[clang] [Clang] Address feedback in PR183010 (PR #185608)

[Matheus Izvekov via cfe-commits]

================
@@ -1185,7 +1154,7 @@ static bool CheckConstraintSatisfaction(
     return false;
   }
 
-  if (TemplateArgsLists.isAnyArgInstantiationDependent(S.Context)) {
+  if (TemplateArgsLists.isAnyArgDependent(S.Context)) {
----------------
mizvekov wrote:

Yes it is, that goes back to repeating everything I said, but the point is that the change in #183010 is incorrect, the existing code skipping the satisfaction check when instantiation dependence is at play is the correct approach. Changing the check for plain dependence causes these problems.

I think this calls the patch into question. It didn't deliberately call the fact that it was attempting to fix a regression by introducing another regression.

I can't honestly say which regression is worse, and I think at this point that it's too risky to backport it.

https://github.com/llvm/llvm-project/pull/185608