[RISCV] Support for varargs

[asb]

Summary

Includes support for expanding va_copy. Also adds support for using 'aligned' registers when necessary for vararg calls, and ensure the frame pointer always points to the bottom of the vararg spill region. This is necessary to ensure that the saved return address and stack pointer are always available at fixed known offsets of the frame pointer.

See https://github.com/riscv/riscv-elf-psabi-doc/blob/master/riscv-elf.md for calling convention documentation.

[sabuasal]

test/CodeGen/RISCV/vararg.ll

145

"where the first"

[asb]

Rebase patch and fix typo in test code comment.

[asb]

Rebasing patch. Ping!

[asb]

Update patch after r320884 (MachineFunction::getFunction now returns a reference).

[xiangzhai]

lib/Target/RISCV/RISCVISelLowering.cpp

446

RegIdx % 2 == 1 is an odd register, not an even, but the comment is "ensure it is assigned to an even register".

[asb]

lib/Target/RISCV/RISCVISelLowering.cpp

446

The comment and code are correct. If we would use an odd register, we 'waste' it by mark it as allocated to ensure that an even register will be used. I'll add an extra comment to make it a bit clearer - thanks for the feedback.

[xiangzhai]

lib/Target/RISCV/RISCVISelLowering.cpp

446

Thanks for your teaching!

[efriedma]

lib/Target/RISCV/RISCVISelLowering.cpp

701

Could you clarify what "ensure even-numbered registers are always 2*XLEN-aligned" means? The varargs save area is always right next to any arguments passed on the stack, so I'm not sure what exactly you need to align.

test/CodeGen/RISCV/vararg.ll

4

Could you generate 64-bit tests as well? (It probably isn't that different, but nice to have.)

298

This looks a little weird; why are we loading sp+24 into a scratch register?

[xiangzhai]

lib/Target/RISCV/RISCVISelLowering.cpp

701

Hi Eli,

I am reviewing Compiler Principle chapter 7.2 Stack Allocation, then consider about this question, because Strict Alignment? just like GCC https://gcc.gnu.org/ml/gcc-patches/2017-05/msg00011.html Please teach me, thanks a lot!

Regards,
Leslie Zhai

[efriedma]

lib/Target/RISCV/RISCVISelLowering.cpp

701

Yes, in general, some things need to be aligned, for a variety of reasons (including the issue you noted). I'm not questioning that part.

The question is more about the stack layout. Why do we need to allocate an extra object here to ensure alignment? What stack objects are we ensuring the alignment of?

[xiangzhai]

lib/Target/RISCV/RISCVISelLowering.cpp

701

Perhaps I need to review Compiler Principle's chapter 7.2.4 VarArg in stack and 7.3.5 Access Link more carefully :)

[asb]

Update code comments to address feedback.

lib/Target/RISCV/RISCVISelLowering.cpp

701

This is a requirement of the calling convention, see here. I believe the intention is to ensure that when the vararg registers are written to the stack they are done so with correct alignment, meaning va_arg can assume correct alignment regardless of whether a parameter was originally passed in an argument register or the stack. Assigning 'aligned' registers is no use if a2 is unaligned on the stack.

test/CodeGen/RISCV/vararg.ll

4

RV64I codegen isn't upstream yet.

[efriedma]

lib/Target/RISCV/RISCVISelLowering.cpp

701

I'm still not following; even-numbered registers are always 2*XLEN-aligned because the vararg save area must be adjacent to arguments passed on the stack. Like so:

stack arg

stack arg (XLEN*2 aligned)

saved a7

saved a6 (XLEN*2 aligned)

saved a5

saved a4 (XLEN*2 aligned)

saved a3

saved a2 (XLEN*2 aligned)

saved a1

saved a0 (XLEN*2 aligned)

This is the one and only layout, no matter how many argregs we save. If some of these arguments are fixed, you can skip saving them (and use the space for other stack objects, or just let the space extend beyond the stack pointer), but the layout doesn't fundamentally change. For example, if we save five registers (because there are three fixed registers), the layout is something like this:

stack arg

stack arg (XLEN*2 aligned)

saved a7

saved a6 (XLEN*2 aligned)

saved a5

saved a4 (XLEN*2 aligned)

saved a3 (va_start points here)

PADDING (XLEN*2 aligned) (frame pointer points here?)

(unused space)

(unused space) (XLEN*2 aligned)

Assuming my model is right, what purpose does the padding serve? The only effect I can see is that it changes the alignment of the frame pointer (which the comment doesn't mention at all).

[asb]

Updated to clarify a comment based on Eli's feedback (thanks!).

lib/Target/RISCV/RISCVISelLowering.cpp

701

Your analysis is correct, this code doesn't change the alignment of the data on the stack, only the aligment of the frame pointer (and hence the offsets codegen uses to access it). I've updated the comment accordingly. Thanks for your careful reading of this, I really appreciate it.

test/CodeGen/RISCV/vararg.ll

298

That's the frame pointer.

[efriedma]

LGTM

[kparzysz]

lib/Target/RISCV/RISCVISelLowering.cpp

719

You could use MachinePointerInfo::getFixedStack instead of an empty pointer info to get more accurate aliasing.

[kparzysz]

lib/Target/RISCV/RISCVISelLowering.cpp

719

Other than that, LGTM.

[asb]

lib/Target/RISCV/RISCVISelLowering.cpp

719

Good spot, making that change and applying now. Thanks for the reviews everyone.

lib/Target/RISCV/RISCVFrameLowering.cpp

 //===-- RISCVFrameLowering.cpp - RISCV Frame Information ------------------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file contains the RISCV implementation of TargetFrameLowering class.
 //
 //===----------------------------------------------------------------------===//
 
 #include "RISCVFrameLowering.h"
+#include "RISCVMachineFunctionInfo.h"
 #include "RISCVSubtarget.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 
 using namespace llvm;
 
void RISCVFrameLowering::emitPrologue(MachineFunction &MF,
 
   if (!hasFP(MF)) {
     report_fatal_error(
         "emitPrologue doesn't support framepointer-less functions");
   }
 
   MachineFrameInfo &MFI = MF.getFrameInfo();
   const RISCVInstrInfo *TII = STI.getInstrInfo();
+  auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>();
   MachineBasicBlock::iterator MBBI = MBB.begin();
 
   unsigned FPReg = getFPReg(STI);
   unsigned SPReg = getSPReg(STI);
 
   // Debug location must be unknown since the first debug location is used
   // to determine the end of the prologue.
   DebugLoc DL;
void RISCVFrameLowering::emitPrologue(MachineFunction &MF,
   // FIXME: assumes exactly one instruction is used to save each callee-saved
   // register.
   const std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo();
   std::advance(MBBI, CSI.size());
 
