[llvm] AMDGPU: Add SubArchSpelling override to the TargetParser TableGen (PR #218500)

[Stanislav Mekhanoshin via llvm-commits]

================
@@ -612,19 +643,29 @@ static void emitAMDGPUSubArchNames(raw_ostream &OS, const RecordKeeper &RK,
       continue;
     SubArchEntry Entry;
     Entry.GPUName = E.Rec->getValueAsString("Name");
-    {
-      raw_svector_ostream SubArchOS(Entry.Suffix);
-      emitSubArchSuffix(SubArchOS, Entry.GPUName);
-    }
 
-    // A "gfxN-generic" target maps to the major-family subarch, so it takes the
-    // family triple name; a concrete GPU derives it from the ISA version.
     SmallString<16> TripleName;
     raw_svector_ostream TripleOS(TripleName);
-    if (isGenericTarget(E.Rec))
-      emitFamilySubArchTripleName(TripleOS, Entry.Suffix);
-    else
-      emitConcreteSubArchTripleName(TripleOS, E.Rec);
+
+    // An explicit subarch spelling supplies the enumerator suffix and triple
+    // name, rather than the name/ISA version.
+    if (std::optional<StringRef> Spelling = getSubArchSpelling(E.Rec)) {
+      raw_svector_ostream SubArchOS(Entry.Suffix);
+      emitSpellingSuffix(SubArchOS, *Spelling);
+      TripleOS << "amdgpu" << *Spelling;
+    } else {
+      {
----------------
rampitec wrote:

Formatting looks strange.

https://github.com/llvm/llvm-project/pull/218500