[clang] [Clang][Sema] Improve support for explicit specializations of constrained member functions & member function templates (PR #88963)

[Richard Smith via cfe-commits]

================
@@ -10339,24 +10339,53 @@ Sema::CheckMemberSpecialization(NamedDecl *Member, LookupResult &Previous) {
   if (Previous.empty()) {
     // Nowhere to look anyway.
   } else if (FunctionDecl *Function = dyn_cast<FunctionDecl>(Member)) {
+    SmallVector<FunctionDecl *> Candidates;
+    bool Ambiguous = false;
     for (LookupResult::iterator I = Previous.begin(), E = Previous.end();
            I != E; ++I) {
-      NamedDecl *D = (*I)->getUnderlyingDecl();
-      if (CXXMethodDecl *Method = dyn_cast<CXXMethodDecl>(D)) {
-        QualType Adjusted = Function->getType();
-        if (!hasExplicitCallingConv(Adjusted))
-          Adjusted = adjustCCAndNoReturn(Adjusted, Method->getType());
-        // This doesn't handle deduced return types, but both function
-        // declarations should be undeduced at this point.
-        if (Context.hasSameType(Adjusted, Method->getType())) {
-          FoundInstantiation = *I;
-          Instantiation = Method;
-          InstantiatedFrom = Method->getInstantiatedFromMemberFunction();
-          MSInfo = Method->getMemberSpecializationInfo();
-          break;
-        }
+      CXXMethodDecl *Method =
+          dyn_cast<CXXMethodDecl>((*I)->getUnderlyingDecl());
+      if (!Method)
+        continue;
+      QualType Adjusted = Function->getType();
+      if (!hasExplicitCallingConv(Adjusted))
+        Adjusted = adjustCCAndNoReturn(Adjusted, Method->getType());
+      // This doesn't handle deduced return types, but both function
+      // declarations should be undeduced at this point.
+      if (!Context.hasSameType(Adjusted, Method->getType()))
+        continue;
+      if (ConstraintSatisfaction Satisfaction;
+          Method->getTrailingRequiresClause() &&
+          (CheckFunctionConstraints(Method, Satisfaction,
+                                    /*UsageLoc=*/Member->getLocation(),
+                                    /*ForOverloadResolution=*/true) ||
+           !Satisfaction.IsSatisfied))
+        continue;
+      Candidates.push_back(Method);
+      FunctionDecl *MoreConstrained =
+          Instantiation ? getMoreConstrainedFunction(
+                              Method, cast<FunctionDecl>(Instantiation))
+                        : Method;
+      if (!MoreConstrained) {
+        Ambiguous = true;
+        continue;
+      }
+      if (MoreConstrained == Method) {
+        Ambiguous = false;
----------------
zygoloid wrote:

This doesn't look right to me. Suppose we see declaration `A` then `B` then `C`, and `A` and `B` are ambiguous and `C` is more constrained than `A`. This code will pick `C` as the found instantiation, even though `C` might not be more constrained than `B`.

I think we need to do the same two-pass dance we do in overload resolution: first do a walk through the declarations tracking the best so far, then go through them again and make sure the best so far is actually better than everything else.

https://github.com/llvm/llvm-project/pull/88963