Add support for artifact size profiling
This adds support for profiling artifact file sizes (incremental compilation artifacts and query cache to begin with).
Eventually we want to track this in perf.rlo so we can ensure that file sizes do not change dramatically on each pull request.
This relies on support in measureme: https://github.com/rust-lang/measureme/pull/169. Once that lands we can update this PR to not point to a git dependency.
This was worked on together with `@michaelwoerister.`
r? `@wesleywiser`
The support for runtime multi-threading was removed from LLVM. Calls to
`LLVMStartMultithreaded` became no-ops equivalent to checking if LLVM
was compiled with support for threads http://reviews.llvm.org/D4216.
This change adds a new compiler flag that can help reduce the size of
ELF binaries that contain many functions.
By default, when enabling function sections (which is the default for most
targets), the LLVM backend will generate different section names for each
function. For example, a function "func" would generate a section called
".text.func". Normally this is fine because the linker will merge all those
sections into a single one in the binary. However, starting with LLVM 12
(llvm/llvm-project@ee5d1a0), the backend will
also generate unique section names for exception handling, resulting in
thousands of ".gcc_except_table.*" sections ending up in the final binary
because some linkers don't currently merge or strip these EH sections.
This can bloat the ELF headers and string table significantly in
binaries that contain many functions.
The new option is analogous to Clang's -fno-unique-section-names, and
instructs LLVM to generate the same ".text" and ".gcc_except_table"
section for each function, resulting in smaller object files and
potentially a smaller final binary.
Create more accurate debuginfo for vtables.
Before this PR all vtables would have the same name (`"vtable"`) in debuginfo. Now they get an unambiguous name that identifies the implementing type and the trait that is being implemented.
This is only one of several possible improvements:
- This PR describes vtables as arrays of `*const u8` pointers. It would nice to describe them as structs where function pointer is represented by a field with a name indicative of the method it maps to. However, this requires coming up with a naming scheme that avoids clashes between methods with the same name (which is possible if the vtable contains multiple traits).
- The PR does not update the debuginfo we generate for the vtable-pointer field in a fat `dyn` pointer. Right now there does not seem to be an easy way of getting ahold of a vtable-layout without also knowing the concrete self-type of a trait object.
r? `@wesleywiser`
Add new tier-3 target: armv7-unknown-linux-uclibceabihf
This change adds a new tier-3 target: armv7-unknown-linux-uclibceabihf
This target is primarily used in embedded linux devices where system resources are slim and glibc is deemed too heavyweight. Cross compilation C toolchains are available [here](https://toolchains.bootlin.com/) or via [buildroot](https://buildroot.org).
The change is based largely on a previous PR #79380 with a few minor modifications. The author of that PR was unable to push the PR forward, and graciously allowed me to take it over.
Per the [target tier 3 policy](https://github.com/rust-lang/rfcs/blob/master/text/2803-target-tier-policy.md), I volunteer to be the "target maintainer".
This is my first PR to Rust itself, so I apologize if I've missed things!
Before this commit all vtables would have the same name "vtable" in
debuginfo. Now they get a name that identifies the implementing type
and the trait that is being implemented.
Implement `#[link_ordinal(n)]`
Allows the use of `#[link_ordinal(n)]` with `#[link(kind = "raw-dylib")]`, allowing Rust to link against DLLs that export symbols by ordinal rather than by name. As long as the ordinal matches, the name of the function in Rust is not required to match the name of the corresponding function in the exporting DLL.
Part of #58713.
Enable AutoFDO.
This largely involves implementing the options debug-info-for-profiling
and profile-sample-use and forwarding them on to LLVM.
AutoFDO can be used on x86-64 Linux like this:
rustc -O -Clink-arg='Wl,--no-rosegment' -Cdebug-info-for-profiling main.rs -o main
perf record -b ./main
create_llvm_prof --binary=main --out=code.prof
rustc -O -Cprofile-sample-use=code.prof main.rs -o main2
Now `main2` will have feedback directed optimization applied to it.
