[clang] Allow constexpr-unknown values pre C++23

[tbaederr]

Let's check the CI.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Timm Baeder (tbaederr)

Changes

Let's check the CI.

---

Full diff: https://github.com/llvm/llvm-project/pull/129646.diff

1 Files Affected:

• (modified) clang/lib/AST/ExprConstant.cpp (+2-4)

diff --git a/clang/lib/AST/ExprConstant.cpp b/clang/lib/AST/ExprConstant.cpp
index 6ccb6e23f8d2f..9ea42b308f42c 100644
--- a/clang/lib/AST/ExprConstant.cpp
+++ b/clang/lib/AST/ExprConstant.cpp
@@ -3470,8 +3470,7 @@ static bool evaluateVarDeclInit(EvalInfo &Info, const Expr *E,
                                 unsigned Version, APValue *&Result) {
   // C++23 [expr.const]p8 If we have a reference type allow unknown references
   // and pointers.
-  bool AllowConstexprUnknown =
-      Info.getLangOpts().CPlusPlus23 && VD->getType()->isReferenceType();
+  bool AllowConstexprUnknown = VD->getType()->isReferenceType();
 
   APValue::LValueBase Base(VD, Frame ? Frame->Index : 0, Version);
 
@@ -8893,8 +8892,7 @@ bool LValueExprEvaluator::VisitDeclRefExpr(const DeclRefExpr *E) {
 bool LValueExprEvaluator::VisitVarDecl(const Expr *E, const VarDecl *VD) {
   // C++23 [expr.const]p8 If we have a reference type allow unknown references
   // and pointers.
-  bool AllowConstexprUnknown =
-      Info.getLangOpts().CPlusPlus23 && VD->getType()->isReferenceType();
+  bool AllowConstexprUnknown = VD->getType()->isReferenceType();
   // If we are within a lambda's call operator, check whether the 'VD' referred
   // to within 'E' actually represents a lambda-capture that maps to a
   // data-member/field within the closure object, and if so, evaluate to the

[tbaederr]

Looks like the differences are mainly that we don't emit the "unknown/invalid initializer" diagnostics anymore.

[cor3ntin]

@frederick-vs-ja AFAIK, this would be an extension, NOT a DR.

We probably want to be reasonably conservative here, because there are known issues with the implementation
(#127475)

[frederick-vs-ja]

@cor3ntin I see... My opinion is because N4916 said the P2280R4 was accepted as a DR and GCC is so treating it. Not sure whether the divergence would be desired.

[cor3ntin]

@cor3ntin I see... My opinion is because N4916 said the P2280R4 was accepted as a DR and GCC is so treating it. Not sure whether the divergence would be desired.

You are right, I updated https://clang.llvm.org/cxx_status.html#cxx23

I still think we should wait a bit for the codegen bugs to be fixed first

[shafik]

@cor3ntin I see... My opinion is because N4916 said the P2280R4 was accepted as a DR and GCC is so treating it. Not sure whether the divergence would be desired.

You are right, I updated https://clang.llvm.org/cxx_status.html#cxx23

I still think we should wait a bit for the codegen bugs to be fixed first

I am still recovering from being sick but I agree, we should probably hold off for a bit more.

[frederick-vs-ja]

The reduction of diagnostic messages in OpenMP tests seems like a good thing to me. IMO there's no initialization in #pragma omp parallel masked taskloop simd aligned(f:j) and its friends.
But other test failures probably indicated remaining bugs.

Edit: The behavioral change of clangd is also probably a good thing!

[efriedma-quic]

Reduced testcase for the microsoft-abi-member-pointers.cpp failure:

constinit int &b = b;

This should produce an error; clang -std=c++23 currently accepts it.

[efriedma-quic]

The other test changes look mostly okay, but worth investigating to see if we can improve the diagnostic notes for illegal uses of constexpr-unknown variables.