[WebAssembly] Varargs support

[sunfish]

Summary

This is a preliminary patch for varargs support. It doesn't all work yet, and is quite untested. I'm just sending out the patch for comments.

[jfb]

Looks simpler than I thought it would be!

lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

321

This is just extra validation, unrelated to varargs? No need to split it out, I just want to make sure I'm not missing something.

347

This is pretty much the same approach as PNaCl, but done right. IIUC: varargs are effectively in a naturally-aligned struct at the top of stack, and the callee gets a pointer to the stark. Is that correct? If so, could the comment explain that part?

380

This just appends the fixed args, and doesn't append anything related to varags, correct?

464

Duplicated above. Factor out?

[sunfish]

Roughly the same patch, but with some of the uninteresting bits split out, and some extra comments.

[sunfish]

One thing slightly different about this varargs implementation from the PNaCl LowerVarArgs pass is that this implementation doesn't pass a pointer argument; it just puts the varargs arguments at the top of the stack, so they are pointed to by the stack pointer (or from the callee's perspective, by the frame pointer). We could go either way, but I don't think it makes a difference in performance (and besides, this is varargs ;-)), and I think it's a little tidier this way.

lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

321

That's right. I decided to just split it out now, so it's out of the way.

347

Comments added.

380

I added a comment explaining this.

464

Done.

[sunfish]

Various fixes and cleanups. And add a testcase. Note that va_start and actual arguments don't work yet; they depend on frame pointer support. All the tests pass with this patch applied, so I think it's ready to land.

[jfb]

lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

338

Add a comment string to the assert.

363

Ditto.

444

Typo?

lib/Target/WebAssembly/WebAssemblyRegStackify.cpp

207

Does this happen because of varargs? Could you comment.

test/CodeGen/WebAssembly/varargs.ll

126

No need for the attributes.

[sunfish]

Update for review comments.

[sunfish]

Review comments addressed.

[jfb]

lgtm

lib/Target/WebAssembly/WebAssemblyISelLowering.h

SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
                                SmallVectorImpl<SDValue> &InVals) const override;
 
   // Custom lowering hooks.
   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
   SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
 };
 
 namespace WebAssembly {
 FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
                          const TargetLibraryInfo *libInfo);
 } // end namespace WebAssembly
 
 } // end namespace llvm
 
 #endif

lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

WebAssemblyTargetLowering::WebAssemblyTargetLowering(
   addRegisterClass(MVT::f64, &WebAssembly::F64RegClass);
   // Compute derived properties from the register classes.
   computeRegisterProperties(Subtarget->getRegisterInfo());
 
   setOperationAction(ISD::GlobalAddress, MVTPtr, Custom);
   setOperationAction(ISD::ExternalSymbol, MVTPtr, Custom);
   setOperationAction(ISD::JumpTable, MVTPtr, Custom);
 
+  // Take the default expansion for va_arg, va_copy, and va_end. There is no
+  // default action for va_start, so we do that custom.
+  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 T : {MVT::f32, MVT::f64}) {
     // Don't expand the floating-point types to constant pools.
     setOperationAction(ISD::ConstantFP, T, Legal);
     // Expand floating-point comparisons.
     for (auto CC : {ISD::SETO, ISD::SETUO, ISD::SETUEQ, ISD::SETONE,
                     ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE})
       setCondCodeAction(CC, T, Expand);
     // Expand floating-point library function operators.
for (const ISD::OutputArg &Out : Outs) {
     assert(!Out.Flags.isByVal() && "byval is not valid for return values");
     assert(!Out.Flags.isNest() && "nest is not valid for return values");
     if (Out.Flags.isInAlloca())
       fail(DL, DAG, "WebAssembly hasn't implemented inalloca results");
     if (Out.Flags.isInConsecutiveRegs())
       fail(DL, DAG, "WebAssembly hasn't implemented cons regs results");
     if (Out.Flags.isInConsecutiveRegsLast())
       fail(DL, DAG, "WebAssembly hasn't implemented cons regs last results");
   }
 
   bool IsVarArg = CLI.IsVarArg;
-  if (IsVarArg)
-    fail(DL, DAG, "WebAssembly doesn't support varargs yet");
+  unsigned NumFixedArgs = CLI.NumFixedArgs;
+  auto PtrVT = getPointerTy(MF.getDataLayout());
 
