `async fn f() -> T` to be considered `impl Future<Output = T>`, where before it was considered to have a concrete `T` type. ## How will this affect future changes to rustdoc? - Any new changes to rustdoc will not be able to perform type checking without bringing back resolution errors in function bodies. + As a corollary, any new lints cannot require or perform type checking. In some cases this may require refactoring other parts of the compiler to perform type-checking only on-demand, see for example rust-lang#73743. + As a corollary, rustdoc can never again call `tcx.analysis()` unless this PR is reverted altogether. ## Current status - ~~I am not yet sure how to bring back `missing_docs` without running typeck. @eddyb suggested allowing lints to opt-out of type-checking, which would probably be another rabbit hole.~~ The opt-out was implemented in rust-lang#73743. However, of the rustc lints, now _only_ missing_docs is run and no other lints: rust-lang#73566 (comment). We need a team decision on whether that's an acceptable tradeoff. Note that all rustdoc lints are still run (`intra_doc_link_resolution_failure`, etc). **UPDATE**: This was deemed acceptable in rust-lang#73566 (comment) - ~~The implementation of optional errors in `rustc_resolve` is very brute force, it should probably be moved from `LateResolver` to `Resolver` to avoid duplicating the logic in many places.~~ I'm mostly happy with it now. - This no longer allows errors in `async fn f() -> T`. This caused breakage in 50 crates out of a full crater run, all of which (that I looked at) didn't compile when run with rustc directly. In other words, it used to be that they could not be compiled but could still be documented; now they can't be documented either. This needs a decision from the rustdoc team on whether this is acceptable breakage. **UPDATE**: This was deemed acceptable in rust-lang#73566 (comment) - ~~This makes `fn typeck_tables_of` in `rustc_typeck` public. This is not desired behavior, but needs the changes from rust-lang#74070 in order to be fixed.~~ Reverted.
Manishearth
added a commit
to Manishearth/rust
that referenced
this pull request
Jul 16, 2020
Don't run `everybody_loops` for rustdoc; instead ignore resolution errors r? @eddyb cc @petrochenkov, @GuillaumeGomez, @Manishearth, @ecstatic-morse, @marmeladema ~~Blocked on rust-lang#73743 Merged. ~~Blocked on crater run.~~ Crater popped up some ICEs ([now fixed](rust-lang#73566 (comment))). See [crater run](https://crater-reports.s3.amazonaws.com/pr-73566/index.html), [ICEs](rust-lang#73566 (comment)). ~~Blocked on rust-lang#74070 so that we don't make typeck_tables_of public when it shouldn't be.~~ Merged. Closes rust-lang#71820, closes rust-lang#71104, closes rust-lang#65863. ## What is the motivation for this change? As seen from a lengthy trail of PRs and issues (rust-lang#73532, rust-lang#73103, rust-lang#71820, rust-lang#71104), `everybody_loops` is causing bugs in rustdoc. The main issue is that it does not preserve the validity of the `DefId` tree, meaning that operations on DefIds may unexpectedly fail when called later. This is blocking intra-doc links (see rust-lang#73101). This PR starts by removing `everybody_loops`, fixing rust-lang#71104 and rust-lang#71820. However, that brings back the bugs seen originally in rust-lang#43348: Since libstd documents items for all platforms, the function bodies sometimes do not type check. Here are the errors from documenting `libstd` with `everybody_loops` disabled and no other changes: ```rust error[E0433]: failed to resolve: could not find `handle` in `sys` --> src/libstd/sys/windows/ext/process.rs:13:27 | 13 | let handle = sys::handle::Handle::new(handle as *mut _); | ^^^^^^ could not find `handle` in `sys` error[E0425]: cannot find function `symlink_inner` in module `sys::fs` --> src/libstd/sys/windows/ext/fs.rs:544:14 | 544 | sys::fs::symlink_inner(src.as_ref(), dst.as_ref(), false) | ^^^^^^^^^^^^^ not found in `sys::fs` error[E0425]: cannot find function `symlink_inner` in module `sys::fs` --> src/libstd/sys/windows/ext/fs.rs:564:14 | 564 | sys::fs::symlink_inner(src.as_ref(), dst.as_ref(), true) | ^^^^^^^^^^^^^ not found in `sys::fs` ``` ## Why does this need changes to `rustc_resolve`? Normally, this could be avoided by simply not calling the `typeck_item_bodies` pass. However, the errors above happen before type checking, in name resolution itself. Since name resolution is intermingled with macro expansion, and rustdoc needs expansion