[clang] [clang-format] Fix a bug in annotating angles containing FatArrow (PR #108671)

[Nico Weber via cfe-commits]

================
@@ -243,14 +244,16 @@ class AnnotatingParser {
       // operator that was misinterpreted because we are parsing template
       // parameters.
       // FIXME: This is getting out of hand, write a decent parser.
-      if (InExpr && !Line.startsWith(tok::kw_template) &&
+      if (InExpr && !SeenFatArrow && !Line.startsWith(tok::kw_template) &&
           Prev.is(TT_BinaryOperator)) {
         const auto Precedence = Prev.getPrecedence();
         if (Precedence > prec::Conditional && Precedence < prec::Relational)
           return false;
       }
       if (Prev.isOneOf(tok::question, tok::colon) && !Style.isProto())
         SeenTernaryOperator = true;
+      else if (Prev.is(TT_FatArrow))
----------------
nico wrote:

When we update clang-format in chromium, we format a chunk of the codebase with old and new clang-format versions and look at the diff. We also ran into this regression: https://chromium-review.googlesource.com/c/chromium/src/+/6174847/1..2/base/allocator/partition_allocator/src/partition_alloc/shim/allocator_shim_default_dispatch_to_partition_alloc.cc

We reported #123144 which pointed us to here.

While it's true that `a<b | c>(d)` could be `a < b | c > (d)`, using `|` (instead of `||`) with comparisons and using parenthesized one-element expressions seems less likely than this being a template function call with explicit template arguments that are a bitwise or, right? Maybe the heuristics could be tweaked to pick the template personality of <> when that seems "more likely"?

https://github.com/llvm/llvm-project/pull/108671