- Add verbose logging of payloads
- Add public logging of request summaries
- fix non-logging of messages in request scopes (oops!)
- add test for public/non-public logging, extending pipeline_helper a bit.
We've accumulated quite a lot of duplication in the request handlers by now.
I should factor that out, but not in this patch...
Differential Revision: https://reviews.llvm.org/D90654
Avoid requiring an actual MemoryBuffer in ComputePreambleBounds, when
a MemoryBufferRef will do just fine.
Differential Revision: https://reviews.llvm.org/D90890
Previously a corrupted index shard could cause us to resize arrays to an
arbitrary int32. This tends to be a huge number, and can render the
system unresponsive.
Instead, cap this at the amount of data that might reasonably be read
(e.g. the #bytes in the file). If the specified length is more than that,
assume the data is corrupt.
Differential Revision: https://reviews.llvm.org/D91258
This is a try to improve clangd-indexer tool performance:
- avoid processing already processed files.
- use different mutexes for different entities (e.g. do not block insertion of references while symbols are inserted)
Results for LLVM project indexing:
- before: ~30 minutes
- after: ~10 minutes
Reviewed By: kadircet
Differential Revision: https://reviews.llvm.org/D91051
By iterating backwards over the globs we can exit the loop as soon as we find a match.
While we're here:
- Regex doesn't need to be mutable.
- We can reserve the amount of Globs needed ahead of time.
- Using a SmallVector with size 0 is slightly more space efficient than a std::vector.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D91033
https://reviews.llvm.org/D89670 changed the Ref structure, we need to
bump the version to invalidate all stored stale data, otherwise we will
get ` Error while reading shard: malformed or truncated refs` when
building the background index.
Differential Revision: https://reviews.llvm.org/D91131
* Even though remote index is still somewhat experimental, it can now be
used withing clangd itself: this should be the primary way of trying
it out
* Remove `protobuf-compiler` from list of needed Debian packages as it
`protobuf-compiler-grpc` already depends on it
* Bump recommended gRPC version to 1.32.3
We plan to eliminate error-prone and obsolete Clang-Rename API from Clangd. To
do that, we will introduce Decl canonicalization rules that will make renaming
code simpler and easier to maintain (D71880).
To ensure smooth transition to the new implementation, many Clang-Rename tests
will be adopted in Clangd to prevent any possible regressions. This patch is
the first in the chain of test migration patches. It improves existing tests
and adopts tests from Clang-Rename's alias, class and enum testing files.
Reviewed By: hokein
Differential Revision: https://reviews.llvm.org/D91102
With this patch, we reject the rename if the new name would conflict with
any other decls in the decl context of the renamed decl.
Differential Revision: https://reviews.llvm.org/D89790
There is not reason to check `std::make_unique<...>(..)` return value,
but `clangd::clang::loadIndex()` returns `nullptr` if an index file could not be loaded (e.g. incorrect version).
Reviewed By: kadircet
Differential Revision: https://reviews.llvm.org/D91049
The altera kernel name restriction check finds kernel files and include
directives whose filename is "kernel.cl", "Verilog.cl", or "VHDL.cl".
Such kernel file names cause the Altera Offline Compiler to generate
intermediate design files that have the same names as certain internal
files, which leads to a compilation error.
As per the "Guidelines for Naming the Kernel" section in the "Intel FPGA
SDK for OpenCL Pro Edition: Programming Guide."
This reverts the reversion from 43a38a6523.
If an enum has different names for the same constant, make sure only the first one declared gets added into the switch. Failing to do so results in a compiler error as 2 case labels can't represent the same value.
```
lang=c
enum Numbers{
One,
Un = One,
Two,
Deux = Two,
Three,
Trois = Three
};
// Old behaviour
switch (<Number>) {
case One:
case Un:
case Two:
case Duex:
case Three:
case Trois: break;
}
// New behaviour
switch (<Number>) {
case One:
case Two:
case Three: break;
}
```
Reviewed By: sammccall
Differential Revision: https://reviews.llvm.org/D90555
While here, clean up ParsedAST::build a bit:
- remove FIXMEs that were fixed long ago by ReplayPreamble
- remove redundant if, ClangTidyContext is not actually optional
Differential Revision: https://reviews.llvm.org/D90975
The config providers that look for configuration files currently take a pointer to a FileSystem in the constructor.
