[NVPTX] Unify vectorization of load/stores of aggregate arguments and return values.

[tra]

Summary

Original code only used vector loads/stores for explicit vector arguments.
It could also do more loads/stores than necessary (e.g v5f32 would
touch 8 f32 values). Aggregate types were loaded one element at a time,
even the vectors contained within.

This change attempts to generalize (and simplify) parameter space
loads/stores so that vector loads/stores can be used more broadly.
Functionality of the patch has been verified by compiling thrust
test suite and manually checking the differences between PTX
generated by llvm with and without the patch.

General algorithm:

• ComputePTXValueVTs() flattens input/output argument into a flat list of scalars to load/store and returns their types and offsets.
• VectorizePTXValueVTs() uses that data to create vectorization plan which returns an array of flags marking boundaries of vectorized load/stores. Scalars are represented as 1-element vectors.
• Code that generates loads/stores implements a simple state machine that constructs a vector according to the plan.

NOTE: the tests for this patch have found a problem with the way we load/store unaligned fields. This will be addressed separately.

[jlebar]

lib/Target/NVPTX/NVPTXISelLowering.cpp

190

All right, now that I understand this, it's doing something deceptively simple. Perhaps:

Check whether we can merge loads/stores of some of the pieces of a flattened function parameter or return value into a single vector load/store.

The flattened parameter is represented as a list of EVTs and offsets, and the whole structure is aligned to ParamAlignment. This function determines whether we can load/store pieces of the parameter starting at index Idx using a single vectorized op of size AccessSize. If so, it returns the number of param pieces covered by the vector op. Otherwise, it returns 1.

Perhaps "CanVectorizeAt" is not a great name. I think we want "Param" in the name. Maybe CanMergeParamLoadStoresStartingAt? Ideally we'd also pick a name that doesn't sound like this returns a boolean, but I don't have a suggestion.

208

Actually I'm not sure why we bail here if the sizes are equal. Don't we want to vectorize {i32, i32}, even if AccessSize is 4 bytes?

217

Can we move this right below the NumElts computation? Long head-scratcher trying to figure out how we knew that the division above is exact. :)

220

Nit, "2- and 4-element vector ops"

224

Maybe it would make sense to write this as

for (unsigned j = Idx + 1; j < Idx + NumElts; ++j)  { ... }

237

Please add a brief comment. I'm also not sure about the name, for the same reasons as CanVectorizeAt -- it's very generic. When we change the enumeration name, we should probably also change PTX_LDST to something else.

248

Perhaps

Computes whether and how we can vectorize the loads/stores of a flattened function parameter or return value.

The flattened parameter is represented as the list of ValueVTs and Offsets, and is aligned to ParamAlignment bytes. We return a vector of the same size as ValueVTs indicating how each piece should be loaded/stored (i.e. as a scalar, or as part of a vector load/store).

252

Can we return the SmallVector instead of using an outparam? We'll get RVO, so it's the same to the compiler.

1417

Does this comment still hold?

1456

Is this the right name? We're dealing with function params, not a return value, I think?

1461

StoreOperands or StNodeOperands? "Ops" is a bad abbreviation for "operands". :) And it looks like this is specifically for stores?

1461

Can we assert somewhere that this vector is empty when we're done with the loop?

1463

"New store."?

1480

s/load-stores/stores/?

1507

s#load/store#store#?

2400

This looks oddly familiar from the parameter handling above. We really can't factor it out without making this into more of a hairball? If not, same comments apply down here.

[tra]

Addressed Justin's comments.

lib/Target/NVPTX/NVPTXISelLowering.cpp

208

If EltSize == AccessSize, then we can only do 1-element load/store and there's nothing to vectorize.
{i32,i32} needs AccessSize of at least 8 so we can fold it into ld/st.v2 .

2400

Alas, I've found no good way to generalize this further. We have 4 cases that we need to vectorize:

• Load of function args
• Storing return value.
• Storing function args before a call
• Loading call result.

Loads are currently rather different, so there's not much to factor out between them.
Stores are indeed similar, but create different nodes that take different set of operands. Combining them would probably make things more confusing that it already is.

[jlebar]

lib/Target/NVPTX/NVPTXISelLowering.cpp

190

Nit, please add blank line to separate paragraphs.

253

Newline to separate para

[tra]

Fixed last two nits.

lib/Target/NVPTX/NVPTXISelLowering.cpp

for (unsigned i = 0, e = TempVTs.size(); i != e; ++i) {
     else {
       ValueVTs.push_back(VT);
       if (Offsets)
         Offsets->push_back(Off);
     }
   }
 }
 