   // Analyze operands of the call, assigning locations to each operand.
   SmallVector<CCValAssign, 16> ArgLocs;
   CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
-  unsigned NumBytes = CCInfo.getNextStackOffset();
 
-  auto PtrVT = getPointerTy(MF.getDataLayout());
-  auto Zero = DAG.getConstant(0, DL, PtrVT, true);
+  if (IsVarArg) {
+    // Outgoing non-fixed arguments are placed at the top of the stack. First
+    // compute their offsets and the total amount of argument stack space
+    // needed.
+    for (SDValue Arg :
+         make_range(OutVals.begin() + NumFixedArgs, OutVals.end())) {
+      EVT VT = Arg.getValueType();
+      assert(VT != MVT::iPTR);
+      Type *Ty = VT.getTypeForEVT(*DAG.getContext());
+      unsigned Offset =
+          CCInfo.AllocateStack(MF.getDataLayout().getTypeAllocSize(Ty),
+                               MF.getDataLayout().getABITypeAlignment(Ty));
+      CCInfo.addLoc(CCValAssign::getMem(ArgLocs.size(), VT.getSimpleVT(),
+                                        Offset, VT.getSimpleVT(),
+                                        CCValAssign::Full));
+    }
+  }
+
+  unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
+
   auto NB = DAG.getConstant(NumBytes, DL, PtrVT, true);
   Chain = DAG.getCALLSEQ_START(Chain, NB, DL);
 
+  if (IsVarArg) {
+    // For non-fixed arguments, next emit stores to store the argument values
+    // to the stack at the offsets computed above.
+    SDValue SP = DAG.getCopyFromReg(
+        Chain, DL, getStackPointerRegisterToSaveRestore(), PtrVT);
+    unsigned ValNo = 0;
+    SmallVector<SDValue, 8> Chains;
+    for (SDValue Arg :
+         make_range(OutVals.begin() + NumFixedArgs, OutVals.end())) {
+      assert(ArgLocs[ValNo].getValNo() == ValNo);
+      unsigned Offset = ArgLocs[ValNo++].getLocMemOffset();
+      SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, SP,
+                                DAG.getConstant(Offset, DL, PtrVT));
+      Chains.push_back(DAG.getStore(Chain, DL, Arg, Add,
+                                    MachinePointerInfo::getStack(MF, Offset),
+                                    false, false, 0));
+    }
+    if (!Chains.empty())
+      Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
+  }
+
+  // Compute the operands for the CALLn node.
   SmallVector<SDValue, 16> Ops;
   Ops.push_back(Chain);
   Ops.push_back(Callee);
-  Ops.append(OutVals.begin(), OutVals.end());
+
+  // Add all fixed arguments. Note that for non-varargs calls, NumFixedArgs
+  // isn't reliable.
+  Ops.append(OutVals.begin(),
+             IsVarArg ? OutVals.begin() + NumFixedArgs : OutVals.end());
 
   SmallVector<EVT, 8> Tys;
   for (const auto &In : Ins) {
     if (In.Flags.isByVal())
       fail(DL, DAG, "WebAssembly hasn't implemented byval arguments");
     if (In.Flags.isInAlloca())
       fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
     if (In.Flags.isNest())
SDValue Res =
                   DL, TyList, Ops);
   if (Ins.empty()) {
     Chain = Res;
   } else {
     InVals.push_back(Res);
     Chain = Res.getValue(1);
   }
 