   // Generate new FP.
   BuildMI(MBB, MBBI, DL, TII->get(RISCV::ADDI), FPReg)
       .addReg(SPReg)
-      .addImm(StackSize)
+      .addImm(StackSize - RVFI->getVarArgsSaveSize())
       .setMIFlag(MachineInstr::FrameSetup);
 }
 
 void RISCVFrameLowering::emitEpilogue(MachineFunction &MF,
                                       MachineBasicBlock &MBB) const {
   if (!hasFP(MF)) {
     report_fatal_error(
         "emitEpilogue doesn't support framepointer-less functions");
   }
 
   MachineBasicBlock::iterator MBBI = MBB.getLastNonDebugInstr();
   const RISCVInstrInfo *TII = STI.getInstrInfo();
   const RISCVRegisterInfo *RI = STI.getRegisterInfo();
   MachineFrameInfo &MFI = MF.getFrameInfo();
+  auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>();
   DebugLoc DL = MBBI->getDebugLoc();
   unsigned FPReg = getFPReg(STI);
   unsigned SPReg = getSPReg(STI);
 
   // Skip to before the restores of callee-saved registers
   // FIXME: assumes exactly one instruction is used to restore each
   // callee-saved register.
   MachineBasicBlock::iterator LastFrameDestroy = MBBI;
   std::advance(LastFrameDestroy, -MFI.getCalleeSavedInfo().size());
 
   uint64_t StackSize = MFI.getStackSize();
 
   // Restore the stack pointer using the value of the frame pointer. Only
   // necessary if the stack pointer was modified, meaning the stack size is
   // unknown.
   if (RI->needsStackRealignment(MF) || MFI.hasVarSizedObjects()) {
     BuildMI(MBB, LastFrameDestroy, DL, TII->get(RISCV::ADDI), SPReg)
         .addReg(FPReg)
-        .addImm(-StackSize)
+        .addImm(-StackSize + RVFI->getVarArgsSaveSize())
         .setMIFlag(MachineInstr::FrameDestroy);
   }
 
   if (!isInt<12>(StackSize)) {
     report_fatal_error("Stack adjustment won't fit in signed 12-bit immediate");
   }
 
   // Deallocate stack
   BuildMI(MBB, MBBI, DL, TII->get(RISCV::ADDI), SPReg)
       .addReg(SPReg)
       .addImm(StackSize)
       .setMIFlag(MachineInstr::FrameDestroy);
 }
 
 int RISCVFrameLowering::getFrameIndexReference(const MachineFunction &MF,
                                                int FI,
                                                unsigned &FrameReg) const {
   const MachineFrameInfo &MFI = MF.getFrameInfo();
   const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo();
+  const auto *RVFI = MF.getInfo<RISCVMachineFunctionInfo>();
 
   // Callee-saved registers should be referenced relative to the stack
   // pointer (positive offset), otherwise use the frame pointer (negative
   // offset).
   const std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo();
   int MinCSFI = 0;
   int MaxCSFI = -1;
 
   int Offset = MFI.getObjectOffset(FI) - getOffsetOfLocalArea() +
                MFI.getOffsetAdjustment();
 
   if (CSI.size()) {
     MinCSFI = CSI[0].getFrameIdx();
     MaxCSFI = CSI[CSI.size() - 1].getFrameIdx();
   }
 
-  FrameReg = RI->getFrameRegister(MF);
   if (FI >= MinCSFI && FI <= MaxCSFI) {
     FrameReg = RISCV::X2;
     Offset += MF.getFrameInfo().getStackSize();
+  } else {
+    FrameReg = RI->getFrameRegister(MF);
+    assert(hasFP(MF) && "Offset calculation incorrect if !hasFP");
+    Offset += RVFI->getVarArgsSaveSize();
   }
   return Offset;
 }
 
 void RISCVFrameLowering::determineCalleeSaves(MachineFunction &MF,
                                               BitVector &SavedRegs,
                                               RegScavenger *RS) const {
   TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS);
   // TODO: Once frame pointer elimination is implemented, don't
   // unconditionally spill the frame pointer and return address.
   SavedRegs.set(RISCV::X1);
   SavedRegs.set(RISCV::X8);
 }

lib/Target/RISCV/RISCVISelLowering.h

EmitInstrWithCustomInserter(MachineInstr &MI,
                               MachineBasicBlock *BB) const override;
 
 private:
   void analyzeInputArgs(MachineFunction &MF, CCState &CCInfo,
                         const SmallVectorImpl<ISD::InputArg> &Ins,
                         bool IsRet) const;
   void analyzeOutputArgs(MachineFunction &MF, CCState &CCInfo,
                          const SmallVectorImpl<ISD::OutputArg> &Outs,
-                         bool IsRet) const;
+                         bool IsRet, CallLoweringInfo *CLI) const;
   // Lower incoming arguments, copy physregs into vregs
   SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
                                bool IsVarArg,
                                const SmallVectorImpl<ISD::InputArg> &Ins,
                                const SDLoc &DL, SelectionDAG &DAG,
                                SmallVectorImpl<SDValue> &InVals) const override;
   bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
                       bool IsVarArg,
                       const SmallVectorImpl<ISD::OutputArg> &Outs,
                       LLVMContext &Context) const override;
   SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
                       const SmallVectorImpl<ISD::OutputArg> &Outs,
                       const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
                       SelectionDAG &DAG) const override;
   SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
                     SmallVectorImpl<SDValue> &InVals) const override;
   bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
                                          Type *Ty) const override {
     return true;
   }
   SDValue lowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerSELECT(SDValue Op, SelectionDAG &DAG) const;
+  SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const;
 };
 }
 
 #endif

lib/Target/RISCV/RISCVISelLowering.cpp

 //===-- RISCVISelLowering.cpp - RISCV DAG Lowering Implementation  --------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file defines the interfaces that RISCV uses to lower LLVM code into a
 // selection DAG.
 //
 //===----------------------------------------------------------------------===//
 
 #include "RISCVISelLowering.h"
 #include "RISCV.h"
+#include "RISCVMachineFunctionInfo.h"
 #include "RISCVRegisterInfo.h"
 #include "RISCVSubtarget.h"
 #include "RISCVTargetMachine.h"
 #include "llvm/CodeGen/CallingConvLower.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
   setOperationAction(ISD::BR_JT, MVT::Other, Expand);
   setOperationAction(ISD::BR_CC, XLenVT, Expand);
   setOperationAction(ISD::SELECT, XLenVT, Custom);
   setOperationAction(ISD::SELECT_CC, XLenVT, Expand);
 
   setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
   setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
 
+  setOperationAction(ISD::VASTART, MVT::Other, Custom);
+  setOperationAction(ISD::VAARG, MVT::Other, Expand);
+  setOperationAction(ISD::VACOPY, MVT::Other, Expand);
+  setOperationAction(ISD::VAEND, MVT::Other, Expand);
+
   for (auto VT : {MVT::i1, MVT::i8, MVT::i16})
     setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
 
   setOperationAction(ISD::ADDC, XLenVT, Expand);
   setOperationAction(ISD::ADDE, XLenVT, Expand);
   setOperationAction(ISD::SUBC, XLenVT, Expand);
   setOperationAction(ISD::SUBE, XLenVT, Expand);
 
