[llvm] [AMDGPU] Move kernarg preload logic to AMDGPU Attributor (PR #123547)

[Matt Arsenault via llvm-commits]

================
@@ -1314,19 +1523,64 @@ struct AAAMDGPUNoAGPR
 
 const char AAAMDGPUNoAGPR::ID = 0;
 
-static void addPreloadKernArgHint(Function &F, TargetMachine &TM) {
-  const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
-  for (unsigned I = 0;
-       I < F.arg_size() &&
-       I < std::min(KernargPreloadCount.getValue(), ST.getMaxNumUserSGPRs());
-       ++I) {
-    Argument &Arg = *F.getArg(I);
-    // Check for incompatible attributes.
-    if (Arg.hasByRefAttr() || Arg.hasNestAttr())
-      break;
+static void markKernelArgsAsInreg(SetVector<Function *> &Functions,
+                                  TargetMachine &TM) {
+  SmallVector<Function *, 4> FunctionsToErase;
+  for (auto *F : Functions) {
+    const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(*F);
+    if (!ST.hasKernargPreload() ||
+        F->getCallingConv() != CallingConv::AMDGPU_KERNEL || F->arg_empty())
----------------
arsenm wrote:

Not really worth handling arg_empty a s a special case, it should fall out naturally and never happens in practice 

https://github.com/llvm/llvm-project/pull/123547