[clang] [clang] deprecate frelaxed-template-template-args, make it on by default (PR #89807)

[Matheus Izvekov via cfe-commits]

================
@@ -519,13 +571,45 @@ DeduceTemplateArguments(Sema &S, TemplateParameterList *TemplateParams,
     return TemplateDeductionResult::Success;
   }
 
-  if (TemplateTemplateParmDecl *TempParam
-        = dyn_cast<TemplateTemplateParmDecl>(ParamDecl)) {
+  if (auto *TempParam = dyn_cast<TemplateTemplateParmDecl>(ParamDecl)) {
     // If we're not deducing at this depth, there's nothing to deduce.
     if (TempParam->getDepth() != Info.getDeducedDepth())
       return TemplateDeductionResult::Success;
 
-    DeducedTemplateArgument NewDeduced(S.Context.getCanonicalTemplateName(Arg));
+    auto NewDeduced = DeducedTemplateArgument(Arg);
+    // Provisional resolution for CWG2398: If Arg is also a template template
+    // param, and it names a template specialization, then we deduce a
+    // synthesized template template parameter based on A, but using the TS's
+    // arguments as defaults.
+    if (auto *TempArg = dyn_cast_or_null<TemplateTemplateParmDecl>(
+            Arg.getAsTemplateDecl())) {
+      assert(Arg.getKind() == TemplateName::Template);
+      assert(!TempArg->isExpandedParameterPack());
+
+      TemplateParameterList *As = TempArg->getTemplateParameters();
+      if (DefaultArguments.size() != 0) {
+        assert(DefaultArguments.size() <= As->size());
+        SmallVector<NamedDecl *, 4> Params(As->size());
+        for (unsigned I = 0; I < DefaultArguments.size(); ++I)
+          Params[I] =
+              DeduceTemplateArguments(S, As->getParam(I), DefaultArguments[I]);
+        for (unsigned I = DefaultArguments.size(); I < As->size(); ++I)
+          Params[I] = As->getParam(I);
+        // FIXME: We could unique these, and also the parameters, but we don't
+        // expect programs to contain a large enough amount of these deductions
+        // for that to be worthwhile.
----------------
mizvekov wrote:

Yeah, doesn't affect identity. The results from deduction are assumed to be sugared, this was one of the changes I made to the deduction machinery back in 2021 or so.

Where identity is concerned, the types should be canonicalized, and we have a special step for canonicalizing template template parameters.

https://github.com/llvm/llvm-project/pull/89807