SDValue RISCVTargetLowering::LowerOperation(SDValue Op,
   default:
     report_fatal_error("unimplemented operand");
   case ISD::GlobalAddress:
     return lowerGlobalAddress(Op, DAG);
   case ISD::BlockAddress:
     return lowerBlockAddress(Op, DAG);
   case ISD::SELECT:
     return lowerSELECT(Op, DAG);
+  case ISD::VASTART:
+    return lowerVASTART(Op, DAG);
   }
 }
 
 SDValue RISCVTargetLowering::lowerGlobalAddress(SDValue Op,
                                                 SelectionDAG &DAG) const {
   SDLoc DL(Op);
   EVT Ty = Op.getValueType();
   GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
SDValue RISCVTargetLowering::lowerSELECT(SDValue Op, SelectionDAG &DAG) const {
   SDValue SetNE = DAG.getConstant(ISD::SETNE, DL, XLenVT);
 
   SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
   SDValue Ops[] = {CondV, Zero, SetNE, TrueV, FalseV};
 
   return DAG.getNode(RISCVISD::SELECT_CC, DL, VTs, Ops);
 }
 
+SDValue RISCVTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
+  MachineFunction &MF = DAG.getMachineFunction();
+  RISCVMachineFunctionInfo *FuncInfo = MF.getInfo<RISCVMachineFunctionInfo>();
+
+  SDLoc DL(Op);
+  SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
+                                 getPointerTy(MF.getDataLayout()));
+
+  // vastart just stores the address of the VarArgsFrameIndex slot into the
+  // memory location argument.
+  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
+  return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
+                      MachinePointerInfo(SV));
+}
+
 MachineBasicBlock *
 RISCVTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
                                                  MachineBasicBlock *BB) const {
   const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
   DebugLoc DL = MI.getDebugLoc();
 
   assert(MI.getOpcode() == RISCV::Select_GPR_Using_CC_GPR &&
          "Unexpected instr type to insert");
static bool CC_RISCVAssign2XLen(unsigned XLen, CCState &State, CCValAssign VA1,
   }
 
   return false;
 }
 
 // Implements the RISC-V calling convention. Returns true upon failure.
 static bool CC_RISCV(const DataLayout &DL, unsigned ValNo, MVT ValVT, MVT LocVT,
                      CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
-                     CCState &State, bool IsFixed, bool IsRet) {
+                     CCState &State, bool IsFixed, bool IsRet, Type *OrigTy) {
   unsigned XLen = DL.getLargestLegalIntTypeSizeInBits();
   assert(XLen == 32 || XLen == 64);
   MVT XLenVT = XLen == 32 ? MVT::i32 : MVT::i64;
   assert(ValVT == XLenVT && "Unexpected ValVT");
   assert(LocVT == XLenVT && "Unexpected LocVT");
-  assert(IsFixed && "Vararg support not yet implemented");
 
   // Any return value split in to more than two values can't be returned
   // directly.
   if (IsRet && ValNo > 1)
     return true;
 
+  // If this is a variadic argument, ensure it is assigned an even register
+  // if it has 8-byte alignment (RV32) or 16-byte alignment (RV64)
+  // An aligned register should be used regardless of
+  // whether the original argument is 'split' or not. The argument will not be
+  // passed by registers if the original type is larger than 2*XLEN, so don't
+  // bother aligning for that case.
+  unsigned TwoXLenInBytes = (2 * XLen) / 8;
+  if (!IsFixed && ArgFlags.getOrigAlign() == TwoXLenInBytes &&
+      DL.getTypeAllocSize(OrigTy) == TwoXLenInBytes) {
+    unsigned RegIdx = State.getFirstUnallocated(ArgGPRs);
+    if (RegIdx != array_lengthof(ArgGPRs) && RegIdx % 2 == 1) {
+      State.AllocateReg(ArgGPRs);
+    }
+  }
+
   SmallVectorImpl<CCValAssign> &PendingLocs = State.getPendingLocs();
   SmallVectorImpl<ISD::ArgFlagsTy> &PendingArgFlags =
       State.getPendingArgFlags();
 
   assert(PendingLocs.size() == PendingArgFlags.size() &&
          "PendingLocs and PendingArgFlags out of sync");
 
   // Split arguments might be passed indirectly, so keep track of the pending
static bool CC_RISCV(const DataLayout &DL, unsigned ValNo, MVT ValVT, MVT LocVT,
   }
   return false;
 }
 
 void RISCVTargetLowering::analyzeInputArgs(
     MachineFunction &MF, CCState &CCInfo,
     const SmallVectorImpl<ISD::InputArg> &Ins, bool IsRet) const {
   unsigned NumArgs = Ins.size();
+  FunctionType *FType = MF.getFunction()->getFunctionType();
 
   for (unsigned i = 0; i != NumArgs; ++i) {
     MVT ArgVT = Ins[i].VT;
     ISD::ArgFlagsTy ArgFlags = Ins[i].Flags;
 
+    Type *ArgTy = nullptr;
+    if (IsRet)
+      ArgTy = FType->getReturnType();
+    else if (Ins[i].isOrigArg())
+      ArgTy = FType->getParamType(Ins[i].getOrigArgIndex());
+
     if (CC_RISCV(MF.getDataLayout(), i, ArgVT, ArgVT, CCValAssign::Full,
-                 ArgFlags, CCInfo, /*IsRet=*/true, IsRet)) {
+                 ArgFlags, CCInfo, /*IsRet=*/true, IsRet, ArgTy)) {
       DEBUG(dbgs() << "InputArg #" << i << " has unhandled type "
                    << EVT(ArgVT).getEVTString() << '\n');
       llvm_unreachable(nullptr);
     }
   }
 }
 
 void RISCVTargetLowering::analyzeOutputArgs(
     MachineFunction &MF, CCState &CCInfo,
-    const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsRet) const {
+    const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsRet,
+    CallLoweringInfo *CLI) const {
   unsigned NumArgs = Outs.size();
 
   for (unsigned i = 0; i != NumArgs; i++) {
     MVT ArgVT = Outs[i].VT;
     ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
+    Type *OrigTy = CLI ? CLI->getArgs()[Outs[i].OrigArgIndex].Ty : nullptr;
 
     if (CC_RISCV(MF.getDataLayout(), i, ArgVT, ArgVT, CCValAssign::Full,
-                 ArgFlags, CCInfo, Outs[i].IsFixed, IsRet)) {
+                 ArgFlags, CCInfo, Outs[i].IsFixed, IsRet, OrigTy)) {
       DEBUG(dbgs() << "OutputArg #" << i << " has unhandled type "
                    << EVT(ArgVT).getEVTString() << "\n");
       llvm_unreachable(nullptr);
     }
   }
 }
 
 // The caller is responsible for loading the full value if the argument is
SDValue RISCVTargetLowering::LowerFormalArguments(
   case CallingConv::C:
   case CallingConv::Fast:
     break;
   }
 