-  Chain = DAG.getCALLSEQ_END(Chain, NB, Zero, SDValue(), DL);
+  SDValue Unused = DAG.getUNDEF(PtrVT);
+  Chain = DAG.getCALLSEQ_END(Chain, NB, Unused, SDValue(), DL);
 
   return Chain;
 }
 
 bool WebAssemblyTargetLowering::CanLowerReturn(
     CallingConv::ID /*CallConv*/, MachineFunction & /*MF*/, bool /*IsVarArg*/,
     const SmallVectorImpl<ISD::OutputArg> &Outs,
     LLVMContext & /*Context*/) const {
   // WebAssembly can't currently handle returning tuples.
   return Outs.size() <= 1;
 }
 
 SDValue WebAssemblyTargetLowering::LowerReturn(
-    SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
+    SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/,
     const SmallVectorImpl<ISD::OutputArg> &Outs,
     const SmallVectorImpl<SDValue> &OutVals, SDLoc DL,
     SelectionDAG &DAG) const {
   assert(Outs.size() <= 1 && "WebAssembly can only return up to one value");
   if (!CallingConvSupported(CallConv))
     fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
-  if (IsVarArg)
-    fail(DL, DAG, "WebAssembly doesn't support varargs yet");
 
   SmallVector<SDValue, 4> RetOps(1, Chain);
   RetOps.append(OutVals.begin(), OutVals.end());
   Chain = DAG.getNode(WebAssemblyISD::RETURN, DL, MVT::Other, RetOps);
 
   // Record the number and types of the return values.
   for (const ISD::OutputArg &Out : Outs) {
     assert(!Out.Flags.isByVal() && "byval is not valid for return values");
     assert(!Out.Flags.isNest() && "nest is not valid for return values");
+    assert(Out.IsFixed && "non-fixed return valid is not valid");
     if (Out.Flags.isInAlloca())
       fail(DL, DAG, "WebAssembly hasn't implemented inalloca results");
     if (Out.Flags.isInConsecutiveRegs())
       fail(DL, DAG, "WebAssembly hasn't implemented cons regs results");
     if (Out.Flags.isInConsecutiveRegsLast())
       fail(DL, DAG, "WebAssembly hasn't implemented cons regs last results");
-    if (!Out.IsFixed)
-      fail(DL, DAG, "WebAssembly doesn't support non-fixed results yet");
   }
 
   return Chain;
 }
 
 SDValue WebAssemblyTargetLowering::LowerFormalArguments(
-    SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
+    SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/,
     const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
     SmallVectorImpl<SDValue> &InVals) const {
   MachineFunction &MF = DAG.getMachineFunction();
 
   if (!CallingConvSupported(CallConv))
     fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
-  if (IsVarArg)
-    fail(DL, DAG, "WebAssembly doesn't support varargs yet");
 
   // Set up the incoming ARGUMENTS value, which serves to represent the liveness
   // of the incoming values before they're represented by virtual registers.
   MF.getRegInfo().addLiveIn(WebAssembly::ARGUMENTS);
 
   for (const ISD::InputArg &In : Ins) {
     if (In.Flags.isByVal())
       fail(DL, DAG, "WebAssembly hasn't implemented byval arguments");
     if (In.Flags.isInAlloca())
InVals.push_back(
             ? DAG.getNode(WebAssemblyISD::ARGUMENT, DL, In.VT,
                           DAG.getTargetConstant(InVals.size(), DL, MVT::i32))
             : DAG.getUNDEF(In.VT));
 
     // Record the number and types of arguments.
     MF.getInfo<WebAssemblyFunctionInfo>()->addParam(In.VT);
   }
 
+  // Incoming varargs arguments are on the stack and will be accessed through
+  // va_arg, so we don't need to do anything for them here.
+
   return Chain;
 }
 
 //===----------------------------------------------------------------------===//
 //  Custom lowering hooks.
 //===----------------------------------------------------------------------===//
 
 SDValue WebAssemblyTargetLowering::LowerOperation(SDValue Op,
                                                   SelectionDAG &DAG) const {
   switch (Op.getOpcode()) {
   default:
     llvm_unreachable("unimplemented operation lowering");
     return SDValue();
   case ISD::GlobalAddress:
     return LowerGlobalAddress(Op, DAG);
   case ISD::ExternalSymbol:
     return LowerExternalSymbol(Op, DAG);
   case ISD::JumpTable:
     return LowerJumpTable(Op, DAG);
   case ISD::BR_JT:
     return LowerBR_JT(Op, DAG);
+  case ISD::VASTART:
+    return LowerVASTART(Op, DAG);
   }
 }
 
