| Age | Commit message (Collapse) | Author | Lines |
|---|
|
|
|
rustc: Implement incremental "fat" LTO
Currently the compiler will produce an error if both incremental
compilation and full fat LTO is requested. With recent changes and the
advent of incremental ThinLTO, however, all the hard work is already
done for us and it's actually not too bad to remove this error!
This commit updates the codegen backend to allow incremental full fat
LTO. The semantics are that the input modules to LTO are all produce
incrementally, but the final LTO step is always done unconditionally
regardless of whether the inputs changed or not. The only real
incremental win we could have here is if zero of the input modules
changed, but that's so rare it's unlikely to be worthwhile to implement
such a code path.
cc #57968
cc rust-lang/cargo#6643
|
|
Stabilize linker-plugin based LTO (aka cross-language LTO)
This PR stabilizes [linker plugin based LTO](https://github.com/rust-lang/rust/issues/49879), also known as "cross-language LTO" because it allows for doing inlining and other optimizations across language boundaries in mixed Rust/C/C++ projects.
As described in the tracking issue, it works by making `rustc` emit LLVM bitcode instead of machine code, the same as `clang` does. A linker with the proper plugin (like LLD) can then run (Thin)LTO across all modules.
The feature has been implemented over a number of pull requests and there are various [codegen](https://github.com/rust-lang/rust/blob/master/src/test/codegen/no-dllimport-w-cross-lang-lto.rs) and [run](https://github.com/rust-lang/rust/tree/master/src/test/run-make-fulldeps/cross-lang-lto-clang)-[make](https://github.com/rust-lang/rust/tree/master/src/test/run-make-fulldeps/cross-lang-lto-upstream-rlibs) [tests](https://github.com/rust-lang/rust/tree/master/src/test/run-make-fulldeps/cross-lang-lto) that make sure that it keeps working.
It also works for building big projects like [Firefox](https://treeherder.mozilla.org/#/jobs?repo=try&revision=2ce2d5ddcea6fbff790503eac406954e469b2f5d).
The PR makes the feature available under the `-C linker-plugin-lto` flag. As discussed in the tracking issue it is not cross-language specific and also not LLD specific. `-C linker-plugin-lto` is descriptive of what it does. If someone has a better name, let me know `:)`
|
|
Cosmetic improvements to doc comments
This has been factored out from https://github.com/rust-lang/rust/pull/58036 to only include changes to documentation comments (throughout the rustc codebase).
r? @steveklabnik
Once you're happy with this, maybe we could get it through with r=1, so it doesn't constantly get invalidated? (I'm not sure this will be an issue, but just in case...) Anyway, thanks for your advice so far!
|
|
|
|
|
|
Currently the compiler will produce an error if both incremental
compilation and full fat LTO is requested. With recent changes and the
advent of incremental ThinLTO, however, all the hard work is already
done for us and it's actually not too bad to remove this error!
This commit updates the codegen backend to allow incremental full fat
LTO. The semantics are that the input modules to LTO are all produce
incrementally, but the final LTO step is always done unconditionally
regardless of whether the inputs changed or not. The only real
incremental win we could have here is if zero of the input modules
changed, but that's so rare it's unlikely to be worthwhile to implement
such a code path.
cc #57968
cc rust-lang/cargo#6643
|
|
|
|
|
|
|
|
|
|
|
|
Implement optimize(size) and optimize(speed) attributes
This PR implements both `optimize(size)` and `optimize(speed)` attributes.
While the functionality itself works fine now, this PR is not yet complete: the code might be messy in places and, most importantly, the compiletest must be improved with functionality to run tests with custom optimization levels. Otherwise the new attribute cannot be tested properly. Oh, and not all of the RFC is implemented – attribute propagation is not implemented for example.