   MachineFunction &MF = DAG.getMachineFunction();
   MVT XLenVT = Subtarget.getXLenVT();
   EVT PtrVT = getPointerTy(DAG.getDataLayout());
-
-  if (IsVarArg)
-    report_fatal_error("VarArg not supported");
+  unsigned XLen = DAG.getDataLayout().getLargestLegalIntTypeSizeInBits();
+  unsigned XLenInBytes = XLen / 8;
+  // Used with vargs to acumulate store chains.
+  std::vector<SDValue> OutChains;
 
   // Assign locations to all of the incoming arguments.
   SmallVector<CCValAssign, 16> ArgLocs;
   CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
   analyzeInputArgs(MF, CCInfo, Ins, /*IsRet=*/false);
 
   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
     CCValAssign &VA = ArgLocs[i];
if (VA.getLocInfo() == CCValAssign::Indirect) {
         InVals.push_back(DAG.getLoad(PartVA.getValVT(), DL, Chain, Address,
                                      MachinePointerInfo()));
         ++i;
       }
       continue;
     }
     InVals.push_back(ArgValue);
   }
+
+  if (IsVarArg) {
+    ArrayRef<MCPhysReg> ArgRegs = makeArrayRef(ArgGPRs);
+    unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs);
+    const TargetRegisterClass *RC = &RISCV::GPRRegClass;
+    MachineFrameInfo &MFI = MF.getFrameInfo();
+    MachineRegisterInfo &RegInfo = MF.getRegInfo();
+    RISCVMachineFunctionInfo *RVFI = MF.getInfo<RISCVMachineFunctionInfo>();
+
+    // Offset of the first variable argument from stack pointer, and size of
+    // the vararg save area. For now, the varargs save area is either zero or
+    // large enough to hold a0-a7.
+    int VaArgOffset, VarArgsSaveSize;
+
+    // If all registers are allocated, then all varargs must be passed on the
+    // stack and we don't need to save any argregs.
+    if (ArgRegs.size() == Idx) {
+      VaArgOffset = CCInfo.getNextStackOffset();
+      VarArgsSaveSize = 0;
+    } else {
+      VarArgsSaveSize = XLenInBytes * (ArgRegs.size() - Idx);
+      VaArgOffset = -VarArgsSaveSize;
+    }
+
+    // Record the frame index of the first variable argument
+    // which is a value necessary to VASTART.
+    int FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset, true);
+    RVFI->setVarArgsFrameIndex(FI);
+
+    // If saving an odd-number of registers, create an extra stack slot to
+    // ensure even-numbered registers are always 2*XLEN-aligned.
+    if (Idx % 2) {
+      FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset - (int)XLenInBytes,
+                                 true);
+      VarArgsSaveSize += XLenInBytes;
+    }
+
+    // Copy the integer registers that may have been used for passing varargs
+    // to the vararg save area.
+    for (unsigned I = Idx; I < ArgRegs.size();
+         ++I, VaArgOffset += XLenInBytes) {
+      const unsigned Reg = RegInfo.createVirtualRegister(RC);
+      RegInfo.addLiveIn(ArgRegs[I], Reg);
+      SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, XLenVT);
+      FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset, true);
+      SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
+      SDValue Store =
+          DAG.getStore(Chain, DL, ArgValue, PtrOff, MachinePointerInfo());
+      cast<StoreSDNode>(Store.getNode())
+          ->getMemOperand()
+          ->setValue((Value *)nullptr);
+      OutChains.push_back(Store);
+    }
+    RVFI->setVarArgsSaveSize(VarArgsSaveSize);
+  }
+
+  // All stores are grouped in one node to allow the matching between
+  // the size of Ins and InVals. This only happens for vararg functions.
+  if (!OutChains.empty()) {
+    OutChains.push_back(Chain);
+    Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);
+  }
+
   return Chain;
 }
 
 // Lower a call to a callseq_start + CALL + callseq_end chain, and add input
 // and output parameter nodes.
 SDValue RISCVTargetLowering::LowerCall(CallLoweringInfo &CLI,
                                        SmallVectorImpl<SDValue> &InVals) const {
   SelectionDAG &DAG = CLI.DAG;
   SDLoc &DL = CLI.DL;
   SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
   SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
   SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
   SDValue Chain = CLI.Chain;
   SDValue Callee = CLI.Callee;
   CLI.IsTailCall = false;
   CallingConv::ID CallConv = CLI.CallConv;
   bool IsVarArg = CLI.IsVarArg;
   EVT PtrVT = getPointerTy(DAG.getDataLayout());
   MVT XLenVT = Subtarget.getXLenVT();
 
-  if (IsVarArg) {
-    report_fatal_error("LowerCall with varargs not implemented");
-  }
-
   MachineFunction &MF = DAG.getMachineFunction();
 
   // Analyze the operands of the call, assigning locations to each operand.
   SmallVector<CCValAssign, 16> ArgLocs;
   CCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
-  analyzeOutputArgs(MF, ArgCCInfo, Outs, /*IsRet=*/false);
+  analyzeOutputArgs(MF, ArgCCInfo, Outs, /*IsRet=*/false, &CLI);
 
   // Get a count of how many bytes are to be pushed on the stack.
   unsigned NumBytes = ArgCCInfo.getNextStackOffset();
 
   // Create local copies for byval args
   SmallVector<SDValue, 8> ByValArgs;
   for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
     ISD::ArgFlagsTy Flags = Outs[i].Flags;
bool RISCVTargetLowering::CanLowerReturn(
     CallingConv::ID CallConv, MachineFunction &MF, bool IsVarArg,
     const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
   SmallVector<CCValAssign, 16> RVLocs;
   CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
   for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
     MVT VT = Outs[i].VT;
     ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
     if (CC_RISCV(MF.getDataLayout(), i, VT, VT, CCValAssign::Full, ArgFlags,
-                 CCInfo, /*IsFixed=*/true, /*IsRet=*/true))
+                 CCInfo, /*IsFixed=*/true, /*IsRet=*/true, nullptr))
       return false;
   }
   return true;
 }
 
 SDValue
 RISCVTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
                                  bool IsVarArg,
                                  const SmallVectorImpl<ISD::OutputArg> &Outs,
                                  const SmallVectorImpl<SDValue> &OutVals,
                                  const SDLoc &DL, SelectionDAG &DAG) const {
-  if (IsVarArg) {
-    report_fatal_error("VarArg not supported");
-  }
-
   // Stores the assignment of the return value to a location.
   SmallVector<CCValAssign, 16> RVLocs;
 
   // Info about the registers and stack slot.
   CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
                  *DAG.getContext());
 
-  analyzeOutputArgs(DAG.getMachineFunction(), CCInfo, Outs, /*IsRet=*/true);
+  analyzeOutputArgs(DAG.getMachineFunction(), CCInfo, Outs, /*IsRet=*/true,
+                    nullptr);
 
   SDValue Flag;
   SmallVector<SDValue, 4> RetOps(1, Chain);
 