 SDValue WebAssemblyTargetLowering::LowerGlobalAddress(SDValue Op,
                                                       SelectionDAG &DAG) const {
   SDLoc DL(Op);
   const auto *GA = cast<GlobalAddressSDNode>(Op);
   EVT VT = Op.getValueType();
SDValue WebAssemblyTargetLowering::LowerBR_JT(SDValue Op,
 
   // Add an operand for each case.
   for (auto MBB : MBBs)
     Ops.push_back(DAG.getBasicBlock(MBB));
 
   return DAG.getNode(WebAssemblyISD::TABLESWITCH, DL, MVT::Other, Ops);
 }
 
+SDValue WebAssemblyTargetLowering::LowerVASTART(SDValue Op,
+                                                SelectionDAG &DAG) const {
+  SDLoc DL(Op);
+  EVT PtrVT = getPointerTy(DAG.getMachineFunction().getDataLayout());
+
+  // The incoming non-fixed arguments are placed on the top of the stack, with
+  // natural alignment, at the point of the call, so the base pointer is just
+  // the current frame pointer.
+  DAG.getMachineFunction().getFrameInfo()->setFrameAddressIsTaken(true);
+  unsigned FP =
+      static_cast<const WebAssemblyRegisterInfo *>(Subtarget->getRegisterInfo())
+          ->getFrameRegister(DAG.getMachineFunction());
+  SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, FP, PtrVT);
+  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
+  return DAG.getStore(Op.getOperand(0), DL, FrameAddr, Op.getOperand(1),
+                      MachinePointerInfo(SV), false, false, 0);
+}
+
 //===----------------------------------------------------------------------===//
 //                          WebAssembly Optimization Hooks
 //===----------------------------------------------------------------------===//
 
 MCSection *WebAssemblyTargetObjectFile::SelectSectionForGlobal(
     const GlobalValue *GV, SectionKind /*Kind*/, Mangler & /*Mang*/,
     const TargetMachine & /*TM*/) const {
   // TODO: Be more sophisticated than this.
   return isa<Function>(GV) ? getTextSection() : getDataSection();
 }

lib/Target/WebAssembly/WebAssemblyInstrCall.td

 
 let Defs = [ARGUMENTS] in {
 
 // The call sequence start/end LLVM-isms isn't useful to WebAssembly since it's
 // a virtual ISA.
 let isCodeGenOnly = 1 in {
 def ADJCALLSTACKDOWN : I<(outs), (ins i64imm:$amt),
                          [(WebAssemblycallseq_start timm:$amt)]>;
-def ADJCALLSTACKUP : I<(outs), (ins i64imm:$amt1, i64imm:$amt2),
-                       [(WebAssemblycallseq_end timm:$amt1, timm:$amt2)]>;
+def ADJCALLSTACKUP : I<(outs), (ins i64imm:$amt),
+                       [(WebAssemblycallseq_end timm:$amt, undef)]>;
 } // isCodeGenOnly = 1
 
 multiclass CALL<WebAssemblyRegClass vt> {
   def CALL_#vt : I<(outs vt:$dst), (ins global:$callee, variable_ops),
                    [(set vt:$dst, (WebAssemblycall1 (WebAssemblywrapper tglobaladdr:$callee)))],
                    "call    \t$dst, $callee">;
   def CALL_INDIRECT_#vt : I<(outs vt:$dst), (ins I32:$callee, variable_ops),
                             [(set vt:$dst, (WebAssemblycall1 I32:$callee))],

lib/Target/WebAssembly/WebAssemblyInstrInfo.cpp

 using namespace llvm;
 
 #define DEBUG_TYPE "wasm-instr-info"
 
 #define GET_INSTRINFO_CTOR_DTOR
 #include "WebAssemblyGenInstrInfo.inc"
 
