[llvm] [AMDGPU] Enable kernarg preloading by default on gfx940 (PR #110691)

[Austin Kerbow via llvm-commits]

================
@@ -1014,12 +1014,49 @@ struct AAAMDGPUNoAGPR
 
 const char AAAMDGPUNoAGPR::ID = 0;
 
+static unsigned getMaxNumPreloadArgs(const Function &F, const DataLayout &DL,
+                                     const TargetMachine &TM) {
+  const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
+  unsigned Offset = 0;
+  unsigned ArgsToPreload = 0;
+  for (const auto &Arg : F.args()) {
+    if (Arg.hasByRefAttr())
+      break;
+
+    Type *Ty = Arg.getType();
+    Align ArgAlign = DL.getABITypeAlign(Ty);
+    auto Size = DL.getTypeAllocSize(Ty);
----------------
kerbowa wrote:

What we are calculating is the size in memory, i.e. the start offset to the ending explicit offset. Isn't this method correct for anything that is not byref?

Do you think we should also consider the attributes that are inferred by the AMDGPUAttributor itself when marking things for preloading here? I.e. add inreg at the very end?

https://github.com/llvm/llvm-project/pull/110691