   // Copy the result values into the output registers.
   for (unsigned i = 0, e = RVLocs.size(); i < e; ++i) {
     SDValue Val = OutVals[i];
     CCValAssign &VA = RVLocs[i];

lib/Target/RISCV/RISCVMachineFunctionInfo.h

+//=- RISCVMachineFunctionInfo.h - RISCV machine function info -----*- C++ -*-=//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file declares RISCV-specific per-machine-function information.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_TARGET_RISCV_RISCVMACHINEFUNCTIONINFO_H
+#define LLVM_LIB_TARGET_RISCV_RISCVMACHINEFUNCTIONINFO_H
+
+#include "llvm/CodeGen/MachineFunction.h"
+
+namespace llvm {
+
+/// RISCVMachineFunctionInfo - This class is derived from MachineFunctionInfo
+/// and contains private RISCV-specific information for each MachineFunction.
+class RISCVMachineFunctionInfo : public MachineFunctionInfo {
+
+  /// FrameIndex for start of varargs area
+  int VarArgsFrameIndex = 0;
+  /// Size of the save area used for varargs
+  int VarArgsSaveSize = 0;
+
+public:
+  RISCVMachineFunctionInfo() = default;
+
+  explicit RISCVMachineFunctionInfo(MachineFunction &MF) {}
+
+  int getVarArgsFrameIndex() const { return VarArgsFrameIndex; }
+  void setVarArgsFrameIndex(int Index) { VarArgsFrameIndex = Index; }
+
+  unsigned getVarArgsSaveSize() const { return VarArgsSaveSize; }
+  void setVarArgsSaveSize(int Size) { VarArgsSaveSize = Size; }
+};
+
+} // end namespace llvm
+
+#endif // LLVM_LIB_TARGET_RISCV_RISCVMACHINEFUNCTIONINFO_H