 WebAssemblyInstrInfo::WebAssemblyInstrInfo(const WebAssemblySubtarget &STI)
-    : RI(STI.getTargetTriple()) {}
+    : WebAssemblyGenInstrInfo(WebAssembly::ADJCALLSTACKDOWN,
+                              WebAssembly::ADJCALLSTACKUP),
+      RI(STI.getTargetTriple()) {}
 
 void WebAssemblyInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
                                        MachineBasicBlock::iterator I,
                                        DebugLoc DL, unsigned DestReg,
                                        unsigned SrcReg, bool KillSrc) const {
   const TargetRegisterClass *RC =
       MBB.getParent()->getRegInfo().getRegClass(SrcReg);
 

lib/Target/WebAssembly/WebAssemblyRegStackify.cpp

#ifndef NDEBUG
   // Verify that pushes and pops are performed in FIFO order.
   SmallVector<unsigned, 0> Stack;
   for (MachineBasicBlock &MBB : MF) {
     for (MachineInstr &MI : MBB) {
       for (MachineOperand &MO : reverse(MI.explicit_operands())) {
         if (!MO.isReg())
           continue;
         unsigned VReg = MO.getReg();
+        if (!TargetRegisterInfo::isVirtualRegister(VReg))
+          continue;
 
         if (MFI.isVRegStackified(VReg)) {
           if (MO.isDef())
             Stack.push_back(VReg);
           else
             assert(Stack.pop_back_val() == VReg);
         }
       }

lib/Target/WebAssembly/WebAssemblyRegisterInfo.h

public:
   const MCPhysReg *getCalleeSavedRegs(const MachineFunction *MF) const override;
   BitVector getReservedRegs(const MachineFunction &MF) const override;
   void eliminateFrameIndex(MachineBasicBlock::iterator MI, int SPAdj,
                            unsigned FIOperandNum,
                            RegScavenger *RS = nullptr) const override;
 
   // Debug information queries.
   unsigned getFrameRegister(const MachineFunction &MF) const override;
+
+  const TargetRegisterClass *
+  getPointerRegClass(const MachineFunction &MF,
+                     unsigned Kind = 0) const override;
 };
 
 } // end namespace llvm
 
 #endif

lib/Target/WebAssembly/WebAssemblyRegisterInfo.cpp

 WebAssemblyRegisterInfo::getFrameRegister(const MachineFunction &MF) const {
   static const unsigned Regs[2][2] = {
       /*            !isArch64Bit       isArch64Bit      */
       /* !hasFP */ {WebAssembly::SP32, WebAssembly::SP64},
       /*  hasFP */ {WebAssembly::FP32, WebAssembly::FP64}};
   const WebAssemblyFrameLowering *TFI = getFrameLowering(MF);
   return Regs[TFI->hasFP(MF)][TT.isArch64Bit()];
 }
+
+const TargetRegisterClass *
+WebAssemblyRegisterInfo::getPointerRegClass(const MachineFunction &MF,
+                                            unsigned Kind) const {
+  assert(Kind == 0 && "Only one kind of pointer on WebAssembly");
+  if (MF.getSubtarget<WebAssemblySubtarget>().hasAddr64())
+    return &WebAssembly::I64RegClass;
+  return &WebAssembly::I32RegClass;
+}

