[clang] [clang-format] Keep compound literals stable in macro bodies (PR #173771)

via cfe-commits cfe-commits at lists.llvm.org
Sun Dec 28 06:46:29 PST 2025


================
@@ -528,7 +553,16 @@ void UnwrappedLineParser::calculateBraceTypes(bool ExpectClassBody) {
           Tok->setBlockKind(BK_Block);
         }
       } else {
-        Tok->setBlockKind(BK_Unknown);
+        // In macro bodies we try to keep compound literal expressions like
----------------
Lane0218 wrote:

Thanks for the questions!

Regarding the non-macro path: the mis-formatting reported in #173583 seems to be specific to macro bodies because the macro replacement text often ends at `eof` (no trailing `;`, `)`, `,`, etc.), which makes brace-type inference more likely to misclassify the `{...}` as a block/function body. In “normal” code (non-macro), a compound literal like `&(type){...}` is typically embedded in a larger syntactic construct (e.g. followed by `;`, `)`, `,`), so the existing heuristics tend to classify it as a braced-init already. That’s why I kept the change restricted to `Line->InMacroBody` to minimize impact outside macros.

On using `MatchingParen`: I agree that using structured pairing information would be preferable. However, at this point (`UnwrappedLineParser::calculateBraceTypes`) we run before `TokenAnnotator::annotate()` (where `MatchingParen` is established), so it isn’t reliably available here. I’m happy to follow up with a more general refactor if we want to move this logic to a stage where `MatchingParen` is guaranteed, but that would be a larger change than this targeted regression fix.

https://github.com/llvm/llvm-project/pull/173771


More information about the cfe-commits mailing list