Previously, the static-libstdcpp setting was tied to llvm-tools such
that enabling the latter always enabled the latter. This seems
unfortunate, since it is entirely reasonable for someone to want to
_not_ statically link stdc++, but _also_ want to build the llvm-tools.
This patch therefore separates the two settings such that neither
implies the other.
On its own, that would change the default behavior in a way that's
likely to surprise users. Specifically, users who build llvm-tools
_likely_ want those tools to be statically compiled against libstdc++,
since otherwise users with older GLIBCXX will be unable to run the
vended tools. So, we also flip the default for the `static-libstdcpp`
setting such that builds always link statically against libstdc++ by
default, but it's _possible_ to opt out.
See also #94719.
PR #93918 made it so that `-static-libstdc++` was only set in one place,
and was only set during linking, but accidentally also made it so that
it is no longer passed when building LLD or sanitizers, only when
building LLVM itself. This moves the logic for setting
`-static-libstdc++` in the linker flags back to `configure_cmake` so
that it takes effect for all CMake invocations in `native.rs`.
As a side-effect, this also causes libstdc++ to be statically compiled
into sanitizers and LLD if `llvm-tools-enabled` is set but
`llvm-static-stdcpp` is not, even though previously it was only linked
statically if `llvm-static-stdcpp` was set explicitly. But that seems
more like the expected behavior anyway.
rustbuild: support RelWithDebInfo for lld
r? ``@alexcrichton``
LLVM has flags that control the level of debuginfo generated when building via rustbuild. Since LLD is built separately, it currently has no way of generating any debuginfo. This change re-uses the same flags as LLVM for LLD to ensure it has the same level of debuginfo generated as LLVM.
The majority of the code is only used by either rustbuild or
rustc_llvm's build script. Rust_build is compiled once for rustbuild and
once for every stage. This means that the majority of the code in this
crate is needlessly compiled multiple times. By moving only the code
actually used by the respective crates to rustbuild and rustc_llvm's
build script, this needless duplicate compilation is avoided.
First, this reverts the `CFLAGS`/`CXXFLAGS` of #93918. Those flags are
already read by `cc` and populated into `Build` earlier on in the
process. We shouldn't be overriding that based on `CFLAGS`, since `cc`
also respects overrides like `CFLAGS_{TARGET}` and `HOST_CFLAGS`, which
we want to take into account.
Second, this adds the same capability to specify target-specific
versions of `LDFLAGS` as we have through `cc` for the `C*` flags:
https://github.com/alexcrichton/cc-rs#external-configuration-via-environment-variables
Note that this also necessitated an update to compiletest to treat
CXXFLAGS separately from CFLAGS.
This tidies up the logic in `src/bootstrap/native.rs` such that:
- `CMAKE_*_LINKER_FLAGS` is not overridden if we add to it twice.
- `CMAKE_*_FLAGS` also include the standard `*FLAGS` environment
variables, which CMake respects when we _don't_ set `CMAKE_*_FLAGS`.
- `llvm.ldflags` from `config.toml` appends to the ldflags Rust's
bootstrap logic adds, rather than replacing them.
Fixes#93880.
I'm working on some LLVM patches in concert with a Rust patch, and it's
helping me quite a bit to have this as an option. It doesn't seem that
hard, so I figured I'd formalize it in x.py and send it upstream.
Make sure FileCheck is copied in the LLVM output directory
The tool, which is needed by parts of our test suite, is built as part of LLVM but is *not* copied to the directory containing the output LLVM binaries. This adds a flag to ensure the binary is copied. This shouldn't add any extra built time, as the flag just installs extra binaries that were already compiled.
This is not strictly needed for the test suite to work (as it also checks `build/$target/llvm/build/bin` for the binary), but it allows deleting the `build/$TARGET/llvm/build` directory (which also contains the intermediary build artifacts) without affecting the test suite, saving disk space.
The tool, which is needed by parts of our test suite, is built as part
of LLVM but is *not* copied to the directory containing the output LLVM
binaries. This adds a flag to ensure the binary is copied.
This shouldn't add any extra built time, as the flag just installs extra
binaries that were already compiled.
This shows up to 5% less instruction counts on multiple benchmarks, and up to
19% wins on the -j1 wall times for rustc self-compilation.
We can afford to spend the extra cycles building LLVM essentially once more for
the x86_64-unknown-linux-gnu CI build today. The builder finishes in around 50
minutes on average, and this adds just 10 more minutes. Given the sizeable
improvements in compiler performance, this is definitely worth it.
When LLVM_INCLUDE_TESTS is enabled (by default), LLVM requires
Python 3.6 for the lit test runner, otherwise only Python 3.0 is
required.
As we have many docker images using Ubuntu 16.04, which only has
Python 3.5, this avoids the need to install a newer Python version
for them.
This only updates the submodules the first time they're needed, instead
of unconditionally the first time you run x.py.
Ideally, this would move *all* submodules and not exclude some tools and
backtrace. Unfortunately, cargo requires all `Cargo.toml` files in the
whole workspace to be present to build any crate.