to happen before it knows all items to be documented, there needs to be someway to ignore _resolution_ errors in function bodies. An alternative solution suggested by @petrochenkov was to not run `everybody_loops` on anything containing a nested `DefId`. This would solve some of the immediate issues, but isn't bullet-proof: the following functions still could not be documented if the items in the body failed to resolve: - Functions containing a nested `DefId` (rust-lang#71104) - ~~Functions returning `impl Trait` (rust-lang#43878 These ended up not resolving anyway with this PR. - ~~`const fn`, because `loop {}` in `const fn` is unstable (rust-lang#43636 `const_loop` was just stabilized. This also isn't exactly what rustdoc wants, which is to avoid looking at function bodies in the first place. ## What changes were made? The hack implemented in this PR is to add an option to ignore all resolution errors in function bodies. This is enabled only for rustdoc. Since resolution errors are ignored, the MIR generated will be invalid, as can be seen in the following ICE: ```rust error: internal compiler error: broken MIR in DefId(0:11 ~ doc_cfg[8787]::uses_target_feature[0]) ("return type"): bad type [type error] --> /home/joshua/src/rust/src/test/rustdoc/doc-cfg.rs:51:1 | 51 | / pub unsafe fn uses_target_feature() { 52 | | content::should::be::irrelevant(); 53 | | } | |_^ ``` Fortunately, rustdoc does not need to access MIR in order to generate documentation. Therefore this also removes the call to `analyze()` in `rustdoc::run_core`. This has the side effect of not generating all lints by default. Most lints are safe to ignore (does rustdoc really need to run liveness analysis?) but `missing_docs` in particular is disabled when it should not be. Re-running `missing_docs` specifically does not help, because it causes the typechecking pass to be run, bringing back the errors from rust-lang#24658: ``` error[E0599]: no method named `into_handle` found for struct `sys::unix::pipe::AnonPipe` in the current scope --> src/libstd/sys/windows/ext/process.rs:71:27 | 71 | self.into_inner().into_handle().into_raw() as *mut _ | ^^^^^^^^^^^ method not found in `sys::unix::pipe::AnonPipe` | ``` Because of rust-lang#73743, we only run typeck on demand. So this only causes an issue for functions returning `impl Trait`, which were already special cased by `ReplaceFunctionWithBody`. However, it now considers `async fn f() -> T` to be considered `impl Future<Output = T>`, where before it was considered to have a concrete `T` type. ## How will this affect future changes to rustdoc? - Any new changes to rustdoc will not be able to perform type checking without bringing back resolution errors in function bodies. + As a corollary, any new lints cannot require or perform type checking. In some cases this may require refactoring other parts of the compiler to perform type-checking only on-demand, see for example rust-lang#73743. + As a corollary, rustdoc can never again call `tcx.analysis()` unless this PR is reverted altogether. ## Current status - ~~I am not yet sure how to bring back `missing_docs` without running typeck. @eddyb suggested allowing lints to opt-out of type-checking, which would probably be another rabbit hole.~~ The opt-out was implemented in rust-lang#73743. However, of the rustc lints, now _only_ missing_docs is run and no other lints: rust-lang#73566 (comment). We need a team decision on whether that's an acceptable tradeoff. Note that all rustdoc lints are still run (`intra_doc_link_resolution_failure`, etc). **UPDATE**: This was deemed acceptable in rust-lang#73566 (comment) - ~~The implementation of optional errors in `rustc_resolve` is very brute force, it should probably be moved from `LateResolver` to `Resolver` to avoid duplicating the logic in many places.~~ I'm mostly happy with it now. - This no longer allows errors in `async fn f() -> T`. This caused breakage in 50 crates out of a full crater run, all of which (that I looked at) didn't compile when run with rustc directly. In other words, it used to be that they could not be compiled but could still be documented; now they can't be documented either. This needs a decision from the rustdoc team on whether this is acceptable breakage. **UPDATE**: This was deemed acceptable in rust-lang#73566 (comment) - ~~This makes `fn typeck_tables_of` in `rustc_typeck` public. This is not desired behavior, but needs the changes from rust-lang#74070 in order to be fixed.~~ Reverted.