The create_llvm_prof tool can be obtained from this github repository:
https://github.com/google/autofdo
The option -Clink-arg='Wl,--no-rosegment' is necessary to avoid lld
putting an extra RO segment before the executable code, which would make
the binary silently incompatible with create_llvm_prof.
This largely involves implementing the options debug-info-for-profiling
and profile-sample-use and forwarding them on to LLVM.
AutoFDO can be used on x86-64 Linux like this:
rustc -O -Cdebug-info-for-profiling main.rs -o main
perf record -b ./main
create_llvm_prof --binary=main --out=code.prof
rustc -O -Cprofile-sample-use=code.prof main.rs -o main2
Now `main2` will have feedback directed optimization applied to it.
The create_llvm_prof tool can be obtained from this github repository:
https://github.com/google/autofdoFixes#64892.
[aarch64] add target feature outline-atomics
Enable outline-atomics by default as enabled in clang by the following commit
https://reviews.llvm.org/rGc5e7e649d537067dec7111f3de1430d0fc8a4d11
Performance improves by several orders of magnitude when using the LSE instructions
instead of the ARMv8.0 compatible load/store exclusive instructions.
Tested on Graviton2 aarch64-linux with
x.py build && x.py install && x.py test
Fix clippy lints
I'm currently working on allowing clippy to run on librustdoc after a discussion I had with `@Mark-Simulacrum.` So in the meantime, I fixed a few lints on the compiler crates.
Fix use after drop in self-profile with llvm events
self-profile with `-Z self-profile-events=llvm` have failed with a segmentation fault due to this use after drop.
this type of events can be more useful now that the new passmanager is the default.
Enable outline-atomics by default as enabled in clang by the following commit
https://reviews.llvm.org/rGc5e7e649d537067dec7111f3de1430d0fc8a4d11
Performance improves by several orders of magnitude when using the LSE instructions
instead of the ARMv8.0 compatible load/store exclusive instructions.
Tested on Graviton2 aarch64-linux with
x.py build && x.py install && x.py test
The new pass manager is enabled by default in clang since
Clang/LLVM 13. While the discussion about this is still ongoing
(https://lists.llvm.org/pipermail/llvm-dev/2021-August/152305.html)
it's expected that support for the legacy pass manager will be
dropped either in LLVM 14 or 15.
This switches us to use the new pass manager if LLVM >= 13 is used.
Avoid a couple of Symbol::as_str calls in cg_llvm
This should improve performance a tiny bit. Also remove `Symbol::len` and make `SymbolIndex` private.
Migrate in-tree crates to 2021
This replaces #89075 (cherry picking some of the commits from there), and closes#88637 and fixes#89074.
It excludes a migration of the library crates for now (see tidy diff) because we have some pending bugs around macro spans to fix there.
I instrumented bootstrap during the migration to make sure all crates moved from 2018 to 2021 had the compatibility warnings applied first.
Originally, the intent was to support cargo fix --edition within bootstrap, but this proved fairly difficult to pull off. We'd need to architect the check functionality to support running cargo check and cargo fix within the same x.py invocation, and only resetting sysroots on check. Further, it was found that cargo fix doesn't behave too well with "not quite workspaces", such as Clippy which has several crates. Bootstrap runs with --manifest-path ... for all the tools, and this makes cargo fix only attempt migration for that crate. We can't use e.g. --workspace due to needing to maintain sysroots for different phases of compilation appropriately.
It is recommended to skip the mass migration of Cargo.toml's to 2021 for review purposes; you can also use `git diff d6cd2c6c87 -I'^edition = .20...$'` to ignore the edition = 2018/21 lines in the diff.
This fixes compiling things like the `snap` crate after
https://reviews.llvm.org/D105462. I added a test that verifies the
additional attribute gets specified, and confirmed that I can build
cargo with both LLVM 13 and 14 with this change applied.
This just applies the suggested fixes from the compatibility warnings,
leaving any that are in practice spurious in. This is primarily intended to
provide a starting point to identify possible fixes to the migrations (e.g., by
avoiding spurious warnings).
A secondary commit cleans these up where they are false positives (as is true in
many of the cases).
Querify `FnAbi::of_{fn_ptr,instance}` as `fn_abi_of_{fn_ptr,instance}`.