# TODO
* [x] Improve compiletest so that tests can be written;
* [x] Assign a proper error number (E9999 currently, no idea how to allocate a number properly);
* [ ] Perhaps reduce the duplication in LLVM attribute assignment code…
|
|
|
|
|
|
Querify `entry_fn`
Analogous to https://github.com/rust-lang/rust/pull/57570 but this will also require few fixups in Miri so I decided to separate that (and it seems [CI doesn't let us break tools anymore](https://github.com/rust-lang/rust/pull/57392#issuecomment-453801540)? Or was that because it was a rollup PR?)
r? @nikomatsakis
|
|
Add a target option "merge-functions", and a corresponding -Z flag (works around #57356)
This commit adds a target option "merge-functions", which takes values in ("disabled", "trampolines", or "aliases" (default is "aliases")), to allow targets to opt out of the MergeFunctions LLVM pass. Additionally, the latest commit also adds an optional -Z flag, "merge-functions", which takes the same values and has precedence over the target option when both are specified.
This works around https://github.com/rust-lang/rust/issues/57356.
cc @eddyb @japaric @oli-obk @nox @nagisa
Also thanks to @denzp and @gnzlbg for discussing this on rust-cuda!
### Motivation
Basically, the problem is that the MergeFunctions pass, which rustc currently enables by default at -O2 and -O3 [1], and `extern "ptx-kernel"` functions (specific to the NVPTX target) are currently not compatible with each other. If the MergeFunctions pass is allowed to run, rustc can generate invalid PTX assembly (i.e. a PTX file that is not accepted by the native PTX assembler `ptxas`). Therefore we would like a way to opt out of the MergeFunctions pass, which is what our target option does.
### Related work
The current behavior of rustc is to enable MergeFunctions at -O2 and -O3 [1], and also to enable the use of function aliases within MergeFunctions [2] [3]. MergeFunctions seems to have some benefits, such as reducing code size and fixing a crash [4], which is why it is enabled. However, MergeFunctions both with and without function aliases is incompatible with the NVPTX target; a more detailed example for both cases is given below.
clang's "solution" is to have a "-fmerge-functions" flag that opts in to the MergeFunctions pass, but it is not enabled by default.
### Examples/more details
Consider an example Rust lib using `extern "ptx-kernel"` functions: https://github.com/peterhj/nvptx-mergefunc-bug/blob/master/nocore.rs. If we try to compile this with nightly rustc, we get the following compiler error:
LLVM ERROR: Module has aliases, which NVPTX does not support.
This error happens because: (1) functions `foo` and `bar` have the same body, so are candidates to be merged by MergeFunctions; and (2) rustc configures MergeFunctions to generate function aliases using the "mergefunc-use-aliases" LLVM option [2] [3], but the NVPTX backend does not support those aliases.
Okay, so we can try omitting "mergefunc-use-aliases", and then rustc will happily emit PTX assembly: https://github.com/peterhj/nvptx-mergefunc-bug/blob/master/nocore-mergefunc-nousealiases-bad.ptx. However, this PTX is invalid! When we try to assemble it with `ptxas` (I'm on the CUDA 9.2 toolchain), we get an assembler error:
ptxas nocore-mergefunc-nousealiases-bad.ptx, line 38; error : Illegal call target, device function expected
ptxas fatal : Ptx assembly aborted due to errors
What's happening is that MergeFunctions rewrites the `bar` function to call `foo`. However, directly calling an `extern "ptx-kernel"` function from another `extern "ptx-kernel"` is wrong.
If we disable the MergeFunctions pass from running at all, rustc generates correct PTX assembly: https://github.com/peterhj/nvptx-mergefunc-bug/blob/master/nocore-nomergefunc-ok.ptx
[1] https://github.com/rust-lang/rust/blob/a36b960df626cbb8bea74f01243318b73f0bd201/src/librustc_codegen_ssa/back/write.rs#L155
[2] https://github.com/rust-lang/rust/blob/a36b960df626cbb8bea74f01243318b73f0bd201/src/librustc_codegen_llvm/llvm_util.rs#L64
[3] https://github.com/rust-lang/rust/pull/56358
[4] https://github.com/rust-lang/rust/pull/49479
|
|
|
|
|
|
|
|
"trampolines", or "aliases (the default)) to allow targets to opt out of
the MergeFunctions LLVM pass. Also add a corresponding -Z option with
the same name and values.
This works around: https://github.com/rust-lang/rust/issues/57356
Motivation:
Basically, the problem is that the MergeFunctions pass, which rustc
currently enables by default at -O2 and -O3, and `extern "ptx-kernel"`
functions (specific to the NVPTX target) are currently not compatible
with each other. If the MergeFunctions pass is allowed to run, rustc can
generate invalid PTX assembly (i.e. a PTX file that is not accepted by
the native PTX assembler ptxas). Therefore we would like a way to opt
out of the MergeFunctions pass, which is what our target option does.
