[llvm] [SystemZ] Handle IR struct arguments correctly. (PR #169583)

[Jonas Paulsson via llvm-commits]

================
@@ -2074,16 +2096,26 @@ SDValue SystemZTargetLowering::LowerFormalArguments(
                                    MachinePointerInfo()));
       // If the original argument was split (e.g. i128), we need
       // to load all parts of it here (using the same address).
-      unsigned ArgIndex = Ins[I].OrigArgIndex;
-      assert (Ins[I].PartOffset == 0);
-      while (I + 1 != E && Ins[I + 1].OrigArgIndex == ArgIndex) {
-        CCValAssign &PartVA = ArgLocs[I + 1];
-        unsigned PartOffset = Ins[I + 1].PartOffset;
-        SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, ArgValue,
-                                      DAG.getIntPtrConstant(PartOffset, DL));
-        InVals.push_back(DAG.getLoad(PartVA.getValVT(), DL, Chain, Address,
-                                     MachinePointerInfo()));
-        ++I;
+      MVT PartVT;
+      unsigned NumParts;
+      if (analyzeArgSplit(Ins, ArgLocs, I, PartVT, NumParts)) {
+        // TODO: It is strange that while LowerCallTo() sets the PartOffset
+        // relative to the first split part LowerArguments() sets the offset
+        // from the beginning of the struct. So with {i32, i256}, the
+        // PartOffset for the i256 parts are differently handled. Try to
+        // remove that difference and use PartOffset directly here (instead
+        // of SplitBaseOffs).
----------------
JonPsson1 wrote:

I tried removing it, and found only a single test case to fail, and this was in SystemZ, so it seems doable, if considering only the public targets, at least. Given that it is "weird", maybe it would be worth trying as a follow up?

https://github.com/llvm/llvm-project/pull/169583