*Note: opening this PR as draft because it's based on #88499*
This more or less replicates the `LayoutOf::layout_of` setup from #88499, to replace `FnAbi::of_{fn_ptr,instance}` with `FnAbiOf::fn_abi_of_{fn_ptr,instance}`, and also route them through queries (which `layout_of` has used for a while).
The two changes at the use sites (other than the names) are:
* return type is now wrapped in `&'tcx`
* the value *is* interned, which may affect performance
* the `extra_args` list is now an interned `&'tcx ty::List<Ty<'tcx>>`
* should be cheap (it's empty for anything other than C variadics)
Theoretically, a `FnAbiOfHelpers` implementer could choose to keep the `Result<...>` instead of eagerly erroring, but the only existing users of these APIs are codegen backends, so they don't (want to) take advantage of this.
At least miri could make use of this, since it prefers propagating errors (it "just" doesn't use `FnAbi` yet - cc `@RalfJung).`
The way this is done is probably less efficient than what is possible, because the queries handle the correctness-oriented API (i.e. the split into `fn` pointers vs instances), whereas a lower-level query could end up with more reuse between different instances with identical signatures.
r? `@nagisa` cc `@oli-obk` `@bjorn3`
Use a separate interner type for UniqueTypeId
Using symbol::Interner makes it very easy to mixup UniqueTypeId symbols
with the global interner. In fact the Debug implementation of
UniqueTypeId did exactly this.
Using a separate interner type also avoids prefilling the interner with
unused symbols and allow for optimizing the symbol interner for parallel
access without negatively affecting the single threaded module codegen.
Using symbol::Interner makes it very easy to mixup UniqueTypeId symbols
with the global interner. In fact the Debug implementation of
UniqueTypeId did exactly this.
Using a separate interner type also avoids prefilling the interner with
unused symbols and allow for optimizing the symbol interner for parallel
access without negatively affecting the single threaded module codegen.
Add -Z panic-in-drop={unwind,abort} command-line option
This PR changes `Drop` to abort if an unwinding panic attempts to escape it, making the process abort instead. This has several benefits:
- The current behavior when unwinding out of `Drop` is very unintuitive and easy to miss: unwinding continues, but the remaining drops in scope are simply leaked.
- A lot of unsafe code doesn't expect drops to unwind, which can lead to unsoundness:
- https://github.com/servo/rust-smallvec/issues/14
- https://github.com/bluss/arrayvec/issues/3
- There is a code size and compilation time cost to this: LLVM needs to generate extra landing pads out of all calls in a drop implementation. This can compound when functions are inlined since unwinding will then continue on to process drops in the callee, which can itself unwind, etc.
- Initial measurements show a 3% size reduction and up to 10% compilation time reduction on some crates (`syn`).
One thing to note about `-Z panic-in-drop=abort` is that *all* crates must be built with this option for it to be sound since it makes the compiler assume that dropping `Box<dyn Any>` will never unwind.
cc https://github.com/rust-lang/lang-team/issues/97
Move *_max methods back to util
change to inline instead of inline(always)
Remove valid_range_exclusive from scalar
Use WrappingRange instead
implement always_valid_for in a safer way
Fix accidental edit
Provide `layout_of` automatically (given tcx + param_env + error handling).
After #88337, there's no longer any uses of `LayoutOf` within `rustc_target` itself, so I realized I could move the trait to `rustc_middle::ty::layout` and redesign it a bit.
This is similar to #88338 (and supersedes it), but at no ergonomic loss, since there's no funky `C: LayoutOf<Ty = Ty>` -> `Ty: TyAbiInterface<C>` generic `impl` chain, and each `LayoutOf` still corresponds to one `impl` (of `LayoutOfHelpers`) for the specific context.
