WIP: refactor(cpp)#195
Open
wzekin wants to merge 8 commits into
Open
Conversation
Replace fragile C++ text heuristics with standard LSP requests where
possible.
Changes:
- lsp: add DocumentSymbol.Detail field, propagate Detail + SelectionRange
through flattenDocumentSymbols (previously stripped).
- cpp/lsp: new CppProvider implementing PrepareTypeHierarchy /
TypeHierarchySupertypes / TypeHierarchySubtypes / Hover /
Implementation / WorkspaceSearchSymbols. Registered for uniast.Cpp.
- collect: collectImpl now resolves C++ bases via typeHierarchy as the
primary path; falls back to text-level BaseClassRefs only when
typeHierarchy returns nothing (typical for unresolved external
templates). Saves ~50 lines of fragile Definition+alias-follow logic
per class.
- collect: extractCppCallSig prefers clangd-canonical signature from
sym.Detail ("void(int)") over source-scanned text ("void(int x)"),
which is more stable across multi-line parameter lists.
- collect: dedupCppFunction uses signature-aware key (cppBaseName on
full mid.Name) — overloads no longer collapse in Type.Methods, and
multi-line .cpp parameter formatting no longer breaks .h/.cpp merge.
Bug fixes carried in this commit:
1. (T,Text)-keyed positive dep cache (collectDepsForEntity) collapsed
distinct same-name calls onto one receiver. Cache reduced to negative
only.
2. BaseClassRefs preserves the full qualified base name instead of just
the last segment, so unresolved third_party::Provider and
other_pkg::Provider remain distinguishable.
3. IsUsingAlias kind-aware: using-declarations importing
functions/members (using Base::foo;) are no longer mis-routed
through the alias-redirect path.
4. dedupCppFunction uses bracket-depth-aware cppFnHasBody helper instead
of strings.Contains(content, "{"), so header decls with
lambda/braced-init default arguments no longer get falsely classified
as definitions.
5. baseSym/sym self-resolution detected by Location.Include containment
rather than short-name equality, so app::Provider extending
common::Provider isn't dropped as "resolved to self".
6. NVI synthesis dedup keyed by local signature (handle(int)) instead of
short name (handle), so overloads survive inheritance.
7. Sysroot (cstdlib) + build-time codegen (build_generated) modules emit
only edges, no nodes — scanner skips build64_release; addSymbol drops
sysroot/codegen syms before they enter c.syms; export skips body
emission for the same buckets.
Testdata + tests:
- testdata/cpp/{2_inheritance, 3_aliases, 4_inline_methods,
5_external_calls, 6_nvi_dispatch, 7_provider_chain,
8_overloaded_methods, 9_crossns_basename, 10_unresolved_basens,
11_using_decl, 13_methods_key_consistency} for regression coverage.
- 15-case unit test for cppFnHasBody covering declarations vs
definitions, lambda defaults, braced-init defaults, strings, comments.
- Integration tests in lang/parse_cpp_test.go for each fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…t LSP calls
Make the C++ collector strictly LSP-based: drop every text-heuristic
fallback (BaseClassRefs / parseBaseSpecifiers / declarationText /
IsTypedefSymbol / IsUsingAlias / AliasTargetTokenIndex / aliasRedirect),
delete the unresolved-base post-passes (resolveCppExternalImplements /
resolveCppUnresolvedBasesActive), and run the typeHierarchy/AST
extension as the only source of truth.
Tradeoff: dependent template bases (`class Foo : public Tmpl<X,Y>`) no
longer get Implements edges — clangd's typeHierarchy doesn't expose
supertypes for unbound template bases, and we no longer try to recover
them via source-text fallback. ModPath="external" placeholders are
eliminated; consumers see clean cross-module Implements only when
typeHierarchy could fully resolve them.
Notable changes:
- lsp.ASTNode: HasFunctionBody() method + ASTRoleDeclaration /
ASTKind{Compound,Typedef,TypeAlias,TypeAliasTemplate,Using,
UsingShadow,UsingPack} constants.
- LSPClient.Call: retry-go-based transient retry (3 attempts, 50ms
fixed delay) with a fast path that only enters the retry framework
when the first call fails. MethodNotFound (-32601) and context
cancel/deadline are terminal.
- IsJSONRPCMethodNotFound: exported from lang/lsp so collect doesn't
duplicate the helper.
- collect: cppFileAST caches the whole-file AST per URI; cppASTForSymbol
walks the cached tree to find the smallest containing declaration
(returns nil on miss, NOT the TranslationUnit root) and whole-file
range now covers the last line even when it has no trailing newline.