Manishearth
added a commit
to Manishearth/rust
that referenced
this pull request
Jul 16, 2020
Don't run `everybody_loops` for rustdoc; instead ignore resolution errors r? @eddyb cc @petrochenkov, @GuillaumeGomez, @Manishearth, @ecstatic-morse, @marmeladema ~~Blocked on rust-lang#73743 Merged. ~~Blocked on crater run.~~ Crater popped up some ICEs ([now fixed](rust-lang#73566 (comment))). See [crater run](https://crater-reports.s3.amazonaws.com/pr-73566/index.html), [ICEs](rust-lang#73566 (comment)). ~~Blocked on rust-lang#74070 so that we don't make typeck_tables_of public when it shouldn't be.~~ Merged. Closes rust-lang#71820, closes rust-lang#71104, closes rust-lang#65863. ## What is the motivation for this change? As seen from a lengthy trail of PRs and issues (rust-lang#73532, rust-lang#73103, rust-lang#71820, rust-lang#71104), `everybody_loops` is causing bugs in rustdoc. The main issue is that it does not preserve the validity of the `DefId` tree, meaning that operations on DefIds may unexpectedly fail when called later. This is blocking intra-doc links (see rust-lang#73101). This PR starts by removing `everybody_loops`, fixing rust-lang#71104 and rust-lang#71820. However, that brings back the bugs seen originally in rust-lang#43348: Since libstd documents items for all platforms, the function bodies sometimes do not type check. Here are the errors from documenting `libstd` with `everybody_loops` disabled and no other changes: ```rust error[E0433]: failed to resolve: could not find `handle` in `sys` --> src/libstd/sys/windows/ext/process.rs:13:27 | 13 | let handle = sys::handle::Handle::new(handle as *mut _); | ^^^^^^ could not find `handle` in `sys` error[E0425]: cannot find function `symlink_inner` in module `sys::fs` --> src/libstd/sys/windows/ext/fs.rs:544:14 | 544 | sys::fs::symlink_inner(src.as_ref(), dst.as_ref(), false) | ^^^^^^^^^^^^^ not found in `sys::fs` error[E0425]: cannot find function `symlink_inner` in module `sys::fs` --> src/libstd/sys/windows/ext/fs.rs:564:14 | 564 | sys::fs::symlink_inner(src.as_ref(), dst.as_ref(), true) | ^^^^^^^^^^^^^ not found in `sys::fs` ``` ## Why does this need changes to `rustc_resolve`? Normally, this could be avoided by simply not calling the `typeck_item_bodies` pass. However, the errors above happen before type checking, in name resolution itself. Since name resolution is intermingled with macro expansion, and rustdoc needs expansion to happen before it knows all items to be documented, there needs to be someway to ignore _resolution_ errors in function bodies. An alternative solution suggested by @petrochenkov was to not run `everybody_loops` on anything containing a nested `DefId`. This would solve some of the immediate issues, but isn't bullet-proof: the following functions still could not be documented if the items in the body failed to resolve: - Functions containing a nested `DefId` (rust-lang#71104) - ~~Functions returning `impl Trait` (rust-lang#43878 These ended up not resolving anyway with this PR. - ~~`const fn`, because `loop {}` in `const fn` is unstable (rust-lang#43636 `const_loop` was just stabilized. This also isn't exactly what rustdoc wants, which is to avoid looking at function bodies in the first place. ## What changes were made? The hack implemented in this PR is to add an option to ignore all resolution errors in function bodies. This is enabled only for rustdoc. Since resolution errors are ignored, the MIR generated will be invalid, as can be seen in the following ICE: ```rust error: internal compiler error: broken MIR in DefId(0:11 ~ doc_cfg[8787]::uses_target_feature[0]) ("return type"): bad type [type error] --> /home/joshua/src/rust/src/test/rustdoc/doc-cfg.rs:51:1 | 51 | / pub unsafe fn uses_target_feature() { 52 | | content::should::be::irrelevant(); 53 | | } | |_^ ``` Fortunately, rustdoc does not need to access MIR in