After this PR, this is what's needed to get `trait LayoutOf` (with the `layout_of` method) implemented on some context type:
* `TyCtxt`, via `HasTyCtxt`
* `ParamEnv`, via `HasParamEnv`
* a way to transform `LayoutError`s into the desired error type
* an error type of `!` can be paired with having `cx.layout_of(...)` return `TyAndLayout` *without* `Result<...>` around it, such as used by codegen
* this is done through a new `LayoutOfHelpers` trait (and so is specifying the type of `cx.layout_of(...)`)
When going through this path (and not bypassing it with a manual `impl` of `LayoutOf`), the end result is that only the error case can be customized, the query itself and the success paths are guaranteed to be uniform.
(**EDIT**: just noticed that because of the supertrait relationship, you cannot actually implement `LayoutOf` yourself, the blanket `impl` fully covers all possible context types that could ever implement it)
Part of the motivation for this shape of API is that I've been working on querifying `FnAbi::of_*`, and what I want/need to introduce for that looks a lot like the setup in this PR - in particular, it's harder to express the `FnAbi` methods in `rustc_target`, since they're much more tied to `rustc` concepts.
r? `@nagisa` cc `@oli-obk` `@bjorn3`
Include debug info for the allocator shim
Issue Details:
In some cases it is necessary to generate an "allocator shim" to forward various Rust allocation functions (e.g., `__rust_alloc`) to an underlying function (e.g., `malloc`). However, since this allocator shim is a manually created LLVM module it is not processed via the normal module processing code and so no debug info is generated for it (if debugging info is enabled).
Fix Details:
* Modify the `debuginfo` code to allow creating debug info for a module without a `CodegenCx` (since it is difficult, and expensive, to create one just to emit some debug info).
* After creating the allocator shim add in basic debug info.
Path remapping: Make behavior of diagnostics output dependent on presence of --remap-path-prefix.
This PR fixes a regression (#87745) with `--remap-path-prefix` where the flag stopped causing diagnostic messages to be remapped as well. The regression was introduced in https://github.com/rust-lang/rust/pull/83813 where we erroneously assumed that remapping of diagnostic messages was not desired anymore (because #70642 partially undid that functionality with nobody objecting).
The issue is fixed by making `--remap-path-prefix` remap diagnostic messages again, including for paths that have been remapped in upstream crates (e.g. the standard library). This means that "sysroot-localization" (implemented in #70642) is also disabled if `rustc` is invoked with `--remap-path-prefix`. The assumption is that once someone starts explicitly remapping paths they also don't want paths to their local Rust installation in their build output.
In the future we might want to give more fine-grained control over this behavior via compiler flags (see https://github.com/rust-lang/rfcs/pull/3127 for a related RFC). For now this PR is intended as a regression fix.
This PR is an alternative to https://github.com/rust-lang/rust/pull/88191, which makes diagnostic messages be remapped unconditionally. That approach, however, would effectively revert #70642.
Fixes https://github.com/rust-lang/rust/issues/87745.
cc `@cbeuw`
r? `@ghost`
Issue Details:
In some cases it is necessary to generate an "allocator shim" to forward various Rust allocation functions (e.g., `__rust_alloc`) to an underlying function (e.g., `malloc`). However, since this allocator shim is a manually created LLVM module it is not processed via the normal module processing code and so no debug info is generated for it (if debugging info is enabled).
Fix Details:
* Modify the `debuginfo` code to allow creating debug info for a module without a `CodegenCx` (since it is difficult, and expensive, to create one just to emit some debug info).
* After creating the allocator shim add in basic debug info.
rustc_target: `TyAndLayout::field` should never error.
This refactor (making `TyAndLayout::field` return `TyAndLayout` without any `Result` around it) is based on a simple observation, regarding `TyAndLayout::field`:
If `cx.layout_of(ty)` succeeds (for some `cx` and `ty`), then `.field(cx, i)` on the resulting `TyAndLayout` should *always* succeed in computing `cx.layout_of(field_ty)` (where `field_ty` is the type of the `i`th field of `ty`).
The reason for this is that no matter which field is chosen, `cx.layout_of(field_ty)` *will have already been computed*, as part of computing `cx.layout_of(ty)`, as we cannot determine the layout of *any* type without considering the layouts of *all* of its fields.
And so it should be fine to turn any errors into ICEs, since they likely indicate a `cx` mismatch, or some other edge case that is due to a compiler bug (as opposed to ever being an user-facing error).
