Allow loading LLVM plugins with both legacy and new pass manager
Opening a draft PR to get feedback and start discussion on this feature. There is already a codegen option `passes` which allow giving a list of LLVM pass names, however we currently can't use a LLVM pass plugin (as described here : https://llvm.org/docs/WritingAnLLVMPass.html), the only available passes are the LLVM built-in ones.
The proposed modification would be to add another codegen option `pass-plugins`, which can be set with a list of paths to shared library files. These libraries are loaded using the LLVM function `PassPlugin::Load`, which calls the expected symbol `lvmGetPassPluginInfo`, and register the pipeline parsing and optimization callbacks.
An example usage with a single plugin and 3 passes would look like this in the `.cargo/config`:
```toml
rustflags = [
"-C", "pass-plugins=/tmp/libLLVMPassPlugin",
"-C", "passes=pass1 pass2 pass3",
]
```
This would give the same functionality as the opt LLVM tool directly integrated in rust build system.
Additionally, we can also not specify the `passes` option, and use a plugin which inserts passes in the optimization pipeline, as one could do using clang.
This doesn't seem to be necessary anymore, although I don't know
at which point or why that changed.
Forcing -O1 makes some tests fail under NewPM, because NewPM also
performs inlining at -O1, so it ends up performing much more
optimization in practice than before.
This commit adds a new stable codegen option to rustc,
`-Csplit-debuginfo`. The old `-Zrun-dsymutil` flag is deleted and now
subsumed by this stable flag. Additionally `-Zsplit-dwarf` is also
subsumed by this flag but still requires `-Zunstable-options` to
actually activate. The `-Csplit-debuginfo` flag takes one of
three values:
* `off` - This indicates that split-debuginfo from the final artifact is
not desired. This is not supported on Windows and is the default on
Unix platforms except macOS. On macOS this means that `dsymutil` is
not executed.
* `packed` - This means that debuginfo is desired in one location
separate from the main executable. This is the default on Windows
(`*.pdb`) and macOS (`*.dSYM`). On other Unix platforms this subsumes
`-Zsplit-dwarf=single` and produces a `*.dwp` file.
* `unpacked` - This means that debuginfo will be roughly equivalent to
object files, meaning that it's throughout the build directory
rather than in one location (often the fastest for local development).
This is not the default on any platform and is not supported on Windows.
Each target can indicate its own default preference for how debuginfo is
handled. Almost all platforms default to `off` except for Windows and
macOS which default to `packed` for historical reasons.
Some equivalencies for previous unstable flags with the new flags are:
* `-Zrun-dsymutil=yes` -> `-Csplit-debuginfo=packed`
* `-Zrun-dsymutil=no` -> `-Csplit-debuginfo=unpacked`
* `-Zsplit-dwarf=single` -> `-Csplit-debuginfo=packed`
* `-Zsplit-dwarf=split` -> `-Csplit-debuginfo=unpacked`
Note that `-Csplit-debuginfo` still requires `-Zunstable-options` for
non-macOS platforms since split-dwarf support was *just* implemented in
rustc.
There's some more rationale listed on #79361, but the main gist of the
motivation for this commit is that `dsymutil` can take quite a long time
to execute in debug builds and provides little benefit. This means that
incremental compile times appear that much worse on macOS because the
compiler is constantly running `dsymutil` over every single binary it
produces during `cargo build` (even build scripts!). Ideally rustc would
switch to not running `dsymutil` by default, but that's a problem left
to get tackled another day.
Closes#79361
llvm-dwp concatenates `DW_AT_comp_dir` with `DW_AT_GNU_dwo_name` (only
when `DW_AT_comp_dir` exists), which can result in it failing to find
the DWARF object files.
In earlier testing, `DW_AT_comp_dir` wasn't present in the final
object and the current directory was the output directory.
When running tests through compiletest, the working directory of the
compilation is different from output directory and that resulted in
`DW_AT_comp_dir` being in the object file (and set to the current
working directory, rather than the output directory), and
`DW_AT_GNU_dwo_name` being set to the full path (rather than just
the filename), so llvm-dwp was failing.
This commit changes the compilation directory provided to LLVM to match
the output directory, where DWARF objects are output; and ensures that
only the filename is used for `DW_AT_GNU_dwo_name`.
Signed-off-by: David Wood <david@davidtw.co>
This commit implements Split DWARF support, wiring up the flag (added in
earlier commits) to the modified FFI wrapper (also from earlier
commits).
Signed-off-by: David Wood <david@davidtw.co>
This commit removes the `TargetMachineFactory` struct and adds a
`TargetMachineFactoryFn` type alias which is used everywhere that the
previous, long type was used.
Signed-off-by: David Wood <david@davidtw.co>
During incremental ThinLTO compilation, we attempt to re-use the
optimized (post-ThinLTO) bitcode file for a module if it is 'safe' to do
so.
Up until now, 'safe' has meant that the set of modules that our current
modules imports from/exports to is unchanged from the previous
compilation session. See PR #67020 and PR #71131 for more details.
However, this turns out be insufficient to guarantee that it's safe
to reuse the post-LTO module (i.e. that optimizing the pre-LTO module
would produce the same result). When LLVM optimizes a module during
ThinLTO, it may look at other information from the 'module index', such
as whether a (non-imported!) global variable is used. If this
information changes between compilation runs, we may end up re-using an
optimized module that (for example) had dead-code elimination run on a
function that is now used by another module.
Fortunately, LLVM implements its own ThinLTO module cache, which is used
when ThinLTO is performed by a linker plugin (e.g. when clang is used to
compile a C proect). Using this cache directly would require extensive
refactoring of our code - but fortunately for us, LLVM provides a
function that does exactly what we need.
The function `llvm::computeLTOCacheKey` is used to compute a SHA-1 hash
from all data that might influence the result of ThinLTO on a module.
In addition to the module imports/exports that we manually track, it
also hashes information about global variables (e.g. their liveness)
which might be used during optimization. By using this function, we
shouldn't have to worry about new LLVM passes breaking our module re-use
behavior.
In LLVM, the output of this function forms part of the filename used to
store the post-ThinLTO module. To keep our current filename structure
intact, this PR just writes out the mapping 'CGU name -> Hash' to a
file. To determine if a post-LTO module should be reused, we compare
hashes from the previous session.
This should unblock PR #75199 - by sheer chance, it seems to have hit
this issue due to the particular CGU partitioning and optimization
decisions that end up getting made.