test/CodeGen/RISCV/vararg.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefix=RV32I %s
+
+declare void @llvm.va_start(i8*)
+declare void @llvm.va_end(i8*)
+
+declare void @notdead(i8*)
+
+; Although frontends are recommended to not generate va_arg due to the lack of
+; support for aggregate types, we test simple cases here to ensure they are
+; lowered correctly
+
+define i32 @va1(i8* %fmt, ...) nounwind {
+; RV32I-LABEL: va1:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 12(sp)
+; RV32I-NEXT:    sw s0, 8(sp)
+; RV32I-NEXT:    addi s0, sp, 16
+; RV32I-NEXT:    sw a1, 4(s0)
+; RV32I-NEXT:    sw a7, 28(s0)
+; RV32I-NEXT:    sw a6, 24(s0)
+; RV32I-NEXT:    sw a5, 20(s0)
+; RV32I-NEXT:    sw a4, 16(s0)
+; RV32I-NEXT:    sw a3, 12(s0)
+; RV32I-NEXT:    sw a2, 8(s0)
+; RV32I-NEXT:    addi a0, s0, 8
+; RV32I-NEXT:    sw a0, -12(s0)
+; RV32I-NEXT:    lw a0, 4(s0)
+; RV32I-NEXT:    lw s0, 8(sp)
+; RV32I-NEXT:    lw ra, 12(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    jalr zero, ra, 0
+  %va = alloca i8*, align 4
+  %1 = bitcast i8** %va to i8*
+  call void @llvm.va_start(i8* %1)
+  %argp.cur = load i8*, i8** %va, align 4
+  %argp.next = getelementptr inbounds i8, i8* %argp.cur, i32 4
+  store i8* %argp.next, i8** %va, align 4
+  %2 = bitcast i8* %argp.cur to i32*
+  %3 = load i32, i32* %2, align 4
+  call void @llvm.va_end(i8* %1)
+  ret i32 %3
+}
+
+define i32 @va1_va_arg(i8* %fmt, ...) nounwind {
+; RV32I-LABEL: va1_va_arg:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 12(sp)
+; RV32I-NEXT:    sw s0, 8(sp)
+; RV32I-NEXT:    addi s0, sp, 16
+; RV32I-NEXT:    sw a1, 4(s0)
+; RV32I-NEXT:    sw a7, 28(s0)
+; RV32I-NEXT:    sw a6, 24(s0)
+; RV32I-NEXT:    sw a5, 20(s0)
+; RV32I-NEXT:    sw a4, 16(s0)
+; RV32I-NEXT:    sw a3, 12(s0)
+; RV32I-NEXT:    sw a2, 8(s0)
+; RV32I-NEXT:    addi a0, s0, 8
+; RV32I-NEXT:    sw a0, -12(s0)
+; RV32I-NEXT:    lw a0, 4(s0)
+; RV32I-NEXT:    lw s0, 8(sp)
+; RV32I-NEXT:    lw ra, 12(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    jalr zero, ra, 0
+  %va = alloca i8*, align 4
+  %1 = bitcast i8** %va to i8*
+  call void @llvm.va_start(i8* %1)
+  %2 = va_arg i8** %va, i32
+  call void @llvm.va_end(i8* %1)
+  ret i32 %2
+}
+
+; Ensure the adjustment when restoring the stack pointer using the frame
+; pointer is correct
+define i32 @va1_va_arg_alloca(i8* %fmt, ...) nounwind {
+; RV32I-LABEL: va1_va_arg_alloca:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 12(sp)
+; RV32I-NEXT:    sw s0, 8(sp)
+; RV32I-NEXT:    sw s1, 4(sp)
+; RV32I-NEXT:    addi s0, sp, 16
+; RV32I-NEXT:    sw a1, 4(s0)
+; RV32I-NEXT:    sw a7, 28(s0)
+; RV32I-NEXT:    sw a6, 24(s0)
+; RV32I-NEXT:    sw a5, 20(s0)
+; RV32I-NEXT:    sw a4, 16(s0)
+; RV32I-NEXT:    sw a3, 12(s0)
+; RV32I-NEXT:    sw a2, 8(s0)
+; RV32I-NEXT:    addi a0, s0, 8
+; RV32I-NEXT:    sw a0, -16(s0)
+; RV32I-NEXT:    lw s1, 4(s0)
+; RV32I-NEXT:    addi a0, s1, 15
+; RV32I-NEXT:    andi a0, a0, -16
+; RV32I-NEXT:    sub a0, sp, a0
+; RV32I-NEXT:    addi sp, a0, 0
+; RV32I-NEXT:    lui a1, %hi(notdead)
+; RV32I-NEXT:    addi a1, a1, %lo(notdead)
+; RV32I-NEXT:    jalr ra, a1, 0
+; RV32I-NEXT:    addi a0, s1, 0
+; RV32I-NEXT:    addi sp, s0, -16
+; RV32I-NEXT:    lw s1, 4(sp)
+; RV32I-NEXT:    lw s0, 8(sp)
+; RV32I-NEXT:    lw ra, 12(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    jalr zero, ra, 0
+  %va = alloca i8*, align 4
+  %1 = bitcast i8** %va to i8*
+  call void @llvm.va_start(i8* %1)
+  %2 = va_arg i8** %va, i32
+  %3 = alloca i8, i32 %2
+  call void @notdead(i8* %3)
+  call void @llvm.va_end(i8* %1)
+  ret i32 %2
+}
+
+define void @va1_caller() nounwind {
+; RV32I-LABEL: va1_caller:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -16
+; RV32I-NEXT:    sw ra, 12(sp)
+; RV32I-NEXT:    sw s0, 8(sp)
+; RV32I-NEXT:    addi s0, sp, 16
+; RV32I-NEXT:    lui a0, 261888
+; RV32I-NEXT:    addi a3, a0, 0
+; RV32I-NEXT:    lui a0, %hi(va1)
+; RV32I-NEXT:    addi a0, a0, %lo(va1)
+; RV32I-NEXT:    addi a4, zero, 2
+; RV32I-NEXT:    addi a2, zero, 0
+; RV32I-NEXT:    jalr ra, a0, 0
+; RV32I-NEXT:    lw s0, 8(sp)
+; RV32I-NEXT:    lw ra, 12(sp)
+; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    jalr zero, ra, 0
+; Pass a double, as a float would be promoted by a C/C++ frontend
+  %1 = call i32 (i8*, ...) @va1(i8* undef, double 1.0, i32 2)
+  ret void
+}
+
+; Ensure that 2x xlen size+alignment varargs are accessed via an "aligned"
+; register pair (where the first register is even-numbered).
+
+define double @va2(i8 *%fmt, ...) nounwind {
+; RV32I-LABEL: va2:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 12(sp)
+; RV32I-NEXT:    sw s0, 8(sp)
+; RV32I-NEXT:    addi s0, sp, 16
+; RV32I-NEXT:    sw a7, 28(s0)
+; RV32I-NEXT:    sw a6, 24(s0)
+; RV32I-NEXT:    sw a5, 20(s0)
+; RV32I-NEXT:    sw a4, 16(s0)
+; RV32I-NEXT:    sw a3, 12(s0)
+; RV32I-NEXT:    sw a2, 8(s0)
+; RV32I-NEXT:    sw a1, 4(s0)
+; RV32I-NEXT:    addi a0, s0, 19
+; RV32I-NEXT:    sw a0, -12(s0)
+; RV32I-NEXT:    addi a0, s0, 11
+; RV32I-NEXT:    andi a1, a0, -8
+; RV32I-NEXT:    lw a0, 0(a1)
+; RV32I-NEXT:    ori a1, a1, 4
+; RV32I-NEXT:    lw a1, 0(a1)
+; RV32I-NEXT:    lw s0, 8(sp)
+; RV32I-NEXT:    lw ra, 12(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    jalr zero, ra, 0
+  %va = alloca i8*, align 4
+  %1 = bitcast i8** %va to i8*
+  call void @llvm.va_start(i8* %1)
+  %2 = bitcast i8** %va to i32*
+  %argp.cur = load i32, i32* %2, align 4
+  %3 = add i32 %argp.cur, 7
+  %4 = and i32 %3, -8
+  %argp.cur.aligned = inttoptr i32 %3 to i8*
+  %argp.next = getelementptr inbounds i8, i8* %argp.cur.aligned, i32 8
+  store i8* %argp.next, i8** %va, align 4
+  %5 = inttoptr i32 %4 to double*
