[llvm] x86: fix musttail sibcall miscompilation (PR #168956)

[Folkert de Vries via llvm-commits]

================
@@ -2146,6 +2201,75 @@ X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
   unsigned NumBytesToPush = NumBytes;
   unsigned NumBytesToPop = NumBytes;
 
+  SDValue StackPtr;
+  const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo();
+  int RetAddrSize = 8;
+
+  // If we are doing a tail-call, any byval arguments will be written to stack
+  // space which was used for incoming arguments. If any the values being used
+  // are incoming byval arguments to this function, then they might be
+  // overwritten by the stores of the outgoing arguments. To avoid this, we
+  // need to make a temporary copy of them in local stack space, then copy back
+  // to the argument area.
+  DenseMap<unsigned, SDValue> ByValTemporaries;
+  SDValue ByValTempChain;
+  if (isTailCall) {
+    SmallVector<SDValue, 8> ByValCopyChains;
+    for (const CCValAssign &VA : ArgLocs) {
+      unsigned ArgIdx = VA.getValNo();
+      SDValue Src = OutVals[ArgIdx];
+      ISD::ArgFlagsTy Flags = Outs[ArgIdx].Flags;
+
+      if (!Flags.isByVal())
+        continue;
+
+      auto PtrVT = getPointerTy(DAG.getDataLayout());
+
+      if (!StackPtr.getNode())
+        StackPtr =
+            DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(), PtrVT);
+
+      // Destination: where this byval should live in the callee’s frame
+      // after the tail call.
+      int32_t Offset = VA.getLocMemOffset() + FPDiff + RetAddrSize;
----------------
folkertdev wrote:

Does adding `RetAddrSize` here make sense? I needed it earlier so that `*rsp` does not get trashed, but now it doesn't seem to make a difference any more.

https://github.com/llvm/llvm-project/pull/168956