+// Check whether we can vectorize argument elements starting at \p
+// Idx, using \p AccessSize large accesses. Returns number of elements
+// we can process in a single ld/st operation (1 for scalar, 2 or 4
+// for vectors).
+static unsigned CanVectorizeAt(unsigned Idx, uint32_t AccessSize,
+                               const SmallVectorImpl<EVT> &ValueVTs,
+                               const SmallVectorImpl<uint64_t> &Offsets,
+                               unsigned ParamAlignment) {
+  assert(isPowerOf2_32(AccessSize) && "must be a power of 2!");
+
+  // Can't vectorize if param alignment is not sufficient.
+  if (AccessSize > ParamAlignment)
+    return 1;
+  // Can't vectorize if offset is not aligned.
+  if (Offsets[Idx] & (AccessSize - 1))
+    return 1;
+
+  EVT EltVT = ValueVTs[Idx];
+  unsigned EltSize = EltVT.getStoreSize();
+
+  // Element is too large to vectorize.
+  if (EltSize >= AccessSize)
+    return 1;
+
+  unsigned NumElts = AccessSize / EltSize;
+  // We don't have enough elements to vectorize.
+  if (Idx + NumElts > ValueVTs.size())
+    return 1;
+
+  // Can't vectorize if AccessBytes if not a multiple of EltSize.
+  if (AccessSize != EltSize * NumElts)
+    return 1;
+
+  // PTX ISA can only deal with 2 and 4 element vector ops.
+  if (NumElts != 4 && NumElts != 2)
+    return 1;
+
+  for (unsigned j = 1; j < NumElts; ++j) {
+    // Types do not match.
+    if (ValueVTs[Idx + j] != EltVT)
+      return 1;
+
+    // Elements are not contiguous.
+    if (Offsets[Idx + j] - Offsets[Idx + j - 1] != EltSize)
+      return 1;
+  }
+  // OK. We can vectorize ValueVTs[i..i+NumElts)
+  return NumElts;
+}
+
+enum PtxVectorInfo {
+  PTX_LDST_VECTORIZED = 0x0, // Middle elements of a vector.
+  PTX_LDST_BEGIN = 0x1,      // First element of the vector.
+  PTX_LDST_END = 0x2,        // Last element of the vector.
+  // Scalar is effectively a 1-element vector.
+  PTX_LDST_SCALAR = PTX_LDST_BEGIN | PTX_LDST_END
+};
+
+// For each pair of elements in \p ValueVTs, \p Offsets produced by
+// ComputePTXValueVTs() for a function argument, and the argument
+// alignment \p ParamAlignment, returns \p VectorInfo with each
+// element indicating whether its load/store can be vectorized.
+static void VectorizePTXValueVTs(const SmallVectorImpl<EVT> &ValueVTs,
+                                 const SmallVectorImpl<uint64_t> &Offsets,
+                                 unsigned ParamAlignment,
+                                 SmallVectorImpl<PtxVectorInfo> &VectorInfo) {
+  // Set vector size to match ValueVTs and mark all elements as
+  // scalars by default.
+  VectorInfo.assign(ValueVTs.size(), PTX_LDST_SCALAR);
+
+  // Check what we can vectorize using 128/64/32-bit accesses.
+  for (int I = 0, E = ValueVTs.size(); I != E; ++I) {
+    // Skip elements we've already processed.
+    assert(VectorInfo[I] == PTX_LDST_SCALAR && "Unexpected vector info state.");
+    for (unsigned AccessSize : {16, 8, 4, 2}) {
+      unsigned NumElts =
+          CanVectorizeAt(I, AccessSize, ValueVTs, Offsets, ParamAlignment);
+      // Mark vectorized elements.
+      switch (NumElts) {
+      default:
+        llvm_unreachable("Unexpected return value");
+      case 1:
+        // Can't vectorize using this size, try next smaller size.
+        continue;
+      case 2:
+        assert(I + 1 < E && "Not enough elements.");
+        VectorInfo[I] = PTX_LDST_BEGIN;
+        VectorInfo[I + 1] = PTX_LDST_END;
+        I += 1;
+        break;
+      case 4:
+        assert(I + 3 < E && "Not enough elements.");
+        VectorInfo[I] = PTX_LDST_BEGIN;
+        VectorInfo[I + 1] = PTX_LDST_VECTORIZED;
+        VectorInfo[I + 2] = PTX_LDST_VECTORIZED;
+        VectorInfo[I + 3] = PTX_LDST_END;
+        I += 3;
+        break;
+      }
+      // Break out of the inner loop because we've already succeeded
+      // using largest possible AccessSize.
+      break;
+    }
+  }
+}
+
 // NVPTXTargetLowering Constructor.
 NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
                                          const NVPTXSubtarget &STI)
     : TargetLowering(TM), nvTM(&TM), STI(STI) {
   // always lower memset, memcpy, and memmove intrinsics to load/store
   // instructions, rather
   // then generating calls to memset, mempcy or memmove.
   MaxStoresPerMemset = (unsigned) 0xFFFFFFFF;
SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
   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;
   bool &isTailCall = CLI.IsTailCall;
   ArgListTy &Args = CLI.getArgs();
-  Type *retTy = CLI.RetTy;
+  Type *RetTy = CLI.RetTy;
   ImmutableCallSite *CS = CLI.CS;
+  const DataLayout &DL = DAG.getDataLayout();
 
   bool isABI = (STI.getSmVersion() >= 20);
   assert(isABI && "Non-ABI compilation is not supported");
   if (!isABI)
     return Chain;
-  MachineFunction &MF = DAG.getMachineFunction();
-  const Function *F = MF.getFunction();
-  auto &DL = MF.getDataLayout();
 
   SDValue tempChain = Chain;
-  Chain = DAG.getCALLSEQ_START(Chain,
-                               DAG.getIntPtrConstant(uniqueCallSite, dl, true),
-                               dl);
+  Chain = DAG.getCALLSEQ_START(
+      Chain, DAG.getIntPtrConstant(uniqueCallSite, dl, true), dl);
   SDValue InFlag = Chain.getValue(1);
 
   unsigned paramCount = 0;
   // Args.size() and Outs.size() need not match.
   // Outs.size() will be larger
   //   * if there is an aggregate argument with multiple fields (each field
   //     showing up separately in Outs)
   //   * if there is a vector argument with more than typical vector-length
   //     elements (generally if more than 4) where each vector element is
   //     individually present in Outs.
   // So a different index should be used for indexing into Outs/OutVals.
   // See similar issue in LowerFormalArguments.
   unsigned OIdx = 0;
   // Declare the .params or .reg need to pass values
   // to the function
   for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
     EVT VT = Outs[OIdx].VT;
     Type *Ty = Args[i].Ty;
 
