[Lldb-commits] [lldb] [lldb] Use TargetParser for AMDGPU architectures (PR #223192)

[satyanarayana reddy janga via lldb-commits]

================
@@ -1031,6 +847,19 @@ void ArchSpec::MergeFrom(const ArchSpec &other) {
     if (other.GetCore() != eCore_uknownMach64)
       UpdateCore();
   }
+  // AMDGPU models share one LLDB core, so merge their triple subarchitectures
+  // explicitly. Keep this ArchSpec's platform fields in the comparison triple
+  // because they are handled separately in this function.
+  if (GetTriple().getArch() == llvm::Triple::amdgpu &&
+      other.GetTriple().getArch() == llvm::Triple::amdgpu) {
+    llvm::Triple other_arch = GetTriple();
+    other_arch.setArch(other.GetTriple().getArch(),
+                       other.GetTriple().getSubArch());
+    if (GetTriple().isCompatibleWith(other_arch)) {
+      llvm::Triple merged(GetTriple().merge(other_arch));
+      GetTriple().setArch(merged.getArch(), merged.getSubArch());
+    }
+  }
----------------
satyajanga wrote:

I still had to keep the AMDGPU guard because `Triple::merge` also rewrites unrelated architectures; without it, LLDB regresses cases such as `arm64` → `aarch64` and `armv7em` → `arm` and the relevant tests are failing. 



https://github.com/llvm/llvm-project/pull/223192