+  %6 = load double, double* %5, align 8
+  call void @llvm.va_end(i8* %1)
+  ret double %6
+}
+
+define double @va2_va_arg(i8 *%fmt, ...) nounwind {
+; RV32I-LABEL: va2_va_arg:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 12(sp)
+; RV32I-NEXT:    sw s0, 8(sp)
+; RV32I-NEXT:    addi s0, sp, 16
+; RV32I-NEXT:    sw a7, 28(s0)
+; RV32I-NEXT:    sw a6, 24(s0)
+; RV32I-NEXT:    sw a5, 20(s0)
+; RV32I-NEXT:    sw a4, 16(s0)
+; RV32I-NEXT:    sw a3, 12(s0)
+; RV32I-NEXT:    sw a2, 8(s0)
+; RV32I-NEXT:    sw a1, 4(s0)
+; RV32I-NEXT:    addi a0, s0, 11
+; RV32I-NEXT:    andi a0, a0, -8
+; RV32I-NEXT:    ori a1, a0, 4
+; RV32I-NEXT:    sw a1, -12(s0)
+; RV32I-NEXT:    lw a0, 0(a0)
+; RV32I-NEXT:    addi a2, a1, 4
+; RV32I-NEXT:    sw a2, -12(s0)
+; RV32I-NEXT:    lw a1, 0(a1)
+; RV32I-NEXT:    lw s0, 8(sp)
+; RV32I-NEXT:    lw ra, 12(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    jalr zero, ra, 0
+  %va = alloca i8*, align 4
+  %1 = bitcast i8** %va to i8*
+  call void @llvm.va_start(i8* %1)
+  %2 = va_arg i8** %va, double
+  call void @llvm.va_end(i8* %1)
+  ret double %2
+}
+
+define void @va2_caller() nounwind {
+; RV32I-LABEL: va2_caller:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -16
+; RV32I-NEXT:    sw ra, 12(sp)
+; RV32I-NEXT:    sw s0, 8(sp)
+; RV32I-NEXT:    addi s0, sp, 16
+; RV32I-NEXT:    lui a0, 261888
+; RV32I-NEXT:    addi a3, a0, 0
+; RV32I-NEXT:    lui a0, %hi(va2)
+; RV32I-NEXT:    addi a0, a0, %lo(va2)
+; RV32I-NEXT:    addi a2, zero, 0
+; RV32I-NEXT:    jalr ra, a0, 0
+; RV32I-NEXT:    lw s0, 8(sp)
+; RV32I-NEXT:    lw ra, 12(sp)
+; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    jalr zero, ra, 0
+ %1 = call double (i8*, ...) @va2(i8* undef, double 1.000000e+00)
+ ret void
+}
+
+; Ensure a named double argument is passed in a1 and a2, while the vararg
+; double is passed in a4 and a5 (rather than a3 and a4)
+
+define double @va3(i32 %a, double %b, ...) nounwind {
+; RV32I-LABEL: va3:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 20(sp)
+; RV32I-NEXT:    sw s0, 16(sp)
+; RV32I-NEXT:    addi s0, sp, 24
+; RV32I-NEXT:    sw a7, 20(s0)
+; RV32I-NEXT:    sw a6, 16(s0)
+; RV32I-NEXT:    sw a5, 12(s0)
+; RV32I-NEXT:    sw a4, 8(s0)
+; RV32I-NEXT:    sw a3, 4(s0)
+; RV32I-NEXT:    addi a0, s0, 19
+; RV32I-NEXT:    sw a0, -12(s0)
+; RV32I-NEXT:    lui a0, %hi(__adddf3)
+; RV32I-NEXT:    addi a5, a0, %lo(__adddf3)
+; RV32I-NEXT:    addi a0, s0, 11
+; RV32I-NEXT:    andi a0, a0, -8
+; RV32I-NEXT:    lw a4, 0(a0)
+; RV32I-NEXT:    ori a0, a0, 4
+; RV32I-NEXT:    lw a3, 0(a0)
+; RV32I-NEXT:    addi a0, a1, 0
+; RV32I-NEXT:    addi a1, a2, 0
+; RV32I-NEXT:    addi a2, a4, 0
+; RV32I-NEXT:    jalr ra, a5, 0
+; RV32I-NEXT:    lw s0, 16(sp)
+; RV32I-NEXT:    lw ra, 20(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    jalr zero, ra, 0
+  %va = alloca i8*, align 4
+  %1 = bitcast i8** %va to i8*
+  call void @llvm.va_start(i8* %1)
+  %2 = bitcast i8** %va to i32*
+  %argp.cur = load i32, i32* %2, align 4
+  %3 = add i32 %argp.cur, 7
+  %4 = and i32 %3, -8
+  %argp.cur.aligned = inttoptr i32 %3 to i8*
+  %argp.next = getelementptr inbounds i8, i8* %argp.cur.aligned, i32 8
+  store i8* %argp.next, i8** %va, align 4
+  %5 = inttoptr i32 %4 to double*
+  %6 = load double, double* %5, align 8
+  call void @llvm.va_end(i8* %1)
+  %7 = fadd double %b, %6
+  ret double %7
+}
+
+define double @va3_va_arg(i32 %a, double %b, ...) nounwind {
+; RV32I-LABEL: va3_va_arg:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 20(sp)
+; RV32I-NEXT:    sw s0, 16(sp)
+; RV32I-NEXT:    addi s0, sp, 24
+; RV32I-NEXT:    sw a7, 20(s0)
+; RV32I-NEXT:    sw a6, 16(s0)
+; RV32I-NEXT:    sw a5, 12(s0)
+; RV32I-NEXT:    sw a4, 8(s0)
+; RV32I-NEXT:    sw a3, 4(s0)
+; RV32I-NEXT:    addi a0, s0, 11
+; RV32I-NEXT:    andi a0, a0, -8
+; RV32I-NEXT:    ori a3, a0, 4
+; RV32I-NEXT:    sw a3, -12(s0)
+; RV32I-NEXT:    lw a4, 0(a0)
+; RV32I-NEXT:    addi a0, a3, 4
+; RV32I-NEXT:    sw a0, -12(s0)
+; RV32I-NEXT:    lui a0, %hi(__adddf3)
+; RV32I-NEXT:    addi a5, a0, %lo(__adddf3)
+; RV32I-NEXT:    lw a3, 0(a3)
+; RV32I-NEXT:    addi a0, a1, 0
+; RV32I-NEXT:    addi a1, a2, 0
+; RV32I-NEXT:    addi a2, a4, 0
+; RV32I-NEXT:    jalr ra, a5, 0
+; RV32I-NEXT:    lw s0, 16(sp)
+; RV32I-NEXT:    lw ra, 20(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    jalr zero, ra, 0
+  %va = alloca i8*, align 4
+  %1 = bitcast i8** %va to i8*
+  call void @llvm.va_start(i8* %1)
+  %2 = va_arg i8** %va, double
+  call void @llvm.va_end(i8* %1)
+  %3 = fadd double %b, %2
+  ret double %3
+}
+
+define void @va3_caller() nounwind {
+; RV32I-LABEL: va3_caller:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -16
+; RV32I-NEXT:    sw ra, 12(sp)
+; RV32I-NEXT:    sw s0, 8(sp)
+; RV32I-NEXT:    addi s0, sp, 16
+; RV32I-NEXT:    lui a0, 261888
+; RV32I-NEXT:    addi a2, a0, 0
+; RV32I-NEXT:    lui a0, 262144
+; RV32I-NEXT:    addi a5, a0, 0
+; RV32I-NEXT:    lui a0, %hi(va3)
+; RV32I-NEXT:    addi a3, a0, %lo(va3)
+; RV32I-NEXT:    addi a0, zero, 2
+; RV32I-NEXT:    addi a1, zero, 0
+; RV32I-NEXT:    addi a4, zero, 0
+; RV32I-NEXT:    jalr ra, a3, 0
+; RV32I-NEXT:    lw s0, 8(sp)
+; RV32I-NEXT:    lw ra, 12(sp)
+; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    jalr zero, ra, 0
+ %1 = call double (i32, double, ...) @va3(i32 2, double 1.000000e+00, double 2.000000e+00)
+ ret void
+}
+
+declare void @llvm.va_copy(i8*, i8*)
+
+define i32 @va4_va_copy(i32 %argno, ...) nounwind {
+; RV32I-LABEL: va4_va_copy:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -64
+; RV32I-NEXT:    sw ra, 28(sp)
+; RV32I-NEXT:    sw s0, 24(sp)
+; RV32I-NEXT:    sw s1, 20(sp)