For some reason this isn't actually used when trying to read those configuration files, Essentially it just follows the behaviour of the real filesystem.
Using clang-tidy standalone this doesn't cause any issue.
But if its used as a library and the user wishes to use say an `InMemoryFileSystem` it will try to read the files from the disc instead.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D90992
Probably not essential as afaik only one check uses this field. but still good to have consistent behaviour.
Reviewed By: sammccall
Differential Revision: https://reviews.llvm.org/D90552
Add IgnoreMainLikeFunctions to the per file config. This can be extended for new options added to the check easily.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D90832
RemoteIndexClient implementations only depends on clangdSupport for
logging functionality and has no dependence on clangDeamon itself. This clears
out that link time dependency and enables depending on it in clangDeamon itself,
so that we can have other index implementations that makes use of the
RemoteIndex.
Differential Revision: https://reviews.llvm.org/D90746
SIG30-C. Call only asynchronous-safe functions within signal handlers
First version of this check, only minimal list of functions is allowed
("strictly conforming" case), for C only.
Differential Revision: https://reviews.llvm.org/D87449
This enables using the arrow operator to access members of the contained item.
```lang=c++
Located<std::string> X;
const char* CStr = X->c_str();
```
This is inline with how classes like `Optional` handle the arrow operator.
Reviewed By: kadircet
Differential Revision: https://reviews.llvm.org/D90682
This introduces a mechanism for providers to interpret paths specified
in a fragment either as absolute or relative to fragment location.
This information should be used during compile stage to handle blocks correctly.
Differential Revision: https://reviews.llvm.org/D90270
Let clang-tidy to read config from specified file.
Example:
$ clang-tidy --config-file=/some/path/myTidyConfig --list-checks --
...this will read config from '/some/path/myTidyConfig'.
ClangTidyMain.cpp reads ConfigFile into string and then assigned read data to 'Config' i.e. makes like '--config' code flow internally.
May speed-up tidy runtime since now it will just look-up <file-path>
instead of searching ".clang-tidy" in parent-dir(s).
Directly specifying config path helps setting build dependencies.
Thanks to @DmitryPolukhin for valuable suggestion. This patch now propose
change only in ClangTidyMain.cpp.
Reviewed By: DmitryPolukhin
Differential Revision: https://reviews.llvm.org/D89936
The index server has access to potentially-sensitive information, e.g. a
sequence of fuzzyFind requests reveals details about code completions in the
editor which in turn reveals details about the code being edited.
This information is necessary to provide the service, and our intention[1] is it
should never be retained beyond the scope of the request (e.g. not logged).
At the same time, some log messages should be exposed:
- server startup/index reloads etc that don't pertain to a user request
- basic request logs (method, latency, #results, error code) for monitoring
- errors while handling requests, without request-specific data
The -log=public design accommodates these by allowing three types of logs:
- those not associated with any user RPC request (via context-propagation)
- those explicitly tagged as [public] in the log line
- logging of format strings only, with no interpolated data (error level only)
[1] Specifically: Google is likely to run public instances of this server
for LLVM and potentially other projects, they will run in this configuration.
The details agreed in a Google-internal privacy review.
As clangd developers, we'd encourage others to use this configuration for public
instances too.
Differential Revision: https://reviews.llvm.org/D90526
- pass required=False to use_clang(), as we don't need it
- fix required=False (which was unused and rotted):
- make derived substitutions conditional on it
- add a feature so we can disable tests that need it
- conditionally disable our one test that depends on %resource_dir.
This doesn't seem right from first principles, but isn't a big deal.
Differential Revision: https://reviews.llvm.org/D90528
The vendor will be prefixed to the "clangd" and can be an arbitrary
string, so account for it in the test.
Reviewed By: sammccall
Differential Revision: https://reviews.llvm.org/D90517
The altera kernel name restriction check finds kernel files and include
directives whose filename is "kernel.cl", "Verilog.cl", or "VHDL.cl".
Such kernel file names cause the Altera Offline Compiler to generate
intermediate design files that have the same names as certain internal
files, which leads to a compilation error.
As per the "Guidelines for Naming the Kernel" section in the "Intel FPGA
SDK for OpenCL Pro Edition: Programming Guide."
