[clang] [llvm] [RISCV][GISel] Add support for quadruple-precision floating point ABIs (PR #195166)

[Sam Elliott via cfe-commits]

================
@@ -524,29 +544,49 @@ static bool CC_RISCV_Impl(unsigned ValNo, MVT ValVT, MVT LocVT,
   assert(PendingLocs.size() == PendingArgFlags.size() &&
          "PendingLocs and PendingArgFlags out of sync");
 
+  // If f128 cannot be passed in an FPR on RV32, it is passed according to the
+  // integer calling convention. Since it is larger than 2*XLEN, pass it by
+  // reference.
+  if (XLen == 32 && LocVT == MVT::f128) {
+    LocVT = XLenVT;
+    LocInfo = CCValAssign::Indirect;
+    if (MCRegister Reg = State.AllocateReg(ArgGPRs))
+      State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
+    else {
+      int64_t StackOffset = State.AllocateStack(4, Align(4));
----------------
lenary wrote:

This Allocation is slightly confusing me. Are we allocating for the pointer, or for the f128 value? If the former, use XLenInBytes, otherwise the values seem quite wrong.

https://github.com/llvm/llvm-project/pull/195166