+; RV32I-NEXT:    addi s0, sp, 32
+; RV32I-NEXT:    sw a1, 4(s0)
+; RV32I-NEXT:    sw a7, 28(s0)
+; RV32I-NEXT:    sw a6, 24(s0)
+; RV32I-NEXT:    sw a5, 20(s0)
+; RV32I-NEXT:    sw a4, 16(s0)
+; RV32I-NEXT:    sw a3, 12(s0)
+; RV32I-NEXT:    sw a2, 8(s0)
+; RV32I-NEXT:    addi a0, s0, 8
+; RV32I-NEXT:    sw a0, -16(s0)
+; RV32I-NEXT:    sw a0, -20(s0)
+; RV32I-NEXT:    lw s1, 4(s0)
+; RV32I-NEXT:    lui a1, %hi(notdead)
+; RV32I-NEXT:    addi a1, a1, %lo(notdead)
+; RV32I-NEXT:    jalr ra, a1, 0
+; RV32I-NEXT:    lw a0, -16(s0)
+; RV32I-NEXT:    addi a0, a0, 3
+; RV32I-NEXT:    andi a0, a0, -4
+; RV32I-NEXT:    addi a1, a0, 4
+; RV32I-NEXT:    sw a1, -16(s0)
+; RV32I-NEXT:    lw a1, 0(a0)
+; RV32I-NEXT:    addi a0, a0, 7
+; RV32I-NEXT:    andi a0, a0, -4
+; RV32I-NEXT:    addi a2, a0, 4
+; RV32I-NEXT:    sw a2, -16(s0)
+; RV32I-NEXT:    lw a2, 0(a0)
+; RV32I-NEXT:    addi a0, a0, 7
+; RV32I-NEXT:    andi a0, a0, -4
+; RV32I-NEXT:    addi a3, a0, 4
+; RV32I-NEXT:    sw a3, -16(s0)
+; RV32I-NEXT:    add a1, a1, s1
+; RV32I-NEXT:    add a1, a1, a2
+; RV32I-NEXT:    lw a0, 0(a0)
+; RV32I-NEXT:    add a0, a1, a0
+; RV32I-NEXT:    lw s1, 20(sp)
+; RV32I-NEXT:    lw s0, 24(sp)
+; RV32I-NEXT:    lw ra, 28(sp)
+; RV32I-NEXT:    addi sp, sp, 64
+; RV32I-NEXT:    jalr zero, ra, 0
+  %vargs = alloca i8*, align 4
+  %wargs = alloca i8*, align 4
+  %1 = bitcast i8** %vargs to i8*
+  %2 = bitcast i8** %wargs to i8*
+  call void @llvm.va_start(i8* %1)
+  %3 = va_arg i8** %vargs, i32
+  call void @llvm.va_copy(i8* %2, i8* %1)
+  %4 = load i8*, i8** %wargs, align 4
+  call void @notdead(i8* %4)
+  %5 = va_arg i8** %vargs, i32
+  %6 = va_arg i8** %vargs, i32
+  %7 = va_arg i8** %vargs, i32
+  call void @llvm.va_end(i8* %1)
+  call void @llvm.va_end(i8* %2)
+  %add1 = add i32 %5, %3
+  %add2 = add i32 %add1, %6
+  %add3 = add i32 %add2, %7
+  ret i32 %add3
+}
+
+; Check 2x*xlen values are aligned appropriately when passed on the stack in a vararg call
+
+define i32 @va5_aligned_stack_callee(i32 %a, ...) nounwind {
+; RV32I-LABEL: va5_aligned_stack_callee:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 12(sp)
+; RV32I-NEXT:    sw s0, 8(sp)
+; RV32I-NEXT:    addi s0, sp, 16
+; RV32I-NEXT:    sw a7, 28(s0)
+; RV32I-NEXT:    sw a6, 24(s0)
+; RV32I-NEXT:    sw a5, 20(s0)
+; RV32I-NEXT:    sw a4, 16(s0)
+; RV32I-NEXT:    sw a3, 12(s0)
+; RV32I-NEXT:    sw a2, 8(s0)
+; RV32I-NEXT:    sw a1, 4(s0)
+; RV32I-NEXT:    addi a0, zero, 1
+; RV32I-NEXT:    lw s0, 8(sp)
+; RV32I-NEXT:    lw ra, 12(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    jalr zero, ra, 0
+  ret i32 1
+}
+
+define void @va5_aligned_stack_caller() nounwind {
+; The double should be 8-byte aligned on the stack, but the two-element array
+; should only be 4-byte aligned
+; RV32I-LABEL: va5_aligned_stack_caller:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -64
+; RV32I-NEXT:    sw ra, 60(sp)
+; RV32I-NEXT:    sw s0, 56(sp)
+; RV32I-NEXT:    addi s0, sp, 64
+; RV32I-NEXT:    addi a0, zero, 17
+; RV32I-NEXT:    sw a0, 24(sp)
+; RV32I-NEXT:    addi a0, zero, 16
+; RV32I-NEXT:    sw a0, 20(sp)
+; RV32I-NEXT:    addi a0, zero, 15
+; RV32I-NEXT:    sw a0, 16(sp)
+; RV32I-NEXT:    lui a0, 262236
+; RV32I-NEXT:    addi a0, a0, 655
+; RV32I-NEXT:    sw a0, 12(sp)
+; RV32I-NEXT:    lui a0, 377487
+; RV32I-NEXT:    addi a0, a0, 1475
+; RV32I-NEXT:    sw a0, 8(sp)
+; RV32I-NEXT:    addi a0, zero, 14
+; RV32I-NEXT:    sw a0, 0(sp)
+; RV32I-NEXT:    lui a0, 262153
+; RV32I-NEXT:    addi a0, a0, 491
+; RV32I-NEXT:    sw a0, -20(s0)
+; RV32I-NEXT:    lui a0, 545260
+; RV32I-NEXT:    addi a0, a0, -1967
+; RV32I-NEXT:    sw a0, -24(s0)
+; RV32I-NEXT:    lui a0, 964690
+; RV32I-NEXT:    addi a0, a0, -328
+; RV32I-NEXT:    sw a0, -28(s0)
+; RV32I-NEXT:    lui a0, 335544
+; RV32I-NEXT:    addi a0, a0, 1311
+; RV32I-NEXT:    sw a0, -32(s0)
+; RV32I-NEXT:    lui a0, 688509
+; RV32I-NEXT:    addi a6, a0, -2048
+; RV32I-NEXT:    lui a0, %hi(va5_aligned_stack_callee)
+; RV32I-NEXT:    addi a5, a0, %lo(va5_aligned_stack_callee)
+; RV32I-NEXT:    addi a0, zero, 1
+; RV32I-NEXT:    addi a1, zero, 11
+; RV32I-NEXT:    addi a2, s0, -32
+; RV32I-NEXT:    addi a3, zero, 12
+; RV32I-NEXT:    addi a4, zero, 13
+; RV32I-NEXT:    addi a7, zero, 4
+; RV32I-NEXT:    jalr ra, a5, 0
+; RV32I-NEXT:    lw s0, 56(sp)
+; RV32I-NEXT:    lw ra, 60(sp)
+; RV32I-NEXT:    addi sp, sp, 64
+; RV32I-NEXT:    jalr zero, ra, 0
+  %1 = call i32 (i32, ...) @va5_aligned_stack_callee(i32 1, i32 11,
+    fp128 0xLEB851EB851EB851F400091EB851EB851, i32 12, i32 13, i64 20000000000,
+    i32 14, double 2.720000e+00, i32 15, [2 x i32] [i32 16, i32 17])
+  ret void
+}
+
+; A function with no fixed arguments is not valid C, but can be
+; specified in LLVM IR. We must ensure the vararg save area is
+; still set up correctly.
+
+define i32 @va6_no_fixed_args(...) nounwind {
+; RV32I-LABEL: va6_no_fixed_args:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 12(sp)
+; RV32I-NEXT:    sw s0, 8(sp)
+; RV32I-NEXT:    addi s0, sp, 16
+; RV32I-NEXT:    sw a0, 0(s0)
+; RV32I-NEXT:    sw a7, 28(s0)
+; RV32I-NEXT:    sw a6, 24(s0)
+; RV32I-NEXT:    sw a5, 20(s0)
+; RV32I-NEXT:    sw a4, 16(s0)
+; RV32I-NEXT:    sw a3, 12(s0)
+; RV32I-NEXT:    sw a2, 8(s0)
+; RV32I-NEXT:    sw a1, 4(s0)
+; RV32I-NEXT:    addi a0, s0, 4
+; RV32I-NEXT:    sw a0, -12(s0)
+; RV32I-NEXT:    lw a0, 0(s0)
+; RV32I-NEXT:    lw s0, 8(sp)
+; RV32I-NEXT:    lw ra, 12(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    jalr zero, ra, 0
+  %va = alloca i8*, align 4
+  %1 = bitcast i8** %va to i8*
+  call void @llvm.va_start(i8* %1)
+  %2 = va_arg i8** %va, i32
+  call void @llvm.va_end(i8* %1)
+  ret i32 %2
+}