Now that clang-tidy supports the --use-color command line option, it's
a better user experience to use --use-color in run-clang-tidy.py and
preserving the colored output.
We were default initializing SymbolIDs before, which would leave
indeterminate values in underlying std::array.
This patch updates the underlying data initalization to be value-init and adds a
way to check for validness of a SymbolID.
Differential Revision: https://reviews.llvm.org/D90397
Introduce a separate thread that will kill `clangd-index-server` after 10 seconds regardless. This helps shut down the test if the server hangs and `stderr.readline()` does not contain inititalizatiton message. It prevents "necessary" waiting delay for the server warm-up and only introduces additional delay if the test fails.
It also makes use of `subprocess.Popen.kill()` which is a portable way of handling process shutdown and avoids using signals.
Reviewed By: kadircet
Differential Revision: https://reviews.llvm.org/D90590
Google test matcher `DeclKind` uses `NamedDecl::getDeclKindName()` to compare its result with expected declaration name.
Both, returned value of this function and the expected kind name argument have type `const char *`, so this matcher effectively
compares two pointers instead of the respective strings.
The test was passing on most platforms because compilers mostly were able to coalesce these string literals.
Patch By: Ilya Golovenko
Reviewed By: hokein
Differential Revision: https://reviews.llvm.org/D90384
On Windows the --use-color option cannot be used for its originally
intended purpose of forcing color when piping stdout, since Windows
does not use ANSI escape codes by default. This change turns on ANSI
escape codes on Windows when forcing color to a non-displayed stdout
(e.g. piped).
Changed two references to developers as "he" or "him" to the more neutral "they".
Reviewed By: JDevlieghere, sylvestre.ledru
Differential Revision: https://reviews.llvm.org/D78807
After https://reviews.llvm.org/D80531 landed, a subtle bug was introduced where the test would fail if `LLVM_ENABLE_WERROR` was set. This just silences that error so the test case runs correctly, down the line it may be worth not enabling `-Werror` for clang-tidy tests even if the cmake flag is passed.
Make check_clang_tidy.py not just pass -format-style=none by default
but a full -config={}. Without this, with a build dir outside of
the llvm root dir and a .clang-tidy config further up that contains
CheckOptions:
- key: modernize-use-default-member-init.UseAssignment
value: 1
these tests would fail:
Clang Tools :: clang-tidy/checkers/cppcoreguidelines-prefer-member-initializer-modernize-use-default-member-init.cpp
Clang Tools :: clang-tidy/checkers/modernize-use-default-member-init-bitfield.cpp
Clang Tools :: clang-tidy/checkers/modernize-use-default-member-init.cpp
After this change, they pass fine, despite the unrelated
.clang-tidy file further up.
With every incremental change, one needs to check-in new model upstream.
This also significantly increases the size of the git repo with every
new model.
Testing and comparing the old and previous model is also not possible as
we run only a single model at any point.
One solution is to have a "staging" decision forest which can be
injected into clangd without pushing it to upstream. Compare the
performance of the staging model with the live model. After a couple of
enhancements have been done to staging model, we can then replace the
live model upstream with the staging model. This reduces upstream churn
and also allows us to compare models with current baseline model.
This is done by having a callback in CodeCompleteOptions which is called
only when we want to use a decision forest ranking model. This allows us
to inject different completion model internally.
Differential Revision: https://reviews.llvm.org/D90014
Remove the need to heap allocate a string for each style option lookup while reading or writing options.p
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D90244
This is an initial attempt to start using Syntax Trees in clangd while improving state of folding ranges feature and experimenting with Syntax Tree capabilities.
Reviewed By: sammccall
Differential Revision: https://reviews.llvm.org/D88553
Currently, this would not correctly associate a category with the related include if it was top-level (i.e. no slashes in the path). This ensures that we explicitly think about that case.
Reviewed By: gribozavr2
Differential Revision: https://reviews.llvm.org/D89608
Both `SymbolKind` and `indexSymbolKindToSymbolKind` support constructors and
separate them into a different category from regular methods.
