When building LLVM/LLD as part of a build that asks LLVM to generate profiles, e.g. when
doing PGO, cmake or clang-cl don't automatically link clang's profiler runtime in,
causing undefined reference errors at link-time.
We do that manually, by adding clang's resource library folder to the library search path:
- for LLVM itself, by extending the linker args that `rustc_llvm`'s build script
uses, to avoid the linker errors when linking `rustc_driver`.
- for LLD, by extending cmake's linker flags during the LLD build step.
Inline assembly uses the target features to determine which registers
are available on the current target. However it needs to be able to
access unstable target features for this.
Fixes#99071
Rollup of 5 pull requests
Successful merges:
- #98882 (explain doc comments in macros a bit)
- #98907 (Deny float const params even when `adt_const_params` is enabled)
- #99091 (Do not mention private types from other crates as impl candidates)
- #99140 (Implement `SourceMap::is_span_accessible`)
- #99147 (Mention similarly named associated type even if it's not clearly in supertrait)
Failed merges:
r? `@ghost`
`@rustbot` modify labels: rollup
Mention similarly named associated type even if it's not clearly in supertrait
Due to query cycle avoidance, we sometimes restrict the candidates in `complain_about_assoc_type_not_found` too much so that we can't detect typo replacements from just supertraits.
This creates a more general note of the existence of a similarly named associated type from _all_ visible traits when possible.
Fixes#55673
Implement `SourceMap::is_span_accessible`
This patch adds `SourceMap::is_span_accessible` and replaces `span_to_snippet(span).is_ok()` and `span_to_snippet(span).is_err()` with it. This removes a `&str` to `String` conversion.
explain doc comments in macros a bit
Open to suggestions on improving this... macro parsing is very foreign to me.
Should we have a structured suggestion to turn them into their regular non-doc comments?
Fixes#92846Fixes#97850
Inline Windows `OsStrExt::encode_wide`
User crates currently produce much more code than necessary because the optimizer fails to make assumptions about this method.
Remove unsupported options in configure.py
I've seen people using `optimize = false` and `full-bootstrap = true` in the past, without knowing
that they're not recommended. Remove `optimize` and a few other options that are always a bad idea,
and document that full-bootstrap is only for testing reproducible builds.
In practice, this doesn't matter very much because the script takes ~no time to run.
But this makes `CARGO_LOG=info` easier to read, and theoretically saves a few milliseconds.
Fix `x build library/std compiler/rustc`
Previously, this was broken because of improper caching:
1. `StepDescription::maybe_run` builds `Compile::Std`, which only built `std` and not `proc_macro`
1. `Std` calls `builder.ensure(StdLink)`
1. `Rustc` calls `ensure(Std)`, which builds all crates, including `proc_macro`
1. `Rustc` calls `ensure(StdLink)`. `ensure` would see that it had already been run and do nothing. <-- bug is here
1. Cargo gives an error that `proc_macro` doesn't exist.
This fixes the caching by adding `crates` to `StdLink`, so it will get rerun if the crates that are built change.
Fixes https://github.com/rust-lang/rust/issues/99129.
Before:
```
thread 'main' panicked at 'Unable to build RLS', dist.rs:42:9
```
After:
```
thread 'main' panicked at 'Unable to build submodule tool RLS (use `missing-tools = true` to ignore this failure)
note: not all tools are available on all nightlies
help: see https://forge.rust-lang.org/infra/toolstate.html for more information', dist.rs:43:9
```
Allow using `download-ci-llvm = true` outside the git checkout
`@bjorn3` noticed that this is already allowed today when download-llvm is disabled, but breaks with it enabled:
```
$ ./rust2/x.py build
fatal: not a git repository (or any of the parent directories): .git
thread 'main' panicked at 'command did not execute successfully: "git" "rev-list" "--author=bors@rust-lang.org" "-n1" "--first-parent" "HEAD" "--" "/home/jnelson/rust-lang/rust2/src/llvm-project" "/home/jnelson/rust-lang/rust2/src/bootstrap/download-ci-llvm-stamp" "/home/jnelson/rust-lang/rust2/src/version"
expected success, got: exit status: 128', src/bootstrap/native.rs:134:20
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
```
Support it too for consistency. It's unclear to me when anyone would need to use this, but `@bjorn3`
feels we should support it, and it's not much additional effort to get it working.
Previously, this was broken because of improper caching:
1. `StepDescription::maybe_run` builds `Compile::Std`, which only built `std` and not `proc_macro`
1. `Std` calls `builder.ensure(StdLink)`
1. `Rustc` calls `ensure(Std)`, which builds all crates, including `proc_macro`
1. `Rustc` calls `ensure(StdLink)`. `ensure` would see that it had already been run and do nothing. <-- bug is here
1. Cargo gives an error that `proc_macro` doesn't exist.
This fixes the caching by adding `crates` to `StdLink`, so it will get rerun if the crates that are
built change. This also does the same for `RustcLink`; it doesn't matter in practice currently
because nothing uses it except `impl Step for Rustc`, but it will avoid bugs if we start using it in
the future (e.g. to build individual crates for rustfmt).
Fix `download-ci-llvm` NixOS patching for binaries
LLVM tools should also be patched, since they are used in some tests, specially,
- src/test/run-make-fulldeps/cross-lang-lto (llvm-ar)
- src/test/run-make-fulldeps/cross-lang-lto-upstream-rlibs (llvm-ar)
- src/test/run-make-fulldeps/issue-64153 (llvm-objdump)
To be more future proof, we should patch all binaries in `bin`, which is done in this PR.