test/CodeGen/WebAssembly/varargs.ll

+; RUN: llc < %s -asm-verbose=false | FileCheck %s
+
+; Test varargs constructs.
+
+target datalayout = "e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32-unknown-unknown"
+
+; Test va_start.
+
+; TODO: Test va_start.
+
+;define void @start(i8** %ap, ...) #0 {
+;entry:
+;  %0 = bitcast i8** %ap to i8*
+;  call void @llvm.va_start(i8* %0)
+;  ret void
+;}
+
+; Test va_end.
+
+; CHECK-LABEL: end:
+; CHECK-NEXT: .param i32{{$}}
+; CHECK-NEXT: return{{$}}
+define void @end(i8** %ap) #0 {
+entry:
+  %0 = bitcast i8** %ap to i8*
+  call void @llvm.va_end(i8* %0)
+  ret void
+}
+
+; Test va_copy.
+
+; CHECK-LABEL: copy:
+; CHECK-NEXT: .param i32, i32{{$}}
+; CHECK-NEXT: i32.load  $push0=, $1{{$}}
+; CHECK-NEXT: i32.store $discard=, $0, $pop0{{$}}
+; CHECK-NEXT: return{{$}}
+define void @copy(i8** %ap, i8** %bp) #0 {
+entry:
+  %0 = bitcast i8** %ap to i8*
+  %1 = bitcast i8** %bp to i8*
+  call void @llvm.va_copy(i8* %0, i8* %1)
+  ret void
+}
+
+; Test va_arg with an i8 argument.
+
+; CHECK-LABEL: arg_i8:
+; CHECK-NEXT: .param     i32{{$}}
+; CHECK-NEXT: .result    i32{{$}}
+; CHECK-NEXT: .local     i32{{$}}
+; CHECK-NEXT: i32.load   $1=, $0{{$}}
+; CHECK-NEXT: i32.const  $push0=, 4{{$}}
+; CHECK-NEXT: i32.add    $push1=, $1, $pop0{{$}}
+; CHECK-NEXT: i32.store  $discard=, $0, $pop1{{$}}
+; CHECK-NEXT: i32.load   $push2=, $1{{$}}
+; CHECK-NEXT: return     $pop2{{$}}
+define i8 @arg_i8(i8** %ap) #0 {
+entry:
+  %t = va_arg i8** %ap, i8
+  ret i8 %t
+}
+
+; Test va_arg with an i32 argument.
+
+; CHECK-LABEL: arg_i32:
+; CHECK-NEXT: .param     i32{{$}}
+; CHECK-NEXT: .result    i32{{$}}
+; CHECK-NEXT: .local     i32{{$}}
+; CHECK-NEXT: i32.load   $push0=, $0{{$}}
+; CHECK-NEXT: i32.const  $push1=, 3{{$}}
+; CHECK-NEXT: i32.add    $push2=, $pop0, $pop1{{$}}
+; CHECK-NEXT: i32.const  $push3=, -4{{$}}
+; CHECK-NEXT: i32.and    $1=, $pop2, $pop3{{$}}
+; CHECK-NEXT: i32.const  $push4=, 4{{$}}
+; CHECK-NEXT: i32.add    $push5=, $1, $pop4{{$}}
+; CHECK-NEXT: i32.store  $discard=, $0, $pop5{{$}}
+; CHECK-NEXT: i32.load   $push6=, $1{{$}}
+; CHECK-NEXT: return     $pop6{{$}}
+define i32 @arg_i32(i8** %ap) #0 {
+entry:
+  %t = va_arg i8** %ap, i32
+  ret i32 %t
+}
+
+; Test va_arg with an i128 argument.
+
+; CHECK-LABEL: arg_i128:
+; CHECK-NEXT: .param i32, i32{{$}}
+; CHECK-NEXT: .local
+; CHECK: i32.and
+; CHECK: i64.load
+; CHECK: i64.load
+; CHECK: return{{$}}
+define i128 @arg_i128(i8** %ap) #0 {
+entry:
+  %t = va_arg i8** %ap, i128
+  ret i128 %t
+}
+
+; Test a varargs call with no actual arguments.
+
+declare void @callee(...) #0
+
+; CHECK-LABEL: caller_none:
+; CHECK-NEXT: call callee{{$}}
+; CHECK-NEXT: return{{$}}
+define void @caller_none() #0 {
+  call void (...) @callee()
+  ret void
+}
+
+; TODO: Test a varargs call with actual arguments.
+
+;define void @caller_some() #0 {
+;  call void (...) @callee(i32 0, double 2.0)
+;  ret void
+;}
+
+declare void @llvm.va_start(i8*) #2
+declare void @llvm.va_end(i8*) #2
+declare void @llvm.va_copy(i8*, i8*) #2
+
+attributes #0 = { norecurse nounwind "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="generic" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { nounwind }