Reviewed By: kadircet
Differential Revision: https://reviews.llvm.org/D89935
Previous attempt (15f6bad6d7) introduced
add_dependencies but unfortunately it does not actually add a dependency
between RemoteIndexProto and RemoteIndexServiceProto. This is likely due
to some requirements of it that clang_add_library violates.
As a workaround, we will link RemoteIndexProto library to
RemoteIndexServiceProto which is logical because the library can not be
without linking to RemoteIndexProto anyway.
TestWorkspace allows easily writing tests involving multiple
files that can have inclusion relationships between them.
BackgroundIndexTest.RelationsMultiFile is refactored to use
TestWorkspace, and moved to FileIndexTest as it no longer
depends on BackgroundIndex.
Differential Revision: https://reviews.llvm.org/D89297
In `ReplayPreamble::replay`, use `getFileRef` instead of `getFile`, and
then use that `FileEntryRef` later to avoid needing
`FileEntryRef::FileEntryRef`. The latter is going to become private to
`FileManager` in a later commit.
We only need to version these messages if they actually diverge.
Unlike the service, the namespace name isn't part of the wire format.
clangd-index-server was broken by 81e5f298c4
as the namespace names weren't updated there, this fixes it (by adding
them for the service, and not requiring them elsewhere).
This allows it to have a separate namespace (grpc versioned service) without
putting versioning info on all of the other protos (before we need it).
clang-index-server is still broken (from 81e5f298c4).
Differential Revision: https://reviews.llvm.org/D90031
`llvm::sys::path` is used a lot in the remote index marshalling code. We can save space by avoiding spelling it out explicitly for most functions and times.
Reviewed By: kadircet
Differential Revision: https://reviews.llvm.org/D90016
This allows us to check whether enum field is actually sent over the wire or missing.
Reviewed By: sammccall
Differential Revision: https://reviews.llvm.org/D89882
In memory VFS cannot handle aceesssing the same file with different paths.
This diff just stops using VFS for modulemap files.
Fixes PR47839
Differential Revision: https://reviews.llvm.org/D89886
And also introduce Protobuf package versioning, it will help to deal
with breaking changes. Inroducing package version itself is a breaking
change, clients and servers need to be updated.
Reviewed By: sammccall
Differential Revision: https://reviews.llvm.org/D89862
Replace `ContentCache::getRawBuffer` with `getBufferDataIfLoaded` and
`getBufferIfLoaded`, excising another accessor for the underlying
`MemoryBuffer*` in favour of `StringRef` and `MemoryBufferRef`.
Differential Revision: https://reviews.llvm.org/D89445
Since its call operator is const but can modify the state of its underlying
functor we cannot tell whether the copy is necessary or not.
This avoids false positives.
Reviewed-by: aaron.ballman, gribozavr2
Differential Revision: https://reviews.llvm.org/D89332
Nullability annotations are implmented using attributes; previusly
clangd would skip over AttributedTypeLoc since their location
points to the attribute instead of the modified type.
Also add some test cases for this.
Differential Revision: https://reviews.llvm.org/D89579
The patch adjusts the existing `llvm::DenseMap<unsigned, T>` and
`llvm::DenseSet<unsigned>` objects that store source locations, so
that they use `SourceLocation` directly instead of `unsigned`.
This patch relies on the `DenseMapInfo` trait added in D89719.
It also replaces the construction of `SourceLocation` objects from
the constants -1 and -2 with calls to the trait's methods `getEmptyKey`
and `getTombstoneKey` where appropriate.
Reviewed By: dexonsmith
Differential Revision: https://reviews.llvm.org/D69840
This change creates a `DenseMapInfo` trait specialization for the
SourceLocation class. The empty key, the tombstone key, and the hash
function are identical to `DenseMapInfo<unsigned>`, because we already
have hash maps that use raw the representation of `SourceLocation` as
a key.
The update of existing `DenseMap`s containing raw representation of
`SourceLocation`s will be done in a follow-up patch. As an example
the patch makes use of the new trait in one instance:
clang-tidy/google/UpgradeGoogletestCaseCheck.{h,cpp}
Reviewed By: dexonsmith
Differential Revision: https://reviews.llvm.org/D89719
This reverts commit 1b589f4d4d and relands the D89463
with the fix: update `MappingTraits<FileFilter>::validate()` in ClangTidyOptions.cpp to
match the new signature (change the return type to "std::string" from "StringRef").
