[clang] [llvm] [Clang] Emit DW_TAG_template_alias for template aliases (PR #87623)

Michael Buch via llvm-commits llvm-commits at lists.llvm.org
Wed Apr 17 06:36:20 PDT 2024


================
@@ -1332,6 +1332,54 @@ llvm::DIType *CGDebugInfo::CreateType(const TemplateSpecializationType *Ty,
   auto PP = getPrintingPolicy();
   Ty->getTemplateName().print(OS, PP, TemplateName::Qualified::None);
 
+  SourceLocation Loc = AliasDecl->getLocation();
+
+  if (CGM.getCodeGenOpts().DebugTemplateAlias) {
+    // TemplateSpecializationType doesn't know if its template args are
+    // being substituted into a parameter pack. We can find out if that's
+    // the case now by inspecting the TypeAliasTemplateDecl template
+    // parameters. Insert Ty's template args into SpecArgs, bundling args
+    // passed to a parameter pack into a TemplateArgument::Pack.
+    SmallVector<TemplateArgument> SpecArgs;
+    {
+      ArrayRef SubstArgs = Ty->template_arguments();
+      for (const NamedDecl *P : TD->getTemplateParameters()->asArray()) {
+        if (P->isParameterPack()) {
+          SpecArgs.push_back(TemplateArgument(SubstArgs));
+          break;
+        }
+        // Skip defaulted args.
+        if (SubstArgs.empty()) {
+          // If SubstArgs is now empty (we're taking from it each iteration) and
+          // this template parameter isn't a pack, then that should mean we're
+          // using default values for the remaining template parameters.
+          break;
----------------
Michael137 wrote:

Ah I see, thanks for clarifying! Just thought I'd point it out in case this wasn't expected. The DWARF spec has this to say about the template parameters of an alias:
```
The template alias entry has child entries describing the template actual parameters (see Section 2.23 on page 57).
```
Which to me implies that the parameters of an alias instantiation follow the same rules as those for class template instantiations.

Regarding name reconstruction, defaulted template arguments should have their `IsDefaulted` bit set, so when reconstructing the name we should be able to drop defaulted arguments? If that's too much of a hassle maybe not emitting them at all as you do now is also fine?

https://github.com/llvm/llvm-project/pull/87623


More information about the llvm-commits mailing list