coverage: Rearrange the code for embedding per-function coverage metadata
This is a series of refactorings to the code that prepares and embeds per-function coverage metadata records (“covfun records”) in the `__llvm_covfun` linker section of the final binary. The `llvm-cov` tool reads this metadata from the binary when preparing a coverage report.
Beyond general cleanup, a big motivation behind these changes is to pave the way for re-landing an updated version of #133418.
---
There should be no change in compiler output, as demonstrated by the absence of (meaningful) changes to coverage tests.
The first patch is just moving code around, so I suggest looking at the other patches to see the actual changes.
---
try-job: x86_64-gnu
try-job: x86_64-msvc
try-job: aarch64-apple
rustc_target: ppc64 target string fixes for LLVM 20
LLVM continues to clean these up, and we continue to make this consistent. This is similar to 9caced7bad, e985396145, and
a10e744faf.
```@rustbot``` label: +llvm-main
Add the `power8-crypto` target feature
Add the `power8-crypto` target feature. This will enable adding some new PPC intrinsics in stdarch (specifically AES, SHA and CLMUL intrinsics). The implied target feature is from [here](https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/PowerPC/PPC.td)
```@rustbot``` label A-target-feature O-PowerPC
LLVM continues to clean these up, and we continue to make this
consistent. This is similar to 9caced7bad,
e985396145, and
a10e744faf.
`@rustbot` label: +llvm-main
coverage: Use a query to find counters/expressions that must be zero
As of #133446, this query (`coverage_ids_info`) determines which counter/expression IDs are unused. So with only a little extra work, we can take the code that was using that information to determine which coverage counters/expressions must be zero, and move that inside the query as well.
There should be no change in compiler output.
A bunch of cleanups
These are all extracted from a branch I have to get rid of driver queries. Most of the commits are not directly necessary for this, but were found in the process of implementing the removal of driver queries.
Previous PR: https://github.com/rust-lang/rust/pull/132410
This query (`coverage_ids_info`) already determines which counter/expression
IDs are unused, so it only takes a little extra effort to also determine which
counters/expressions must have a value of zero.
It was inconsistently done (sometimes even within a single function) and
most of the rest of the compiler uses fatal errors instead, which need
to be caught using catch_with_exit_code anyway. Using fatal errors
instead of ErrorGuaranteed everywhere in the driver simplifies things a
bit.
LLVM does not include an implementation of the va_arg instruction for
Xtensa. From what I understand, this is a conscious decision and
instead language frontends are encouraged to implement it themselves.
The rationale seems to be that loading values correctly requires
language and ABI-specific knowledge that LLVM lacks.
This is true of most architectures, and rustc already provides
implementation for a number of them. This commit extends the support to
include Xtensa.
See https://lists.llvm.org/pipermail/llvm-dev/2017-August/116337.html
for some discussion on the topic.
Unfortunately there does not seem to be a reference document for the
semantics of the va_list and va_arg on Xtensa. The most reliable source
is the GCC implementation, which this commit tries to follow. Clang also
provides its own compatible implementation.
This was tested for all the types that rustc allows in variadics.
Co-authored-by: Brian Tarricone <brian@tarricone.org>
Co-authored-by: Jonathan Bastien-Filiatrault <joe@x2a.org>
Co-authored-by: Paul Lietar <paul@lietar.net>
coverage: Use a query to identify which counter/expression IDs are used
Given that we already have a query to identify the highest-numbered counter ID in a MIR body, we can extend that query to also build bitsets of used counter/expression IDs. That lets us avoid some messy coverage bookkeeping during the main MIR traversal for codegen.
This does mean that we fail to treat some IDs as used in certain MIR-inlining scenarios, but I think that's fine, because it means that the results will be consistent across all instantiations of a function.
---
There's some more cleanup I want to do in the function coverage collector, since it isn't really collecting anything any more, but I'll leave that for future work.
Respect verify-llvm-ir option in the backend
We are currently unconditionally verifying the LLVM IR in the backend (twice), ignoring the value of the verify-llvm-ir option. This has substantial compile-time impact for debug builds.
