Repository navigation
Identifiers that are target-language keywords are emitted verbatim: a property named Type or Default generates fn type(...) (Rust) / int default() (C++) that fails to compile #106
Description
What's wrong
No generator checks a spelled identifier against the target language's reserved words. LanguageGeneratorBase and the language generators have no reserved-word set, and names only go through NameStyles.Spell (re-casing). The AST is language-neutral, so a name that is perfectly legal in one target becomes a keyword in another once re-cased. The clearest case is the getter name made up for a computed property, PropertyDeclaration.GetterName (Coder/Ast/PropertyDeclaration.cs ~141), used from StandardLanguageGenerator.cs ~393:
Repro (reproduced with MSTest, output checked with the real compilers)
A struct Box with new PropertyDeclaration("Type", "int") and new PropertyDeclaration("Default", "int"), each with a getter body return 1. Both are ordinary C# property names.
| Target | Generated | Compiler |
|---|---|---|
| Rust | pub fn type(&self) -> i32 |
rustc: expected identifier, found keyword 'type' |
| C++ | int default() const |
g++ -std=c++20: expected unqualified-id before 'default' |
| C | Box_type |
fine, because of the prefix |
The same applies to any other spelled identifier (parameter, local, field, function) that lands on a keyword after casing, e.g. match/impl/fn in Rust, range/func/type in Go, lambda/def in Python, delete/new in C++.
This is independent of #71: ktsu.CaseConverter would still produce type.
Suggested fix / acceptance criteria
- Give each generator a reserved-word set and a single
EscapeIdentifierhook applied wherever a name is spelled:- Rust: raw identifiers (
r#type). Note thatself/Self/crate/supercan't be raw, so suffix those. - C++/C/Go/Python/JS: append
_(e.g.default_,type_). - C#:
@default.
- Rust: raw identifiers (
- Apply it consistently at declaration and use sites so references still resolve.
- Add compile tests with a property named
Type/Defaultand a parameter named after a keyword in each target.
Activity
matt-edmondson commentedon Sep 27, 2026
Triage
Type,Default,Match,Range,Delete, …) produce Rust and C++ output that doesn't compile, and the same gap exists across every target language. For a multi-target code generator this is a systemic correctness hole, not an edge case, because no generator has a reserved-word list at all.LanguageGeneratorBase(a newEscapeIdentifierhook and a per-language reserved-word set),NameStyles.Spellcall sites, andPropertyDeclaration.GetterName. Owner: @matt-edmondsonr#, suffixingself/Self/crate/super) and C++, where the repro fails. Then roll it out to Go, Python, JS, C and C#, with a compile test per target covering a keyword-named property and parameter.Generated by Claude Code
matt-edmondson commentedon Oct 8, 2026
Triage
An automated issue sweep picked this up and set it aside before starting, because it is larger than one implement-and-test cycle:
.Nameand*Name(sites inCGenerator(27),CSharpGenerator(25),CppGenerator(20),GoGenerator(41),JavaScriptGenerator(21),PythonGenerator(24),RustGenerator(39) andStandardLanguageGenerator(10). There's no shared choke point today.NameStyles.Spellonly covers invented names (GetterName/SetterName), and declared names are written verbatim at each declaration and use site.default_,type_) renames the symbol in C/C++/Go/Python/JS, which theNamingStyleremarks say the AST deliberately never does. Raw identifiers (r#type,@default) don't rename. That trade-off is worth an explicit decision.Suggested split so each piece is actionable:
LanguageGeneratorBase.EscapeIdentifier(string)and a per-language reserved-word set, and route the Rust generator through it (r#, plus a suffix forself/Self/crate/super), with a rustc compile test.CFamilyGenerator), with a g++ compile test.Left unassigned so the next sweep can pick up a child issue once the split exists.
Generated by Claude Code