Also allow __is_unsigned to be used as an identifier

[sberg]

Summary

...similar to 068730992cf08d7d7e82e7bb147d85f429fc3ddd "libstdc++ 4.4 uses __is_signed as an identifier [...]"

Starting with https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=5329e1a8e1480d536ff96283a6556e51112ba470 "libstdc++: Make chrono::hh_mm_ss more compact", libstdc++ 13 trunk used __is_signed as

static constexpr bool __is_unsigned
  = __and_v<is_integral<typename _Duration::rep>,
            is_unsigned<typename _Duration::rep>>;

in libstdc++-v3/include/std/chrono. Even though that got reverted with https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=cb363fd9f19eb791e1ee1eb0d5c61f5fdf21af32 "libstdc++: Change names that clash with Win32 or Clang", it might make sense to treat __is_unsigned the same as __is_signed here (and to also make the workaround hit for constexpr in addition to const).

[aaron.ballman]

I'm not opposed to this, but it's unfortunate that this hack is broader than it needs to be because we apply this parsing logic in all circumstances, not just within a system header. However, because we already behave this way for __is_signed, and it seems that parsing hacks do not follow the same pattern as Sema hacks (as in https://github.com/llvm/llvm-project/blob/main/clang/lib/Sema/SemaDeclCXX.cpp#L12523 or https://github.com/llvm/llvm-project/blob/main/clang/lib/Sema/SemaExceptionSpec.cpp#L44), this seems reasonable to me.

Do you know if folks are hitting problems here in practice, or is this speculative?

The changes should have a release note for the fix.

[sberg]

In D139847#4026966, @aaron.ballman wrote:

Do you know if folks are hitting problems here in practice, or is this speculative?

It had hit me when building LibreOffice against libstdc++ 13 trunk, so I started writing this hack. In parallel, libstdc++ 13 trunk got changed again to not use __is_unsigned as an identifier, so there should be no published releases of libstdc++ where this would be an issue. (And I had finished writing this hack anyway, just in case it might be useful. And then forgot about it... But maybe its best to just abandon it now.)

clang/lib/Parse/ParseDecl.cpp

case tok::annot_typename: {
 
       DS.SetRangeEnd(Tok.getAnnotationEndLoc());
       ConsumeAnnotationToken(); // The typename
 
       continue;
     }
 
     case tok::kw___is_signed:
+    case tok::kw___is_unsigned:
       // GNU libstdc++ 4.4 uses __is_signed as an identifier, but Clang
       // typically treats it as a trait. If we see __is_signed as it appears
       // in libstdc++, e.g.,
       //
       //   static const bool __is_signed;
       //
       // then treat __is_signed as an identifier rather than as a keyword.
+      //
+      // Similarly, libstdc++ 13 trunk at one point used __is_unsigned as an
+      // identifier as
+      //
+      //   static constexpr bool __is_unsigned;
+      //
       if (DS.getTypeSpecType() == TST_bool &&
-          DS.getTypeQualifiers() == DeclSpec::TQ_const &&
+          (DS.getTypeQualifiers() == DeclSpec::TQ_const ||
+           DS.getConstexprSpecifier() == ConstexprSpecKind::Constexpr) &&
           DS.getStorageClassSpec() == DeclSpec::SCS_static)
         TryKeywordIdentFallback(true);
 
       // We're done with the declaration-specifiers.
       goto DoneWithDeclSpec;
 
       // typedef-name
     case tok::kw___super:

clang/test/SemaCXX/libstdcxx_is_pod_hack.cpp

 // trait in Embarcadero's compiler but is used as an identifier in
 // libstdc++.
 struct test_is_signed {
   static const bool __is_signed = true; // expected-warning {{keyword '__is_signed' will be made available as an identifier}}
 };
 
 bool check_signed = test_is_signed::__is_signed;
 
+// Another, similar egregious hack for __is_unsigned, which is a type
+// trait in Embarcadero's compiler but is used as an identifier in
+// libstdc++.
+struct test_is_unsigned {
+  static constexpr bool __is_unsigned = true; // expected-warning {{keyword '__is_unsigned' will be made available as an identifier}}
+};
+
 template<bool B> struct must_be_true {};
 template<> struct must_be_true<false>;
 
 void foo() {
   bool b = __is_pod(int);
   must_be_true<__is_pod(int)> mbt;
 }