Support input/output in vector registers of PowerPC inline assembly
This extends currently clobber-only vector registers (`vreg`) support to allow passing `#[repr(simd)]` types as input/output.
| Architecture | Register class | Target feature | Allowed types |
| ------------ | -------------- | -------------- | -------------- |
| PowerPC | `vreg` | `altivec` | `i8x16`, `i16x8`, `i32x4`, `f32x4` |
| PowerPC | `vreg` | `vsx` | `f32`, `f64`, `i64x2`, `f64x2` |
In addition to floats and `core::simd` types listed above, `core::arch` types and custom `#[repr(simd)]` types of the same size and type are also allowed. All allowed types and relevant target features are currently unstable.
r? `@Amanieu`
`@rustbot` label +O-PowerPC +A-inline-assembly
This is currently handled automatically by the fact that codegen doesn't visit
coverage statements in unused functions, but that will no longer be the case
when unused IDs are identified by a separate query instead.
Enable -Zshare-generics for inline(never) functions
This avoids inlining cross-crate generic items when possible that are
already marked inline(never), implying that the author is not intending
for the function to be inlined by callers. As such, having a local copy
may make it easier for LLVM to optimize but mostly just adds to binary
bloat and codegen time. In practice our benchmarks indicate this is
indeed a win for larger compilations, where the extra cost in dynamic
linking to these symbols is diminished compared to the advantages in
fewer copies that need optimizing in each binary.
It might also make sense it expand this with other heuristics (e.g.,
`#[cold]`) in the future, but this seems like a good starting point.
FWIW, I expect that doing cleanup in where we make the decision
what should/shouldn't be shared is also a good idea. Way too
much code needed to be tweaked to check this. But I'm hoping
to leave that for a follow-up PR rather than blocking this on it.
This reduces code sizes and better respects programmer intent when
marking inline(never). Previously such a marking was essentially ignored
for generic functions, as we'd still inline them in remote crates.
coverage: Store coverage source regions as `Span` until codegen
Historically, coverage spans were converted into line/column coordinates during the MIR instrumentation pass.
This PR moves that conversion step into codegen, so that coverage spans spend most of their time stored as `Span` instead.
In addition to being conceptually nicer, this also reduces the size of coverage mappings in MIR, because `Span` is smaller than 4x u32.
---
There should be no changes to coverage output.
Support input/output in vector registers of s390x inline assembly (under asm_experimental_reg feature)
This extends currently clobber-only vector registers (`vreg`) support to allow passing `#[repr(simd)]` types, floats (f32/f64/f128), and integers (i32/i64/i128) as input/output.
This is unstable and gated under new `#![feature(asm_experimental_reg)]` (tracking issue: https://github.com/rust-lang/rust/issues/133416). If the feature is not enabled, only clober is supported as before.
| Architecture | Register class | Target feature | Allowed types |
| ------------ | -------------- | -------------- | -------------- |
| s390x | `vreg` | `vector` | `i32`, `f32`, `i64`, `f64`, `i128`, `f128`, `i8x16`, `i16x8`, `i32x4`, `i64x2`, `f32x4`, `f64x2` |
This matches the list of types that are supported by the vector registers in LLVM:
https://github.com/llvm/llvm-project/blob/llvmorg-19.1.0/llvm/lib/Target/SystemZ/SystemZRegisterInfo.td#L301-L313
In addition to `core::simd` types and floats listed above, custom `#[repr(simd)]` types of the same size and type are also allowed. All allowed types other than i32/f32/i64/f64/i128, and relevant target features are currently unstable.
Currently there is no SIMD type for s390x in `core::arch`, but this is tracked in https://github.com/rust-lang/rust/issues/130869.
cc https://github.com/rust-lang/rust/issues/130869 about vector facility support in s390x
cc https://github.com/rust-lang/rust/issues/125398 & https://github.com/rust-lang/rust/issues/116909 about f128 support in asm
`@rustbot` label +O-SystemZ +A-inline-assembly
Fix asm goto with outputs and move it to a separate feature gate
Tracking issue: #119364
This PR addresses 3 aspects of asm goto with outputs:
* Codegen is fixed. My initial implementation has an oversight which cause the output to be only stored in fallthrough path, but not in label blocks.
* Outputs can now be used with `options(noreturn)` if a label block is given.
* All of this is moved to a new feature gate, because we likely want to stabilise `asm_goto` before asm goto with outputs.
`@rustbot` labels: +A-inline-assembly +F-asm
When labels are present, the `noreturn` option really means that asm block
won't fallthrough -- if labels are present, then outputs can still be
meaningfully used.
