Improvements to clangd
Performance
clangd’s memory usage is significantly reduced on most Linux systems. In particular, memory usage should not increase dramatically over time.
The standard allocator on most systems is glibc’s ptmalloc2, and it creates disproportionately large heaps when handling clangd’s allocation patterns. By default, clangd will now periodically call malloc_trim to release free pages on glibc systems.
Users of other allocators (such as jemalloc or tcmalloc) on glibc systems can disable this using --malloc_trim=0 or the CMake flag -DCLANGD_MALLOC_TRIM=0.
Added the $/memoryUsage request: an LSP extension. This provides a breakdown of the memory clangd thinks it is using (excluding malloc overhead etc). The clangd VSCode extension supports showing the memory usage tree.
Parsing and selection
Improved navigation of broken code in C using Recovery AST. (This has been enabled for C++ since clangd 11).
Types are understood more often in broken code. (This is the first release where Recovery AST preserves speculated types).
Heuristic resolution for dependent names in templates.
Code completion
Higher priority for symbols that were already used in this file, and symbols from namespaces mentioned in this file. (Estimated 3% accuracy improvement)
Introduced a ranking algorithm trained on snippets from a large C++ codebase. Use the flag --ranking-model=decision_forest to try this (Estimated 6% accuracy improvement). This mode is likely to become the default in future.
Note: this is a generic model, not specialized for your code. clangd does not collect any data from your code to train code completion.
Signature help works with functions with template-dependent parameter types.
Go to definition
Selecting an auto or decltype keyword will attempt to navigate to a definition of the deduced type.
Improved handling of aliases: navigate to the underlying entity more often.
Better understanding of declaration vs definition for Objective-C classes and protocols.
Selecting a pure-virtual method shows its overrides.
Find references
Indexes are smarter about not returning stale references when code is deleted.
References in implementation files are always indexed, so results should be more complete.
Find-references on a virtual method shows references to overridden methods.
New navigation features
Call hierarchy (textDocument/callHierarchy) is supported. Only incoming calls are available.
Go to implementation (textDocument/implementation) is supported on abstract classes, and on virtual methods.
Symbol search (workspace/symbol) queries may be partially qualified. That is, typing b::Foo will match the symbol a:🅱️:c::Foo.
Refactoring
New refactoring: populate switch statement with cases. (This acts as a fix for the -Wswitch-enum warning).
Renaming templates is supported, and many other complex cases were fixed.
Attempting to rename to an invalid or conflicting name can produce an error message rather than broken code. (Not all cases are detected!)
The accuracy of many code actions has been improved.
Hover
Hovers for auto and decltype show the type in the same style as other hovers. this is also now supported.
Displayed type names are more consistent and idiomatic.
Semantic highlighting
Inactive preprocessor regions (#ifdef) are highlighted as comments.
clangd 12 is the last release with support for the non-standard textDocument/semanticHighlights notification. Clients sholud migrate to the textDocument/semanticTokens request added in LSP 3.16.
Remote index (alpha)
clangd can now connect to a remote index server instead of building a project index locally. This saves resources in large codebases that are slow to index.
The server program is clangd-index-server, and it consumes index files produced by clangd-indexer.
This feature requires clangd to be built with the CMake flag -DCLANGD_ENABLE_REMOTE=On, which requires GRPC libraries and is not enabled by default. Unofficial releases of the remote-index-enabled client and server tools are at https://github.com/clangd/clangd/releases
Large projects can deploy a shared server, and check in a .clangd file to enable it (in the Index.External section). We hope to provide such a server for llvm-project itself in the near future.
Configuration
Static and remote indexes can be configured in the Index.External section. Different static indexes can now be used for different files. (Obsoletes the flag --index-file).
Diagnostics can be filtered or suppressed in the Diagnostics section.
Clang-tidy checks can be enabled/disabled in the Diagnostics.ClangTidy section. (Obsoletes the flag --clang-tidy-checks).
The compilation database directory can be configured in the CompileFlags section. Different compilation databases can now be specified for different files. (Obsoletes the flag --compile-commands-dir).
Errors in loaded configuration files are published as LSP diagnostics, and so should be shown in your editor.
Full reference of configuration options
System integration
Changes to compile_commands.json and compile_flags.txt will take effect the next time a file is parsed, without restarting clangd.
clangd --check=<filename> can be run on the command-line to simulate opening a file without actually using an editor. This can be useful to reproduce crashes or aother problems.
Various fixes to handle filenames correctly (and case-insensitively) on windows.
If incoming LSP messages are malformed, the logs now contain details.
Miscellaneous
“Show AST” request (textDocument/ast) added as an LSP extension. This displays a simplified view of the clang AST for selected code. The clangd VSCode extension supports this.
clangd should no longer crash while loading old or corrupt index files.
The flags --index, --recovery-ast and -suggest-missing-includes have been retired. These features are now always enabled.
Too many stability and correctness fixes to mention.
895bccdc3c