     if (!Outs[OIdx].Flags.isByVal()) {
-      if (Ty->isAggregateType()) {
-        // aggregate
-        SmallVector<EVT, 16> vtparts;
-        SmallVector<uint64_t, 16> Offsets;
-        ComputePTXValueVTs(*this, DAG.getDataLayout(), Ty, vtparts, &Offsets,
-                           0);
-
-        unsigned align =
-            getArgumentAlignment(Callee, CS, Ty, paramCount + 1, DL);
-        // declare .param .align <align> .b8 .param<n>[<size>];
-        unsigned sz = DL.getTypeAllocSize(Ty);
-        SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
-        SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, dl,
-                                                             MVT::i32),
-                                      DAG.getConstant(paramCount, dl, MVT::i32),
-                                      DAG.getConstant(sz, dl, MVT::i32),
-                                      InFlag };
+      SmallVector<EVT, 16> VTs;
+      ComputePTXValueVTs(*this, DL, Ty, VTs, &Offsets);
+      unsigned ArgAlign =
+      unsigned AllocSize = DL.getTypeAllocSize(Ty);
+      bool NeedAlign;  // Does argument declaration specify alignment?
+      if (Ty->isAggregateType() || Ty->isVectorTy()) {
+        SDValue DeclareParamOps[] = {
+            Chain, DAG.getConstant(ArgAlign, dl, MVT::i32),
+            DAG.getConstant(AllocSize, dl, MVT::i32), InFlag};
         Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
                             DeclareParamOps);
-        InFlag = Chain.getValue(1);
-        for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
-          EVT elemtype = vtparts[j];
-          unsigned ArgAlign = GreatestCommonDivisor64(align, Offsets[j]);
-          if (elemtype.isInteger() && (sz < 8))
-            sz = 8;
-          SDValue StVal = OutVals[OIdx];
-          if (elemtype.getSizeInBits() < 16) {
-            StVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, StVal);
-          }
-          SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
-          SDValue CopyParamOps[] = { Chain,
-                                     DAG.getConstant(paramCount, dl, MVT::i32),
-                                     DAG.getConstant(Offsets[j], dl, MVT::i32),
-                                     StVal, InFlag };
-          Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
-                                          CopyParamVTs, CopyParamOps,
-                                          elemtype, MachinePointerInfo(),
-                                          ArgAlign);
-          InFlag = Chain.getValue(1);
-          ++OIdx;
-        }
-        if (vtparts.size() > 0)
-          --OIdx;
-        ++paramCount;
-        continue;
-      }
-      if (Ty->isVectorTy()) {
-        EVT ObjectVT = getValueType(DL, Ty);
-        unsigned align =
-            getArgumentAlignment(Callee, CS, Ty, paramCount + 1, DL);
-        // declare .param .align <align> .b8 .param<n>[<size>];
-        unsigned sz = DL.getTypeAllocSize(Ty);
-        SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
-        SDValue DeclareParamOps[] = { Chain,
-                                      DAG.getConstant(align, dl, MVT::i32),
-                                      DAG.getConstant(paramCount, dl, MVT::i32),
-                                      DAG.getConstant(sz, dl, MVT::i32),
-                                      InFlag };
-        Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
-                            DeclareParamOps);
-        InFlag = Chain.getValue(1);
-        unsigned NumElts = ObjectVT.getVectorNumElements();
-        EVT EltVT = ObjectVT.getVectorElementType();
-        EVT MemVT = EltVT;
-        bool NeedExtend = false;
-        if (EltVT.getSizeInBits() < 16) {
-          NeedExtend = true;
-          EltVT = MVT::i16;
-        }
-
-        // V1 store
-        if (NumElts == 1) {
-          SDValue Elt = OutVals[OIdx++];
-          if (NeedExtend)
-            Elt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt);
+        NeedAlign = true;
+      } else {
+        if ((VT.isInteger() || VT.isFloatingPoint()) && AllocSize < 4) {
+          // PTX ABI requires integral types to be at least 32 bits in
+          // size. FP16 is loaded/stored using i16, so it's handled
+          // here as well.
+          AllocSize = 4;
+        }
+        SDValue DeclareScalarParamOps[] = {
+            Chain, DAG.getConstant(paramCount, dl, MVT::i32),
+            DAG.getConstant(AllocSize * 8, dl, MVT::i32),
+            DAG.getConstant(0, dl, MVT::i32), InFlag};
+        Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs,
+                            DeclareScalarParamOps);
+        NeedAlign = false;
 
-          SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
-          SDValue CopyParamOps[] = { Chain,
-                                     DAG.getConstant(paramCount, dl, MVT::i32),
-                                     DAG.getConstant(0, dl, MVT::i32), Elt,
-                                     InFlag };
-          Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
-                                          CopyParamVTs, CopyParamOps,
-                                          MemVT, MachinePointerInfo());
-          InFlag = Chain.getValue(1);
-        } else if (NumElts == 2) {
-          SDValue Elt0 = OutVals[OIdx++];
-          SDValue Elt1 = OutVals[OIdx++];
-          if (NeedExtend) {
-            Elt0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt0);
-            Elt1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt1);
-          }
+      // PTX Interoperability Guide 3.3(A): [Integer] Values shorter
+      // than 32-bits are sign extended or zero extended, depending on
+      // whether they are signed or unsigned types. This case applies
+      // only to scalar parameters and not to aggregate values.
+      bool ExtendIntegerRetVal =
+          Ty->isIntegerTy() && DL.getTypeAllocSizeInBits(Ty) < 32;
+
+      SmallVector<PtxVectorInfo, 16> VectorInfo;
+      VectorizePTXValueVTs(VTs, Offsets, ArgAlign, VectorInfo);
+      SmallVector<SDValue, 6> LdStOps;
+      for (unsigned j = 0, je = VTs.size(); j != je; ++j) {
+        // New load/store.
+        if (VectorInfo[j] & PTX_LDST_BEGIN) {
+          assert(LdStOps.empty() && "Orphaned operand list");
+          LdStOps.push_back(Chain);
+          LdStOps.push_back(DAG.getConstant(paramCount, dl, MVT::i32));
+          LdStOps.push_back(DAG.getConstant(Offsets[j], dl, MVT::i32));
 