Original commit message:
This:
Changes the return type of MappingTraits<T>>::validate to std::string
instead of StringRef. It allows to create more complex error messages.
It introduces std::vector<std::pair<StringRef, bool>> getEntries():
a new virtual method of Section, which is the base class for all sections.
It returns names of special section specific keys (e.g. "Entries") and flags that says if them exist in a YAML.
The code in validate() uses this list of entries descriptions to generalize validation.
This approach was discussed in the D89039 thread.
Differential revision: https://reviews.llvm.org/D89463
Without this patch 6 marshalling tests fail on Windows.
This patch contains the following changes:
- Allow paths with Windows slashes (convert to the POSIX style instead of assertion)
- Add support for URI with Windows path.
- Change the value of the second parameter of several `llvm::sys::path::convert_to_slash()` calls: we should use `windows` instead of `posix` to ensure UNIX slashes in the path.
- Port `RemoteMarshallingTest::IncludeHeaderURI` test to Windows.
Reviewed By: kbobyrev
Differential Revision: https://reviews.llvm.org/D89529
Update `Lexer` / `Lexer::Lexer` to use `MemoryBufferRef` instead of
`MemoryBuffer*`. Callers that were acquiring a `MemoryBuffer*` via
`SourceManager::getBuffer` were updated, such that if they checked
`Invalid` they use `getBufferOrNone` and otherwise `getBufferOrFake`.
Differential Revision: https://reviews.llvm.org/D89398
Added option `ScopedEnumConstant(Prefix|Case|Suffix)` to readability-identitied-naming.
This controls the style for constants in scoped enums, declared as enum (class|struct).
If this option is unspecified the EnumConstant style will be used instead.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D89407
new test: parsing and using compile_commands
new test: export fixes to yaml file
old test extended with CHECK-MESSAGES in order to ensure that they "fail as intended"
This was causing duplicate `symbols` components on the path as both the
edge from an index to filesymbols and filesymbols to symbolslabs were named
symbols.
Differential Revision: https://reviews.llvm.org/D89685
Performs a detailed profiling of clangd lsp server and conveys the
result to the client as a json object. It is of the form:
{
"_self": 0,
"_total": 8,
"child1": {
"_self": 4,
"_total": 4,
}
"child2": {
"_self": 2,
"_total": 4,
"child_deep": {
"_self": 2,
"_total": 2,
}
}
}
Differential Revision: https://reviews.llvm.org/D89277
Enables support for transforming loops of the form
```
for (auto I = Cont.rbegin(), E = Cont.rend(); I != E;++I)
```
This is done automatically in C++20 mode using `std::ranges::reverse_view` but there are options to specify a different function to reverse iterator over a container.
This is the first step, down the line I'd like to possibly extend this support for array based loops
```
for (unsigned I = Arr.size() - 1;I >=0;--I) Arr[I]...
```
Currently if you pass a reversing function with no header in the options it will just assume that the function exists, however as we have the ASTContext it may be as wise to check before applying, or at least lower the confidence level if we can't find it.
Reviewed By: alexfh
Differential Revision: https://reviews.llvm.org/D82089
Update IncludeSorter/IncludeInserter to support objective-c google style (part 1):
1) Correctly consider .mm/.m extensions
2) Correctly categorize category headers.
3) Add support for generated files to go in a separate section of imports
Reviewed By: alexfh, gribozavr2
Patch by Joe Turner.
Differential Revision: https://reviews.llvm.org/D89276
Update clang-tools-extra, clang/tools, clang/unittests to migrate from
`SourceManager::getBuffer`, which returns an always dereferenceable
`MemoryBuffer*`, to `getBufferOrNone` or `getBufferOrFake`, both of
which return a `MemoryBufferRef`, depending on whether the call site was
checking for validity of the buffer. No functionality change intended.
Differential Revision: https://reviews.llvm.org/D89416
Remove `ContentCache::getBuffer`, which always returned a
dereferenceable `MemoryBuffer*` and had a `bool*Invalid` out parameter,
and replace it with:
- `ContentCache::getBufferOrNone`, which returns
`Optional<MemoryBufferRef>`. This is the new API that consumers should
use. Later it could be renamed to `getBuffer`, but intentionally using
a different name to root out any unexpected callers.
