[clang] [Clang][Concepts] Fix the constraint equivalence checking involving parameter packs (PR #102131)

[Younan Zhang via cfe-commits]

================
@@ -972,8 +972,30 @@ static const Expr *SubstituteConstraintExpressionWithoutSatisfaction(
   // equivalence.
   LocalInstantiationScope ScopeForParameters(S);
   if (auto *FD = DeclInfo.getDecl()->getAsFunction())
-    for (auto *PVD : FD->parameters())
-      ScopeForParameters.InstantiatedLocal(PVD, PVD);
+    for (auto *PVD : FD->parameters()) {
+      if (!PVD->isParameterPack()) {
+        ScopeForParameters.InstantiatedLocal(PVD, PVD);
+        continue;
+      }
+      // This is hacky: we're mapping the parameter pack to a size-of-1 argument
+      // to avoid building SubstTemplateTypeParmPackTypes for
+      // PackExpansionTypes. The SubstTemplateTypeParmPackType node would
+      // otherwise reference the AssociatedDecl of the template arguments, which
+      // is, in this case, the template declaration.
+      //
+      // However, as we're also calculating the redeclarations of the template,
+      // the canonical declarations thereof are actually themselves at the
+      // moment. So if we didn't expand these packs, we would end up with an
+      // incorrect profile difference because we will be profiling the
+      // canonical types!
+      //
+      // FIXME: Improve the "no-transform" machinery in FindInstantiatedDecl so
+      // that we can eliminate the Scope in the cases where the declarations are
+      // not necessarily instantiated. It would also benefit the noexcept
+      // specifier comparison.
----------------
zyn0217 wrote:

> I wonder if we could temporarily pretend that the declarations are equivalent while checking their constraint so that their canonical type matches.

I tried that approach in https://github.com/llvm/llvm-project/commit/0727e47124af564826a22a718a64a1da79e6131d. While it works, I don't think I like it quite much because it looks unnatural/dirty to me.

https://github.com/llvm/llvm-project/pull/102131