-          SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
-          SDValue CopyParamOps[] = { Chain,
-                                     DAG.getConstant(paramCount, dl, MVT::i32),
-                                     DAG.getConstant(0, dl, MVT::i32), Elt0,
-                                     Elt1, InFlag };
-          Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParamV2, dl,
-                                          CopyParamVTs, CopyParamOps,
-                                          MemVT, MachinePointerInfo());
-          InFlag = Chain.getValue(1);
-        } else {
-          unsigned curOffset = 0;
-          // V4 stores
-          // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
-          // the
-          // vector will be expanded to a power of 2 elements, so we know we can
-          // always round up to the next multiple of 4 when creating the vector
-          // stores.
-          // e.g.  4 elem => 1 st.v4
-          //       6 elem => 2 st.v4
-          //       8 elem => 2 st.v4
-          //      11 elem => 3 st.v4
-          unsigned VecSize = 4;
-          if (EltVT.getSizeInBits() == 64)
-            VecSize = 2;
-
-          // This is potentially only part of a vector, so assume all elements
-          // are packed together.
-          unsigned PerStoreOffset = MemVT.getStoreSizeInBits() / 8 * VecSize;
-
-          for (unsigned i = 0; i < NumElts; i += VecSize) {
-            // Get values
-            SDValue StoreVal;
-            SmallVector<SDValue, 8> Ops;
-            Ops.push_back(Chain);
-            Ops.push_back(DAG.getConstant(paramCount, dl, MVT::i32));
-            Ops.push_back(DAG.getConstant(curOffset, dl, MVT::i32));
-
-            unsigned Opc = NVPTXISD::StoreParamV2;
-
-            StoreVal = OutVals[OIdx++];
-            if (NeedExtend)
-              StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
-            Ops.push_back(StoreVal);
-
-            if (i + 1 < NumElts) {
-              StoreVal = OutVals[OIdx++];
-              if (NeedExtend)
-                StoreVal =
-                    DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
-            } else {
-              StoreVal = DAG.getUNDEF(EltVT);
-            }
-            Ops.push_back(StoreVal);
+        EVT EltVT = VTs[j];
+        SDValue StVal = OutVals[OIdx];
+        if (ExtendIntegerRetVal) {
+          assert(VTs.size() == 1 && "Scalar can't have multiple parts.");
+          // zext/sext to i32
+          StVal = DAG.getNode(Outs[OIdx].Flags.isSExt() ? ISD::SIGN_EXTEND
+                                                        : ISD::ZERO_EXTEND,
+                              dl, MVT::i32, StVal);
+        } else if (EltVT.getSizeInBits() < 16) {
+          // Use 16-bit registers for small load-stores as it's the
+          // smallest general purpose register size supported by NVPTX.
+          StVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, StVal);
 
-            if (VecSize == 4) {
-              Opc = NVPTXISD::StoreParamV4;
-              if (i + 2 < NumElts) {
-                StoreVal = OutVals[OIdx++];
-                if (NeedExtend)
-                  StoreVal =
-                      DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
-              } else {
-                StoreVal = DAG.getUNDEF(EltVT);
-              }
-              Ops.push_back(StoreVal);
+        // Record the value to store.
+        LdStOps.push_back(StVal);
 
-              if (i + 3 < NumElts) {
-                StoreVal = OutVals[OIdx++];
-                if (NeedExtend)
-                  StoreVal =
-                      DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
-              } else {
-                StoreVal = DAG.getUNDEF(EltVT);
-              }
-              Ops.push_back(StoreVal);
-            }
+        if (VectorInfo[j] & PTX_LDST_END) {
+          unsigned NumElts = LdStOps.size() - 3;
+          NVPTXISD::NodeType Op;
+          switch (NumElts) {
+          case 1:
+            Op = NVPTXISD::StoreParam;
+            break;
+          case 2:
+            Op = NVPTXISD::StoreParamV2;
+            break;
+          case 4:
+            Op = NVPTXISD::StoreParamV4;
+            break;
+          default:
+            llvm_unreachable("Invalid vector info.");
 
-            Ops.push_back(InFlag);
+          LdStOps.push_back(InFlag);
 
