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

[Matt Arsenault 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);
----------------
arsenm wrote:

Ideally kernel arguments would only use byref (except for these inreg arguments). Non-byref arguments are really byref, and will be lowered to byref. 

If you're making the values inreg, they should be treated more like ordinary calling convention register arguments. It's probably simplest to just filter out types that aren't trivially register types. That's the case for all implicit kernel arguments anyway. 

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