On my machine, this takes the time for an initial submodule clone (for
`x.py --help`) from 55.70 to 15.87 seconds.
This uses exactly the same logic as the LLVM update used, modulo some
minor cleanups:
- Use a local variable for `src.join(relative_path)`
- Remove unnecessary arrays for `book!` macro and make the macro simpler to use
- Add more comments
Use HTTPS links where possible
While looking at #86583, I wondered how many other (insecure) HTTP links were in `rustc`. This changes most other `http` links to `https`. While most of the links are in comments or documentation, there are a few other HTTP links that are used by CI that are changed to HTTPS.
Notes:
- I didn't change any to or in licences
- Some links don't support HTTPS :(
- Some `http` links were dead, in those cases I upgraded them to their new places (all of which used HTTPS)
BPF target support
This adds `bpfel-unknown-none` and `bpfeb-unknown-none`, two new no_std targets that generate little and big endian BPF. The approach taken is very similar to the cuda target, where `TargetOptions::obj_is_bitcode` is enabled and code generation is done by the linker.
I added the targets to `dist-various-2`. There are [some tests](https://github.com/alessandrod/bpf-linker/tree/main/tests/assembly) in bpf-linker and I'm planning to add more. Those are currently not ran as part of rust CI.
We do not install Zlib on the CI but recent builds somehow started picking it's shared version.
To avoid relying on CI binaries so let's explicitly disable it.
This enables better caching, since LLVM is only updated when needed, not
whenever x.py is run. Before, bootstrap.py had to use heuristics to
guess if LLVM would be needed, and updated the module more often than
necessary as a result.
This syncs the LLVM submodule only just before building the compiler, so
people working on the standard library never have to worry about it.
Example output:
```
Copying stage0 std from stage0 (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu / x86_64-unknown-linux-gnu)
Updating submodule src/llvm-project
Submodule 'src/llvm-project' (https://github.com/rust-lang/llvm-project.git) registered for path 'src/llvm-project'
Submodule path 'src/llvm-project': checked out 'f9a8d70b6e0365ac2172ca6b7f1de0341297458d'
```
- Don't try to update the LLVM submodule when using system LLVM
Previously, this would try to update LLVM unconditionally. Now the
submodule is only initialized if `llvm-config` is not set.
- Don't update LLVM submodule in dry runs
This prevents the following test failures:
```
running 17 tests
fatal: invalid gitfile format: /checkout/src/llvm-project/.git
test builder::tests::defaults::build_cross_compile ... FAILED
---- builder::tests::defaults::build_default stdout ----
thread 'main' panicked at 'command did not execute successfully: "git" "rev-parse" "HEAD"
expected success, got: exit code: 128', src/build_helper/lib.rs:139:9
```
- Try running git without --progress if it fails the first time
This avoids having to do version detection to see if --progress is
supported or not.
- Don't try to update submodules when the source repository isn't managed by git
- Update LLVM submodules that have already been checked out
- Only check for whether the submodule should be updated in lib.rs; update
it unconditionally in native.rs
- Add an assertion that `link-shared = true` when `thin-lto = true`.
Previously, link-shared would be silently overwritten.
- Get rid of `Option<bool>` in bootstrap/config.rs. Set defaults
immediately instead of delaying until later in bootstrap. This makes
it easier to find what the default value is.
- Remove redundant `config.x = false` when the default was already false
- Set defaults for `bindir` in `default_opts()` instead of `parse()`
- Update `download-ci-llvm = if-supported` option to match bootstrap.py
- Remove redundant check for link_shared. Previously, it was checked twice.
- Update various options in config.toml.example to their defaults.
Previously, some options showed an example value instead of the
default value.
- Fix incorrect defaults in config.toml.example
+ `use-libcxx` defaults to false
+ Add missing `check-stage = 0`
+ Update several defaults to be conditional (e.g. `if incremental { 10 } else { 100 }`)
- Remove redundant defaults in prose
- Use the same comment for the default and target-dependent `musl-root`
- Fix typos
- Link to `cc_detect` for `cc` and `cxx`, since the logic is ... complicated.
- Update more defaults to better reflect how they actually get set
- Remove ignored `gpg-password-file` option
This stopped being used in
7704d35acc,
but was never removed from config.toml.
- Remove unused flags from `config.toml`
+ Disallow `infodir` and `localstatedir` in `config.toml`
+ Allow the flags in `./configure`, but give a warning that they will be
ignored.
+ Fix incorrect comment that `datadir` will be ignored.
Example output:
```
$ ./configure --set install.infodir=xxx
configure: processing command line
configure:
configure: install.infodir := xxx
configure: build.configure-args := ['--set', 'install.infodir=xxx']
warning: infodir will be ignored
configure:
configure: writing `config.toml` in current directory
configure:
configure: run `python /home/joshua/rustc3/x.py --help`
configure:
```
- Update CHANGELOG
- Add "as an example" where appropriate
- Link to an issue instead of to ephemeral chats