-            SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
-            Chain = DAG.getMemIntrinsicNode(Opc, dl, CopyParamVTs, Ops,
-                                            MemVT, MachinePointerInfo());
-            InFlag = Chain.getValue(1);
-            curOffset += PerStoreOffset;
-          }
-        }
-        ++paramCount;
+          // Adjust type of load/store op if we've extended the scalar
+          // return value.
+          EVT TheStoreType = ExtendIntegerRetVal ? MVT::i32 : VTs[j];
+          unsigned EltAlign =
+              NeedAlign ? GreatestCommonDivisor64(ArgAlign, Offsets[j]) : 0;
+
+          Chain = DAG.getMemIntrinsicNode(
+              Op, dl, DAG.getVTList(MVT::Other, MVT::Glue), LdStOps,
+              TheStoreType, MachinePointerInfo(), EltAlign);
+
+          // Cleanup.
+          LdStOps.clear();
+        ++OIdx;
+      if (VTs.size() > 0)
         --OIdx;
-        continue;
-      }
-      // Plain scalar
-      // for ABI,    declare .param .b<size> .param<n>;
-      unsigned sz = VT.getSizeInBits();
-      bool needExtend = false;
-      if (VT.isInteger()) {
-        if (sz < 16)
-          needExtend = true;
-        if (sz < 32)
-          sz = 32;
-      } else if (VT.isFloatingPoint() && sz < 32)
-        // PTX ABI requires all scalar parameters to be at least 32
-        // bits in size.  fp16 normally uses .b16 as its storage type
-        // in PTX, so its size must be adjusted here, too.
-        sz = 32;
-      SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
-      SDValue DeclareParamOps[] = { Chain,
-                                    DAG.getConstant(paramCount, dl, MVT::i32),
-                                    DAG.getConstant(sz, dl, MVT::i32),
-                                    DAG.getConstant(0, dl, MVT::i32), InFlag };
-      Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs,
-                          DeclareParamOps);
-      InFlag = Chain.getValue(1);
-      SDValue OutV = OutVals[OIdx];
-      if (needExtend) {
-        // zext/sext i1 to i16
-        unsigned opc = ISD::ZERO_EXTEND;
-        if (Outs[OIdx].Flags.isSExt())
-          opc = ISD::SIGN_EXTEND;
-        OutV = DAG.getNode(opc, dl, MVT::i16, OutV);
-      }
-      SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
-      SDValue CopyParamOps[] = { Chain,
-                                 DAG.getConstant(paramCount, dl, MVT::i32),
-                                 DAG.getConstant(0, dl, MVT::i32), OutV,
-                                 InFlag };
-
-      unsigned opcode = NVPTXISD::StoreParam;
-      if (Outs[OIdx].Flags.isZExt() && VT.getSizeInBits() < 32)
-        opcode = NVPTXISD::StoreParamU32;
-      else if (Outs[OIdx].Flags.isSExt() && VT.getSizeInBits() < 32)
-        opcode = NVPTXISD::StoreParamS32;
-      Chain = DAG.getMemIntrinsicNode(opcode, dl, CopyParamVTs, CopyParamOps,
-                                      VT, MachinePointerInfo());
-
-      InFlag = Chain.getValue(1);
       ++paramCount;
       continue;
     }
-    // struct or vector
-    SmallVector<EVT, 16> vtparts;
+
+    // ByVal arguments
+    SmallVector<EVT, 16> VTs;
     SmallVector<uint64_t, 16> Offsets;
     auto *PTy = dyn_cast<PointerType>(Args[i].Ty);
     assert(PTy && "Type of a byval parameter should be pointer");
-    ComputePTXValueVTs(*this, DAG.getDataLayout(), PTy->getElementType(),
-                       vtparts, &Offsets, 0);
+    ComputePTXValueVTs(*this, DL, PTy->getElementType(), VTs, &Offsets, 0);
 
     // declare .param .align <align> .b8 .param<n>[<size>];
     unsigned sz = Outs[OIdx].Flags.getByValSize();
     SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
     unsigned ArgAlign = Outs[OIdx].Flags.getByValAlign();
     // The ByValAlign in the Outs[OIdx].Flags is alway set at this point,
     // so we don't need to worry about natural alignment or not.
     // See TargetLowering::LowerCallTo().
 
     // Enforce minumum alignment of 4 to work around ptxas miscompile
     // for sm_50+. See corresponding alignment adjustment in
     // emitFunctionParamList() for details.
     if (ArgAlign < 4)
       ArgAlign = 4;
     SDValue DeclareParamOps[] = {Chain, DAG.getConstant(ArgAlign, dl, MVT::i32),
                                  DAG.getConstant(paramCount, dl, MVT::i32),
                                  DAG.getConstant(sz, dl, MVT::i32), InFlag};
     Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
                         DeclareParamOps);
     InFlag = Chain.getValue(1);
-    for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
-      EVT elemtype = vtparts[j];
+    for (unsigned j = 0, je = VTs.size(); j != je; ++j) {
+      EVT elemtype = VTs[j];
       int curOffset = Offsets[j];
       unsigned PartAlign = GreatestCommonDivisor64(ArgAlign, curOffset);
-      auto PtrVT = getPointerTy(DAG.getDataLayout());
+      auto PtrVT = getPointerTy(DL);
       SDValue srcAddr = DAG.getNode(ISD::ADD, dl, PtrVT, OutVals[OIdx],
                                     DAG.getConstant(curOffset, dl, PtrVT));
       SDValue theVal = DAG.getLoad(elemtype, dl, tempChain, srcAddr,
                                    MachinePointerInfo(), PartAlign);
       if (elemtype.getSizeInBits() < 16) {
         theVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, theVal);
       }
       SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
   }
 
   GlobalAddressSDNode *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode());
   unsigned retAlignment = 0;
 
   // Handle Result
   if (Ins.size() > 0) {
     SmallVector<EVT, 16> resvtparts;
-    ComputeValueVTs(*this, DL, retTy, resvtparts);
+    ComputeValueVTs(*this, DL, RetTy, resvtparts);
 