<hr/>
Each commit should probably be reviewed separately, though note that there's some `where` clauses (in `rustc_target::abi::call::*`) that change in most commits.
cc `@nagisa` `@oli-obk`
S390x inline asm
This adds register definitions and constraint codes for the s390x general and floating point registers necessary for fixing #85931; as well as a few tests.
Further testing is needed, but I am a little unsure of what specific tests should be added to `src/test/assembly/asm/s390x.rs` to address this.
lazily "compute" anon const default substs
Continuing the work of #83086, this implements the discussed solution for the [unused substs problem](https://github.com/rust-lang/project-const-generics/blob/master/design-docs/anon-const-substs.md#unused-substs). As of now, anonymous constants inherit all of their parents generics, even if they do not use them, e.g. in `fn foo<T, const N: usize>() -> [T; N + 1]`, the array length has `T` as a generic parameter even though it doesn't use it. These *unused substs* cause some backwards incompatible, and imo incorrect behavior, e.g. #78369.
---
We do not actually filter any generic parameters here and the `default_anon_const_substs` query still a dummy which only checks that
- we now prevent the previously existing query cycles and are able to call `predicates_of(parent)` when computing the substs of anonymous constants
- the default anon consts substs only include the typeflags we assume it does.
Implementing that filtering will be left as future work.
---
The idea of this PR is to delay the creation of the anon const substs until after we've computed `predicates_of` for the parent of the anon const. As the predicates of the parent can however contain the anon const we still have to create a `ty::Const` for it.
We do this by changing the substs field of `ty::Unevaluated` to an option and modifying accesses to instead call the method `unevaluated.substs(tcx)` which returns the substs as before. If the substs - now `substs_` - of `ty::Unevaluated` are `None`, it means that the anon const currently has its default substs, i.e. the substs it has when first constructed, which are the generic parameters it has available. To be able to call `unevaluated.substs(tcx)` in a `TypeVisitor`, we add the non-defaulted method `fn tcx_for_anon_const_substs(&self) -> Option<TyCtxt<'tcx>>`. In case `tcx_for_anon_const_substs` returns `None`, unknown anon const default substs are skipped entirely.
Even when `substs_` is `None` we still have to treat the constant as if it has its default substs. To do this, `TypeFlags` are modified so that it is clear whether they can still change when *exposing* any anon const default substs. A new flag, `HAS_UNKNOWN_DEFAULT_CONST_SUBSTS`, is added in case some default flags are missing.
The rest of this PR are some smaller changes to either not cause cycles by trying to access the default anon const substs too early or to be able to access the `tcx` in previously unused locations.
cc `@rust-lang/project-const-generics`
r? `@nikomatsakis`
Use custom wrap-around type instead of RangeInclusive
Two reasons:
1. More memory is allocated than necessary for `valid_range` in `Scalar`. The range is not used as an iterator and `exhausted` is never used.
2. `contains`, `count` etc. methods in `RangeInclusive` are doing very unhelpful(and dangerous!) things when used as a wrap-around range. - In general this PR wants to limit potentially confusing methods, that have a low probability of working.
Doing a local perf run, every metric shows improvement except for instructions.
Max-rss seem to have a very consistent improvement.
Sorry - newbie here, probably doing something wrong.
Stop emitting the `dso_local` LLVM attribute for external symbols under the static relocation model on macOS.
This matches Clang's behavior:
973cb2c326/clang/lib/CodeGen/CodeGenModule.cpp (L1038-L1040)
Even if `dso_local` were properly supported in this way on macOS, it seems
incorrect to add this annotation as liberally as we did. The `dso_local`
annotation is for symbols that ultimately end up in the same linkage unit, but
we were adding this annotation even for `static` values inside `extern` blocks
marked with `#[link(type="framework")]`, which should be considered dynamically
linked. Note that Clang likewise avoids emitting `dso_local` for `dllimport`
symbols:
973cb2c326/clang/lib/CodeGen/CodeGenModule.cpp (L1005-L1007)
This issue caused breakage in the `ring` crate, which links to a symbol defined
in `Security.framework` that ultimately resolves to address `0x0`:
b94d61e044/src/rand.rs (L390)
For this symbol, the use of `dso_local` causes LLVM to emit a relocation of
type `X86_64_RELOC_SIGNED`, which is a 32-bit signed PC-relative offset. If the
binary is large enough, `0x0` might be out of range, and the link will fail.
