A single framework for gen-kill and generic dataflow problems
This is the prototype implementation discussed in rust-lang/compiler-team#202. You can read a high-level description of it in [the proposal](https://hackmd.io/@39Qr_z9cQhasi25sGjmFnA/Skvd9rztS) for that design meeting. This would eventually supersede the existing `BitDenotation` interface.
r? @ghost
cc @rust-lang/compiler (esp. @eddyb and @pnkfelix)
`TooGeneric` is encountered during WF checking when we cannot determine that a constant involving a generic parameter will always be evaluated successfully (rather than resulting in an error). In these cases, the burden of proof should be with the caller, so that we can avoid post-monomorphisation tim errors (which was the previous previous behaviour). This commit ensures that this situation produces a proper compiler error, rather than silently ignoring it or ICEing.
Rollup of 8 pull requests
Successful merges:
- #67734 (Remove appendix from Apache license)
- #67795 (Cleanup formatting code)
- #68290 (Fix some tests failing in `--pass check` mode)
- #68297 ( Filter and test predicates using `normalize_and_test_predicates` for const-prop)
- #68302 (Fix #[track_caller] and function pointers)
- #68339 (Add `riscv64gc-unknown-linux-gnu` into target list in build-manifest)
- #68381 (Added minor clarification to specification of GlobalAlloc::realloc.)
- #68397 (rustdoc: Correct order of `async` and `unsafe` in `async unsafe fn`s)
Failed merges:
r? @ghost
rustdoc: Correct order of `async` and `unsafe` in `async unsafe fn`s
The order was swapped in #61319 but rustdoc was never updated to match.
r? @GuillaumeGomez
Added minor clarification to specification of GlobalAlloc::realloc.
The specification of `realloc` is slightly unclear:
```
/// * `layout` must be the same layout that was used
/// to allocate that block of memory,
```
https://github.com/rust-lang/rust/blob/master/src/libcore/alloc.rs#L541-L542
In the case of an `alloc` or `alloc_zeroed` this is fairly evidently the `layout` parameter passed into the original call. In the case of a `realloc`, this I assume is `layout` modified to contain `new_size`. However, I could not find this case specified in the documentation. Thus technically in a sequence of calls to `realloc`, it would be valid to provide the second call to `realloc` the same `layout` as the first call to `realloc`, which is almost certainly not going to be handled correctly.
This PR attempts to clarify the specification.
Filter and test predicates using `normalize_and_test_predicates` for const-prop
Fixes#68264
Previously, I attempted to use
`substitute_normalize_and_test_predicates` to detect unsatisfiable
bounds. Unfortunately, since const-prop runs in a generic environment
(we don't have any of the function's generic parameters substituted),
this could lead to cycle errors when attempting to normalize predicates.
This check is replaced with a more precise check. We now only call
`normalize_and_test_predicates` on predicates that have the possibility
of being proved unsatisfiable - that is, predicates that don't depend
on anything local to the function (e.g. generic parameters). This
ensures that we don't hit cycle errors when we normalize said
predicates, while still ensuring that we detect unsatisfiable
predicates.
I haven't been able to come up with a minimization of the Diesel issue - however, I've verified that it compiles successfully.
Fix some tests failing in `--pass check` mode
Warnings reported at codegen or linking time either have to be converted to errors (preferable), or the tests for them need to be marked with `// ignore-pass` (as a last resort).
ecd5852194 turned them from errors to warnings, but that shouldn't be necessary because it's still clear from the `.stderr` output that the errors are lints and not hard-coded.
Cleanup formatting code
This removes a few leftover positional enum variants that were no longer used.
All details that are changed are unstable (and `#[doc(hidden)]`), so this should
not impact downstream code.
Remove appendix from Apache license
Looking at the codebase I noticed an oddity, in that the appendix of how use the Apache licence is still contained in the licence file. We don't put licence headers at the top of all of our files so I don't think we need to keep this. Alternatively we could delete everything above line 191 to have a shorter licence file.
When we're producing an rlib, we do not need anything more than an rmeta file
for each of our dependencies (this is indeed utilized by Cargo for pipelining).
Previously, we were still storing the paths of possible rlib/dylib crates, which
meant that they could still plausibly be accessed. With -Zbinary-dep-depinfo,
that meant that Cargo thought that rustc was using both the rlib and an (earlier
emitted) rmeta, and so needed a recompile, as the rlib may have finished writing
*after* compilation started (for more detail, see issue 68149).
This commit changes metadata loading to not store the filepaths of dylib/rlib if
we're going to end up creating an rlib only.
Stabilize ManuallyDrop::take
Tracking issue: closes#55422
FCP merge: https://github.com/rust-lang/rust/issues/55422#issuecomment-572653619
Reclaims the doc improvements from closed#62198.
-----
Stable version is a simple change if necessary.
Proposal: [relnotes] (this changes how to best take advantage of `ManuallyDrop`, esp. wrt. `Drop::drop` and finalize-by-value members)
fmt::Formatter is still not Send/Sync, but the UI test emitted two errors, for
the dyn Write and the Void inside Formatter. As of this PR, the Void is now
gone, but the dyn Write remains.
These are only called from one place and don't generally support being called
from other places; furthermore, they're the only formatter functions that look
at the `args` field (which a future commit will remove).
Previous implementation used a single mode type to store various pieces
of otherwise loosely related information:
* Whether debuginfo mode is in use or not.
* Which debuggers should run in general.
* Which debuggers are enabled for particular test case.
The new implementation introduces a separation between those aspects.
There is a single debuginfo mode parametrized by a debugger type.
The debugger detection is performed first and a separate configuration
is created for each detected debugger. The test cases are gathered
independently for each debugger which makes it trivial to implement
support for `ignore` / `only` conditions.
Functional changes:
* A single `debuginfo` entry point (rather than `debuginfo-cdb`, `debuginfo-gdb+lldb`, etc.).
* Debugger name is included in the test name.
* Test outputs are placed in per-debugger directory.
* Fixed spurious hash mismatch. Previously, the config mode would change
from `DebugInfoGdbLldb` (when collecting tests) to `DebugInfoGdb` or
`DebugInfoLldb` (when running them) which would affect hash computation.
* PYTHONPATH is additionally included in gdb hash.
* lldb-python and lldb-python-dir are additionally included in lldb hash.
Make sure that all upstream generics get re-exported from Rust dylibs.
This PR contains a fix for #67276. Rust dylibs would not re-export all generic instances when compiling with `-Zshare-generics=on` (=default for debug builds) which could lead to situations where the compiler expected certain generic instances to be available but then the linker would not find them.
### TODO
- [x] Write a regression test based on the description [here](https://github.com/rust-lang/rust/issues/67276#issuecomment-574613457).
- [x] Find out if this also fixes other issues related to https://github.com/rust-lang/rust/issues/64319.
r? @alexcrichton ~~(once the TODOs are done)~~
cc @pnkfelix @AlexKornitzer