- `ContentCache::getBufferPointer`, which returns `MemoryBuffer*` with
"optional" semantics. This is `private` to avoid growing callers and
`SourceManager` has temporarily been made a `friend` to access it.
Later paches will update the transitive callers to not need a raw
pointer, and eventually this will be deleted.
No functionality change intended here.
Differential Revision: https://reviews.llvm.org/D89348
so that we could start experiment for C.
Previously, these flags in clangd were only meaningful for C++. We need
to flip them for C, this patch repurpose these flags.
- if true, just set it.
- if false, just respect the value in clang.
this would allow us to keep flags on for C++, and optionally flip them on for C.
Differential Revision: https://reviews.llvm.org/D89233
Given the following VarTemplateDecl AST,
```
VarTemplateDecl col:26 X
|-TemplateTypeParmDecl typename depth 0 index 0
`-VarDecl X 'bool' cinit
`-CXXBoolLiteralExpr 'bool' true
```
previously, we returned the VarDecl as the top-level decl, which was not
correct, the top-level decl should be VarTemplateDecl.
Differential Revision: https://reviews.llvm.org/D89098
This will enable queries like "clangd::" to find symbols under clangd
namespace, without requiring full "clang::clangd::" qualification.
Since Fuzzyfind performs the search under all scopes and only boosts the symbols
from relevant namespaces, we might get symbols from non-matching namespaces.
This patch chooses to drop those as they clearly do not match the query.
Fixes https://github.com/clangd/clangd/issues/550.
Differential Revision: https://reviews.llvm.org/D88814
This patch introduces hoisting detection logic into prepare state with
a partial AST traversal of the enclosing function.
We had some complaints from the users about this code action being almost always
available but failing most of the time. Hopefully this should reduce that noise.
The latency/correctness tradeoff is a bunch of hand-waving, but at least today
we don't have any other actions that are available on selection of statements,
so when we get to do the traversal it is quite likely that all the other checks
will bail out early. But this is still up for discussion, I am happy to abandon
the patch if you believe this is not practical.
Differential Revision: https://reviews.llvm.org/D85354
Unreachable file distances are represented as
`std::numeric_limits<unsigned>::max()`.
The previous dataset recorded the signals as `signed int` capturing this default
value as `-1`.
A new dataset was regenerated and a new model is trained that
interprets this unreachable as the intended value.
Distribution of `SymbolScopeDistance`:
Value Normalised Frequency
0 46.6184
4294967295 29.5342
6 14.5666
4 6.4433
2 1.4534
8 0.5760
10 0.3581
....
Distribution of `FileProximityDistance`:
Value Normalised Frequency
4294967295 39.9378
12 5.1997
14 4.9828
15 4.4221
16 4.3820
13 4.2765
17 3.8957
11 3.6387
19 3.4799
18 3.4076
....
Differential Revision: https://reviews.llvm.org/D89035
If the NewName is provided, prepareRename would perform a name
validation.
The motivation is to allow our internal embeder implement the customized
"canRenameInto" functionality on top of prepareRename.
Differential Revision: https://reviews.llvm.org/D88881
Up until now, we relied on matching the filename.
This depends on unstable details of libstdc++ and doesn't work well on other
stdlibs. Also we'd like to remove it (see D88204).
Differential Revision: https://reviews.llvm.org/D88885
The default value is 1.3f, but it was cast to true, which is not a good
base for code completion score.
Differential Revision: https://reviews.llvm.org/D88970
With this patch, we don't treat `using ns::X` as a first-class declaration like `using Z = ns::Y`, reference to X that goes through this using-decl is considered a direct reference (without the Underlying bit).
Fix the workaround in https://reviews.llvm.org/D87225 and https://reviews.llvm.org/D74054.
Reviewed By: sammccall
Differential Revision: https://reviews.llvm.org/D88472
The protocol doesn't really incorporate ranking.
As with code completion, most clients respect what the server sends, but
VSCode re-ranks items, with predictable results.
See https://github.com/clangd/vscode-clangd/issues/81
There's no filterText field so we may be unable to construct a good workaround.
But expose the score so we may be able to do this on the client in future.
Differential Revision: https://reviews.llvm.org/D88844
Sam McCall