     // Declare
     //  .param .align 16 .b8 retval0[<size-in-bytes>], or
     //  .param .b<size-in-bits> retval0
-    unsigned resultsz = DL.getTypeAllocSizeInBits(retTy);
+    unsigned resultsz = DL.getTypeAllocSizeInBits(RetTy);
     // Emit ".param .b<size-in-bits> retval0" instead of byte arrays only for
     // these three types to match the logic in
     // NVPTXAsmPrinter::printReturnValStr and NVPTXTargetLowering::getPrototype.
     // Plus, this behavior is consistent with nvcc's.
-    if (retTy->isFloatingPointTy() || retTy->isIntegerTy() ||
-        retTy->isPointerTy()) {
+    if (RetTy->isFloatingPointTy() || RetTy->isIntegerTy() ||
+        RetTy->isPointerTy()) {
       // Scalar needs to be at least 32bit wide
       if (resultsz < 32)
         resultsz = 32;
       SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
       SDValue DeclareRetOps[] = { Chain, DAG.getConstant(1, dl, MVT::i32),
                                   DAG.getConstant(resultsz, dl, MVT::i32),
                                   DAG.getConstant(0, dl, MVT::i32), InFlag };
       Chain = DAG.getNode(NVPTXISD::DeclareRet, dl, DeclareRetVTs,
                           DeclareRetOps);
       InFlag = Chain.getValue(1);
     } else {
-      retAlignment = getArgumentAlignment(Callee, CS, retTy, 0, DL);
+      retAlignment = getArgumentAlignment(Callee, CS, RetTy, 0, DL);
       SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
       SDValue DeclareRetOps[] = { Chain,
                                   DAG.getConstant(retAlignment, dl, MVT::i32),
                                   DAG.getConstant(resultsz / 8, dl, MVT::i32),
                                   DAG.getConstant(0, dl, MVT::i32), InFlag };
       Chain = DAG.getNode(NVPTXISD::DeclareRetParam, dl, DeclareRetVTs,
                           DeclareRetOps);
       InFlag = Chain.getValue(1);
     }
   }
 
   if (!Func) {
     // This is indirect function call case : PTX requires a prototype of the
     // form
     // proto_0 : .callprototype(.param .b32 _) _ (.param .b32 _);
     // to be emitted, and the label has to used as the last arg of call
     // instruction.
     // The prototype is embedded in a string and put as the operand for a
     // CallPrototype SDNode which will print out to the value of the string.
     SDVTList ProtoVTs = DAG.getVTList(MVT::Other, MVT::Glue);
     std::string Proto =
-        getPrototype(DAG.getDataLayout(), retTy, Args, Outs, retAlignment, CS);
+        getPrototype(DL, RetTy, Args, Outs, retAlignment, CS);
     const char *ProtoStr =
       nvTM->getManagedStrPool()->getManagedString(Proto.c_str())->c_str();
     SDValue ProtoOps[] = {
       Chain, DAG.getTargetExternalSymbol(ProtoStr, MVT::i32), InFlag,
     };
     Chain = DAG.getNode(NVPTXISD::CallPrototype, dl, ProtoVTs, ProtoOps);
     InFlag = Chain.getValue(1);
   }
SDValue PrototypeOps[] = { Chain,
                                DAG.getConstant(uniqueCallSite, dl, MVT::i32),
                                InFlag };
     Chain = DAG.getNode(NVPTXISD::Prototype, dl, PrototypeVTs, PrototypeOps);
     InFlag = Chain.getValue(1);
   }
 