Related work:
The current behavior of rustc is to enable MergeFunctions at -O2 and -O3,
and also to enable the use of function aliases within MergeFunctions.
MergeFunctions both with and without function aliases is incompatible with
the NVPTX target.
clang's "solution" is to have a "-fmerge-functions" flag that opts in to
the MergeFunctions pass, but it is not enabled by default.
|
|
This ensures that arguments passed via `-C link-arg` can override the
first ones on the command line, for example allowing configuring of the
stack size.
|
|
|
|
Stabilize `linker-flavor` flag.
Part of #55396.
This commit moves the linker-flavor flag from a debugging option to a
codegen option, thus stabilizing it. There are no feature flags
associated with this flag.
r? @nagisa
|
|
This commit moves the linker-flavor flag from a debugging option to a
codegen option, thus stabilizing it. There are no feature flags
associated with this flag.
|
|
Instead of maybe storing its own sysroot and maybe deferring to the one
in `Session::opts`, just clone the latter when necessary so one is
always directly available. This removes the need for the getter.
|
|
|
|
Use a function to access the Hir map to be able to turn it into a query later
r? @eddyb
|
|
Discard LLVM modules earlier when performing ThinLTO
Currently ThinLTO is performed by first compiling all modules (and keeping them in memory), and then serializing them into ThinLTO buffers in a separate, synchronized step. Modules are later read back from ThinLTO buffers when running the ThinLTO optimization pipeline.
We can also find the following comment in `lto.rs`:
// FIXME: right now, like with fat LTO, we serialize all in-memory
// modules before working with them and ThinLTO. We really
// shouldn't do this, however, and instead figure out how to
// extract a summary from an in-memory module and then merge that
// into the global index. It turns out that this loop is by far
// the most expensive portion of this small bit of global
// analysis!
I don't think that what is suggested here is the right approach: One of the primary benefits of using ThinLTO over ordinary LTO is that it's not necessary to keep all the modules (merged or not) in memory for the duration of the linking step.
However, we currently don't really make use of this (at least for crate-local ThinLTO), because we keep all modules in memory until the start of the LTO step. This PR changes the implementation to instead perform the serialization into ThinLTO buffers directly after the initial optimization step.
Most of the changes here are plumbing to separate out fat and thin lto handling in `write.rs`, as these now use different intermediate artifacts. For fat lto this will be in-memory modules, for thin lto it will be ThinLTO buffers.
r? @alexcrichton
|
|
|
|
cleanup: remove static lifetimes from consts
A follow-up to https://github.com/rust-lang/rust/pull/56497.
|
|
Instead of keeping all modules in memory until thin LTO and only
serializing them then, serialize the module immediately after
it finishes optimizing.
|
|
Fat LTO merges into one module, so only return one module.
|
|
|
|
These are going to have different intermediate artifacts, so
create separate codepaths for them.
|
|
Instead of only determining whether some form of LTO is necessary,
determine whether thin, fat or no LTO is necessary. I've rewritten
the conditions in a way that I think is more obvious, i.e. specified
LTO type + additional preconditions.
|
|
|
|
|
|
|
|
rustc 1.30.0's linker flavor inference is a non-backwards compat change to -Clinker
Part of #55396.
This commit modifies linker flavor inference to only remove the extension
to the linker when performing inference if that extension is a 'exe'.
r? @nagisa
cc @alexcrichton @japaric
|
|
(compiler front-ends).
|
|
This commit modifies linker flavor inference to only remove the extension
to the linker when performing inference if that extension is a 'exe'.
|
|
This should handle recent symbol visibility changes happening, although
we'll likely want to tweak this in the future!
|
|
Our mangling scheme is not C++'s, so tell LLD to not demangle anything
so we can handle Rust-specific demangling ourselves.
|
|
Add SGX target to rustc
This adds the `x86_64-fortanix-unknown-sgx` target specification to the Rust compiler. See #56066 for more details about this target.
|
|
|
|
|
|
|
|
|
|
|