Avoiding `dso_local` causes LLVM to use the GOT instead, emitting a relocation
of type `X86_64_RELOC_GOT_LOAD`, which will properly handle the large offset
and cause the link to succeed.
As a side note, the static relocation model is effectively deprecated for
security reasons on macOS, as it prohibits PIE. It's also completely
unsupported on Apple Silicon, so I don't think it's worth going to the effort
of properly supporting this model on that platform.
Upgrade to LLVM 13
Work in progress update to LLVM 13. Main changes:
* InlineAsm diagnostics reported using SrcMgr diagnostic kind are now handled. Previously these used a separate diag handler.
* Codegen tests are updated for additional attributes.
* Some data layouts have changed.
* Switch `#[used]` attribute from `llvm.used` to `llvm.compiler.used` to avoid SHF_GNU_RETAIN flag introduced in https://reviews.llvm.org/D97448, which appears to trigger a bug in older versions of gold.
* Set `LLVM_INCLUDE_TESTS=OFF` to avoid Python 3.6 requirement.
Upstream issues:
* ~~https://bugs.llvm.org/show_bug.cgi?id=51210 (InlineAsm diagnostic reporting for module asm)~~ Fixed by 1558bb80c0.
* ~~https://bugs.llvm.org/show_bug.cgi?id=51476 (Miscompile on AArch64 due to incorrect comparison elimination)~~ Fixed by 81b106584f.
* https://bugs.llvm.org/show_bug.cgi?id=51207 (Can't set custom section flags anymore). Problematic change reverted in our fork, https://reviews.llvm.org/D107216 posted for upstream revert.
* https://bugs.llvm.org/show_bug.cgi?id=51211 (Regression in codegen for #83623). This is an optimization regression that we may likely have to eat for this release. The fix for #83623 was based on an incorrect premise, and this needs to be properly addressed in the MergeICmps pass.
The [compile-time impact](https://perf.rust-lang.org/compare.html?start=ef9549b6c0efb7525c9b012148689c8d070f9bc0&end=0983094463497eec22d550dad25576a894687002) is mixed, but quite positive as LLVM upgrades go.
The LLVM 13 final release is scheduled for Sep 21st. The current nightly is scheduled for stable release on Oct 21st.
r? `@ghost`
This matches Clang's behavior:
973cb2c326/clang/lib/CodeGen/CodeGenModule.cpp (L1038-L1040)
Even if `dso_local` were properly supported in this way on macOS, it seems
incorrect to add this annotation as liberally as we did. The `dso_local`
annotation is for symbols that ultimately end up in the same linkage unit, but
we were adding this annotation even for `static` values inside `extern` blocks
marked with `#[link(type="framework")]`, which should be considered dynamically
linked. Note that Clang likewise avoids emitting `dso_local` for `dllimport`
symbols:
973cb2c326/clang/lib/CodeGen/CodeGenModule.cpp (L1005-L1007)
This issue caused breakage in the `ring` crate, which links to a symbol defined
in `Security.framework` that ultimately resolves to address `0x0`:
b94d61e044/src/rand.rs (L390)
For this symbol, the use of `dso_local` causes LLVM to emit a relocation of
type `X86_64_RELOC_SIGNED`, which is a 32-bit signed PC-relative offset. If the
binary is large enough, `0x0` might be out of range, and the link will fail.
Avoiding `dso_local` causes LLVM to use the GOT instead, emitting a relocation
of type `X86_64_RELOC_GOT_LOAD`, which will properly handle the large offset
and cause the link to succeed.
As a side note, the static relocation model is effectively deprecated for
security reasons on macOS, as it prohibits PIE. It's also completely
unsupported on Apple Silicon, so I don't think it's worth going to the effort
of properly supporting this model on that platform.
Matthias Krüger