   // Generate loads from param memory/moves from registers for result
   if (Ins.size() > 0) {
-    if (retTy && retTy->isVectorTy()) {
-      EVT ObjectVT = getValueType(DL, retTy);
-      unsigned NumElts = ObjectVT.getVectorNumElements();
-      EVT EltVT = ObjectVT.getVectorElementType();
-      assert(STI.getTargetLowering()->getNumRegisters(F->getContext(),
-                                                      ObjectVT) == NumElts &&
-             "Vector was not scalarized");
-      unsigned sz = EltVT.getSizeInBits();
-      bool needTruncate = sz < 8;
-
-      if (NumElts == 1) {
-        // Just a simple load
-        SmallVector<EVT, 4> LoadRetVTs;
-        if (EltVT == MVT::i1 || EltVT == MVT::i8) {
-          // If loading i1/i8 result, generate
-          //   load.b8 i16
-          //   if i1
-          //   trunc i16 to i1
-          LoadRetVTs.push_back(MVT::i16);
-        } else
-          LoadRetVTs.push_back(EltVT);
-        LoadRetVTs.push_back(MVT::Other);
-        LoadRetVTs.push_back(MVT::Glue);
-        SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
-                                DAG.getConstant(0, dl, MVT::i32), InFlag};
-        SDValue retval = DAG.getMemIntrinsicNode(
-            NVPTXISD::LoadParam, dl,
-            DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
-        Chain = retval.getValue(1);
-        InFlag = retval.getValue(2);
-        SDValue Ret0 = retval;
-        if (needTruncate)
-          Ret0 = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Ret0);
-        InVals.push_back(Ret0);
-      } else if (NumElts == 2) {
-        // LoadV2
-        SmallVector<EVT, 4> LoadRetVTs;
-        if (EltVT == MVT::i1 || EltVT == MVT::i8) {
-          // If loading i1/i8 result, generate
-          //   load.b8 i16
-          //   if i1
-          //   trunc i16 to i1
-          LoadRetVTs.push_back(MVT::i16);
-          LoadRetVTs.push_back(MVT::i16);
-        } else {
-          LoadRetVTs.push_back(EltVT);
-          LoadRetVTs.push_back(EltVT);
-        }
-        LoadRetVTs.push_back(MVT::Other);
-        LoadRetVTs.push_back(MVT::Glue);
-        SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
-                                DAG.getConstant(0, dl, MVT::i32), InFlag};
-        SDValue retval = DAG.getMemIntrinsicNode(
-            NVPTXISD::LoadParamV2, dl,
-            DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
-        Chain = retval.getValue(2);
-        InFlag = retval.getValue(3);
-        SDValue Ret0 = retval.getValue(0);
-        SDValue Ret1 = retval.getValue(1);
-        if (needTruncate) {
-          Ret0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret0);
-          InVals.push_back(Ret0);
-          Ret1 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret1);
-          InVals.push_back(Ret1);
-        } else {
-          InVals.push_back(Ret0);
-          InVals.push_back(Ret1);
-        }
-      } else {
-        // Split into N LoadV4
-        unsigned Ofst = 0;
-        unsigned VecSize = 4;
-        unsigned Opc = NVPTXISD::LoadParamV4;
-        if (EltVT.getSizeInBits() == 64) {
-          VecSize = 2;
-          Opc = NVPTXISD::LoadParamV2;
-        }
-        EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
-        for (unsigned i = 0; i < NumElts; i += VecSize) {
-          SmallVector<EVT, 8> LoadRetVTs;
-          if (EltVT == MVT::i1 || EltVT == MVT::i8) {
-            // If loading i1/i8 result, generate
-            //   load.b8 i16
-            //   if i1
-            //   trunc i16 to i1
-            for (unsigned j = 0; j < VecSize; ++j)
-              LoadRetVTs.push_back(MVT::i16);
-          } else {
-            for (unsigned j = 0; j < VecSize; ++j)
-              LoadRetVTs.push_back(EltVT);
-          }
-          LoadRetVTs.push_back(MVT::Other);
-          LoadRetVTs.push_back(MVT::Glue);
-          SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
-                                  DAG.getConstant(Ofst, dl, MVT::i32), InFlag};
-          SDValue retval = DAG.getMemIntrinsicNode(
-              Opc, dl, DAG.getVTList(LoadRetVTs),
-              LoadRetOps, EltVT, MachinePointerInfo());
-          if (VecSize == 2) {
-            Chain = retval.getValue(2);
-            InFlag = retval.getValue(3);
-          } else {
-            Chain = retval.getValue(4);
-            InFlag = retval.getValue(5);
-          }
-
-          for (unsigned j = 0; j < VecSize; ++j) {
-            if (i + j >= NumElts)
-              break;
-            SDValue Elt = retval.getValue(j);
-            if (needTruncate)
-              Elt = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
-            InVals.push_back(Elt);
-          }
-          Ofst += DL.getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
-        }
-      }
-    } else {
-      SmallVector<EVT, 16> VTs;
-      SmallVector<uint64_t, 16> Offsets;
-      auto &DL = DAG.getDataLayout();
-      ComputePTXValueVTs(*this, DL, retTy, VTs, &Offsets, 0);
-      assert(VTs.size() == Ins.size() && "Bad value decomposition");
-      unsigned RetAlign = getArgumentAlignment(Callee, CS, retTy, 0, DL);
-      for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
-        unsigned sz = VTs[i].getSizeInBits();
-        unsigned AlignI = GreatestCommonDivisor64(RetAlign, Offsets[i]);
-        bool needTruncate = false;
-        if (VTs[i].isInteger() && sz < 8) {
-          sz = 8;
-          needTruncate = true;
-        }
-
-        SmallVector<EVT, 4> LoadRetVTs;
-        EVT TheLoadType = VTs[i];
-        if (retTy->isIntegerTy() && DL.getTypeAllocSizeInBits(retTy) < 32) {
-          // This is for integer types only, and specifically not for
-          // aggregates.
-          LoadRetVTs.push_back(MVT::i32);
-          TheLoadType = MVT::i32;
-          needTruncate = true;
-        } else if (sz < 16) {
-          // If loading i1/i8 result, generate
-          //   load i8 (-> i16)
-          //   trunc i16 to i1/i8
-
-          // FIXME: Do we need to set needTruncate to true here, too?  We could
-          // not figure out what this branch is for in D17872, so we left it
-          // alone.  The comment above about loading i1/i8 may be wrong, as the
-          // branch above seems to cover integers of size < 32.
-          LoadRetVTs.push_back(MVT::i16);
-        } else
-          LoadRetVTs.push_back(Ins[i].VT);
-        LoadRetVTs.push_back(MVT::Other);
-        LoadRetVTs.push_back(MVT::Glue);
+    ComputePTXValueVTs(*this, DL, RetTy, VTs, &Offsets, 0);
+
+    unsigned RetAlign = getArgumentAlignment(Callee, CS, RetTy, 0, DL);
+    SmallVector<PtxVectorInfo, 16> VectorInfo;
+    VectorizePTXValueVTs(VTs, Offsets, RetAlign, VectorInfo);
+
+    SmallVector<EVT, 6> LoadVTs;
+    int VecIdx = -1;  // Index of the first element of the vector.
+
+    // PTX Interoperability Guide 3.3(A): [Integer] Values shorter than
+    // 32-bits are sign extended or zero extended, depending on whether
+    // they are signed or unsigned types.
+    bool ExtendIntegerRetVal =
+        RetTy->isIntegerTy() && DL.getTypeAllocSizeInBits(RetTy) < 32;
+
+      EVT EltType = Ins[i].VT;
+      unsigned EltAlign = GreatestCommonDivisor64(RetAlign, Offsets[i]);
+      if (ExtendIntegerRetVal) {
+        EltType = MVT::i32;
+      } else if (TheLoadType.getSizeInBits() < 16) {
+        if (VTs[i].isInteger())
+          needTruncate = true;
+        EltType = MVT::i16;
+      }
+
+      // Record index of the very first element of the vector.
+      if (VectorInfo[i] & PTX_LDST_BEGIN) {
+        assert(VecIdx == -1 && LoadVTs.empty() && "Orphaned operand list.");
+        VecIdx = i;
+      }
+
+      LoadVTs.push_back(EltType);
+
+      if (VectorInfo[i] & PTX_LDST_END) {
+        unsigned NumElts = LoadVTs.size();
+        LoadVTs.push_back(MVT::Other);
+        LoadVTs.push_back(MVT::Glue);
+        NVPTXISD::NodeType Op;
+        switch (NumElts) {
+        case 1:
+          Op = NVPTXISD::LoadParam;
+          break;
+        case 2:
+          Op = NVPTXISD::LoadParamV2;
+          break;
+        case 4:
+          Op = NVPTXISD::LoadParamV4;
+          break;
+        default:
+          llvm_unreachable("Invalid vector info.");
+        }
 