order to generate documentation. Therefore this also removes the call to `analyze()` in `rustdoc::run_core`. This has the side effect of not generating all lints by default. Most lints are safe to ignore (does rustdoc really need to run liveness analysis?) but `missing_docs` in particular is disabled when it should not be. Re-running `missing_docs` specifically does not help, because it causes the typechecking pass to be run, bringing back the errors from rust-lang#24658: ``` error[E0599]: no method named `into_handle` found for struct `sys::unix::pipe::AnonPipe` in the current scope --> src/libstd/sys/windows/ext/process.rs:71:27 | 71 | self.into_inner().into_handle().into_raw() as *mut _ | ^^^^^^^^^^^ method not found in `sys::unix::pipe::AnonPipe` | ``` Because of rust-lang#73743, we only run typeck on demand. So this only causes an issue for functions returning `impl Trait`, which were already special cased by `ReplaceFunctionWithBody`. However, it now considers `async fn f() -> T` to be considered `impl Future<Output = T>`, where before it was considered to have a concrete `T` type. ## How will this affect future changes to rustdoc? - Any new changes to rustdoc will not be able to perform type checking without bringing back resolution errors in function bodies. + As a corollary, any new lints cannot require or perform type checking. In some cases this may require refactoring other parts of the compiler to perform type-checking only on-demand, see for example rust-lang#73743. + As a corollary, rustdoc can never again call `tcx.analysis()` unless this PR is reverted altogether. ## Current status - ~~I am not yet sure how to bring back `missing_docs` without running typeck. @eddyb suggested allowing lints to opt-out of type-checking, which would probably be another rabbit hole.~~ The opt-out was implemented in rust-lang#73743. However, of the rustc lints, now _only_ missing_docs is run and no other lints: rust-lang#73566 (comment). We need a team decision on whether that's an acceptable tradeoff. Note that all rustdoc lints are still run (`intra_doc_link_resolution_failure`, etc). **UPDATE**: This was deemed acceptable in rust-lang#73566 (comment) - ~~The implementation of optional errors in `rustc_resolve` is very brute force, it should probably be moved from `LateResolver` to `Resolver` to avoid duplicating the logic in many places.~~ I'm mostly happy with it now. - This no longer allows errors in `async fn f() -> T`. This caused breakage in 50 crates out of a full crater run, all of which (that I looked at) didn't compile when run with rustc directly. In other words, it used to be that they could not be compiled but could still be documented; now they can't be documented either. This needs a decision from the rustdoc team on whether this is acceptable breakage. **UPDATE**: This was deemed acceptable in rust-lang#73566 (comment) - ~~This makes `fn typeck_tables_of` in `rustc_typeck` public. This is not desired behavior, but needs the changes from rust-lang#74070 in order to be fixed.~~ Reverted.
jyn514
added
A-cross
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Uses the special
--cfg doxconfiguration passed by rustbuild when runningrustdoc. Changes the#[cfg(platform)]into#[cfg(any(dox, platform))]so that platform-specific API are visible to rustdoc.Since platform-specific implementations often won't compile correctly on other platforms,
rustdocis changed to applyeverybody_loopsto the functions during documentation and doc-test harness.Since platform-specific code are documented on all platforms now, it could confuse users who found a useful API but is non-portable. Also, their examples will be doc-tested, so must be excluded when not testing on the native platform. An undocumented attribute
#[doc(cfg(...))]is introduced to serve the above purposed.Fixes #24658 (Does not fully implement #1998).