AST request failure caches nil to stop further hammering, but
transient failures get retried inside Call before that happens.
- export: dedupCppFunction uses AST-derived hasBody (Compound child of
the function declaration only — lambda defaults no longer trigger
false positives); typedef detection keys off ASTKindTypedef /
ASTKindTypeAlias; alias suppression keys off ASTKindTypeAlias{,Template}
for type aliases and only UsingDecl-family kinds for type-importing
using-declarations.
Tests:
- TestCpp_UnresolvedBaseNamespace and testdata/cpp/10_unresolved_basens
removed (they covered the deleted H4 fallback).
- TestCpp_Inheritance / TestCpp_ProviderChain / TestCpp_Aliases relax
expectations for the now-unreachable cases (dependent template bases,
using-alias redirect).
- cpp_ast_regression_test.go adds two regression tests:
cppASTForSymbol misses must NOT fall back to the TU root, and
whole-file AST range must include the last line without trailing
newline.
freq_service v46 verified: 33 modules / 119 pkgs / 462 fns / 341 types
(unchanged vs the prior typeHierarchy-only run), 0 ModPath="external"
placeholders, abcoder-side token-not-found dropped from 5 (v45) to 1
(remaining issue is unrelated URL-decoding in a system header path).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
wzekin
force-pushed
the
feat/cpp-typehierarchy-only
branch
from
92cca99 to
2be2e33
Compare
Modern LSP servers (gopls, recent clangd, pylsp Hoblovski/abc fork) suppress `semanticTokensProvider` from the initialize response unless the client declares semanticTokens support — at which point our strict capability check rejects the server and parsing aborts: failed to initialize LSP server: server did not provide SemanticTokensProvider Advertise the capability unconditionally so every server we currently target (and any new ones) returns the provider. The empty tokenTypes / tokenModifiers slices and `formats: ["relative"]` are the minimum LSP 3.17 client capability shape that turns the feature on. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
NewURI canonicalizes via filepath.EvalSymlinks so symbol URIs always carry the realpath form. ScannerByJavaIPC's normalizePath only filepath.Cleaned absolute file paths, so c.files was keyed by the symlinked path and downstream `c.files[symbol.Location.URI.File()]` lookups missed (e.g. /data00/<user>/... -> /home/<user>/...). The miss returned a nil *uniast.File, and JavaSpec.NameSpace panicked on the file.Package deref. Apply canonicalizePath inside normalizePath so the Scanner stores the same realpath shape as NewURI. Fixes TestJavaIPC_InterfaceKindAndImplements on systems where the repo is reached via a symlink (typical /home -> /data00 layout). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
wzekin
force-pushed
the
feat/cpp-typehierarchy-only
branch
from
6fef3cf to
f32fb42
Compare
Recent gopls (and likely other modern servers) interpret empty tokenTypes/tokenModifiers arrays as "client supports zero token types" and omit semanticTokensProvider from the initialize response — which trips our strict capability check at parse start: Failed to initialize go LSP client: server did not provide SemanticTokensProvider Advertise the standard LSP 3.17 token type + modifier sets so every server treats us as a fully-capable client. Also flip on dynamicRegistration and the `full.delta` request shape to match what real editors send. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This reverts commit e02b455.
This reverts commit 5ec9265.
initLSPClient rejected any server whose initialize response omitted `semanticTokensProvider` — including modern LSP servers (e.g. recent gopls in CI) that legitimately gate the feature on client-declared capabilities. The hard-fail kept abcoder's own Go path (native parser, never calls SemanticTokens) and Java path (tree-sitter) from starting against such servers; CI's `gopls@latest` panic surfaced as Failed to initialize go LSP client: server did not provide SemanticTokensProvider Treat absence as "this server doesn't support semantic tokens": store empty tokenTypes/tokenModifiers, return success, and let any caller that actually needs the feature surface a method-not-found error from the LSP call itself. The other capability checks (Definition, TypeDefinition, DocumentSymbol, References) stay required. Reverts and supersedes the earlier attempts at advertising semanticTokens client capability — both work for some gopls versions but break against others. Capability-relaxation is the robust fix. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What type of PR is this?
Check the PR title.
(Optional) Translate the PR title into Chinese.
(Optional) More detailed description for this PR(en: English/zh: Chinese).
en:
zh(optional):
(Optional) Which issue(s) this PR fixes:
(optional) The PR that updates user documentation: