[clang] [clang][AArch64] multilib: fix deduction of "-march=" option (PR #81474)

[Dominik Wójt via cfe-commits]

================
@@ -191,7 +191,11 @@ static void getAArch64MultilibFlags(const Driver &D,
   for (const auto &Ext : AArch64::Extensions)
     if (FeatureSet.contains(Ext.NegFeature))
       MArch.push_back(("no" + Ext.Name).str());
-  MArch.insert(MArch.begin(), ("-march=" + Triple.getArchName()).str());
+  StringRef ArchName;
+  for (const auto &ArchInfo : AArch64::ArchInfos)
+    if (FeatureSet.contains(ArchInfo->ArchFeature))
+      ArchName = ArchInfo->Name;
----------------
domin144 wrote:

There should be at least one arch feature in the list: there is an if-else chain with every branch setting one in getAArch64TargetFeatures. I agree it's better to add an assert just in case something changed.

For the case of multiple arch features: I think getAArch64TargetFeatures does not do that, but if it did, it should still be ok. "armv*-a" are ordered in AArch64::ArchInfos in a way that every arch feature represents a superset of previous one. This loop will select the greatest one, which is nice. I think it would be an error to have both "-a" and "-r" features, but checking this would need more code and I think this is not the right place for such check.

https://github.com/llvm/llvm-project/pull/81474