[llvm] Use Module level target-abi to assign target features for codegenerated functions. (PR #100833)

[Eric Christopher via llvm-commits]

================
@@ -387,6 +388,25 @@ Function *Function::Create(FunctionType *Ty, LinkageTypes Linkage,
   return Create(Ty, Linkage, M.getDataLayout().getProgramAddressSpace(), N, &M);
 }
 
+StringRef Function::getDefaultTargetFeatures(const StringRef TargetABI) {
+  Triple T(getParent()->getTargetTriple());
+  StringRef attr = "";
+  if (T.isRISCV64()) {
+    if (TargetABI.equals_insensitive("lp64d"))
+      attr = "+d";
+    else if (TargetABI.equals_insensitive("lp64f"))
+      attr = "+f";
+    else if (TargetABI.equals_insensitive("lp64q"))
+      attr = "+q";
+  } else if (T.isRISCV32() && TargetABI.contains("ilp32f")) {
+    attr = "+f";
+  } else if (T.isARM() || T.isThumb()) {
+    attr = "+thumb-mode";
----------------
echristo wrote:

+1 to Craig's comment. The module creator should be responsible for setting the ABI flags. I'm not certain why whoever is creating these functions isn't able to set subtarget features on a per function basis and I haven't seen a use case yet that makes me think otherwise. If you have a function being created during optimization then you have a TargetMachine available already. Also, is this trying to merge ilp32 and lp64 in the same file at a module merge time? Something feels a bit off there if so also.

If this is a JIT use case then you already have access to what the machine is and so it's not an issue there I don't believe either.

https://github.com/llvm/llvm-project/pull/100833