-        SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
-                                DAG.getConstant(Offsets[i], dl, MVT::i32),
-                                InFlag};
-        SDValue retval = DAG.getMemIntrinsicNode(
-            NVPTXISD::LoadParam, dl,
-            DAG.getVTList(LoadRetVTs), LoadRetOps,
-            TheLoadType, MachinePointerInfo(), AlignI);
-        Chain = retval.getValue(1);
-        InFlag = retval.getValue(2);
-        SDValue Ret0 = retval.getValue(0);
+        SDValue VectorOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
+                               DAG.getConstant(Offsets[VecIdx], dl, MVT::i32),
+        SDValue RetVal = DAG.getMemIntrinsicNode(
+            Op, dl, DAG.getVTList(LoadVTs), VectorOps, TheLoadType,
+            MachinePointerInfo(), EltAlign);
+
+        for (unsigned j = 0; j < NumElts; ++j) {
+          SDValue Ret = RetVal.getValue(j);
           if (needTruncate)
-          Ret0 = DAG.getNode(ISD::TRUNCATE, dl, Ins[i].VT, Ret0);
-        InVals.push_back(Ret0);
+            Ret = DAG.getNode(ISD::TRUNCATE, dl, Ins[VecIdx + j].VT, Ret);
+          InVals.push_back(Ret);
+        }
+        Chain = RetVal.getValue(NumElts);
+        InFlag = RetVal.getValue(NumElts + 1);
+
+        // Cleanup
+        VecIdx = -1;
+        LoadVTs.clear();
       }
     }
   }
 
   Chain = DAG.getCALLSEQ_END(Chain,
                              DAG.getIntPtrConstant(uniqueCallSite, dl, true),
                              DAG.getIntPtrConstant(uniqueCallSite + 1, dl,
                                                    true),
if (theArgs[i]->use_empty()) {
       continue;
     }
 
     // In the following cases, assign a node order of "idx+1"
     // to newly created nodes. The SDNodes for params have to
     // appear in the same order as their order of appearance
     // in the original function. "idx+1" holds that order.
     if (!PAL.hasAttribute(i + 1, Attribute::ByVal)) {
-      if (Ty->isAggregateType()) {
-        SmallVector<EVT, 16> vtparts;
-        SmallVector<uint64_t, 16> offsets;
+      SmallVector<EVT, 16> VTs;
+      SmallVector<uint64_t, 16> Offsets;
+      SmallVector<PtxVectorInfo, 16> VectorInfo;
+
+      ComputePTXValueVTs(*this, DL, Ty, VTs, &Offsets, 0);
+      assert(VTs.size() > 0 && "empty aggregate type not expected");
 
-        // NOTE: Here, we lose the ability to issue vector loads for vectors
-        // that are a part of a struct.  This should be investigated in the
-        // future.
-        ComputePTXValueVTs(*this, DAG.getDataLayout(), Ty, vtparts, &offsets,
-                           0);
-        assert(vtparts.size() > 0 && "empty aggregate type not expected");
-        bool aggregateIsPacked = false;
-        if (StructType *STy = dyn_cast<StructType>(Ty))
-          aggregateIsPacked = STy->isPacked();
 
-        SDValue Arg = getParamSymbol(DAG, idx, PtrVT);
-        for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
-             ++parti) {
-          EVT partVT = vtparts[parti];
-          Value *srcValue = Constant::getNullValue(
-              PointerType::get(partVT.getTypeForEVT(F->getContext()),
-                               ADDRESS_SPACE_PARAM));
-          SDValue srcAddr =
-              DAG.getNode(ISD::ADD, dl, PtrVT, Arg,
-                          DAG.getConstant(offsets[parti], dl, PtrVT));
-          unsigned partAlign = aggregateIsPacked
-                                   ? 1
-                                   : DL.getABITypeAlignment(
-                                         partVT.getTypeForEVT(F->getContext()));
-          SDValue p;
-          if (Ins[InsIdx].VT.getSizeInBits() > partVT.getSizeInBits()) {
-            ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ? 
-                                     ISD::SEXTLOAD : ISD::ZEXTLOAD;
-            p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, srcAddr,
-                               MachinePointerInfo(srcValue), partVT, partAlign);
-          } else {
-            p = DAG.getLoad(partVT, dl, Root, srcAddr,
-                            MachinePointerInfo(srcValue), partAlign);
-          }
-          if (p.getNode())
-            p.getNode()->setIROrder(idx + 1);
-          InVals.push_back(p);
-          ++InsIdx;
-        }
-        if (vtparts.size() > 0)
-          --InsIdx;
-        continue;
-      }
-      if (Ty->isVectorTy()) {
-        EVT ObjectVT = getValueType(DL, Ty);
-        SDValue Arg = getParamSymbol(DAG, idx, PtrVT);
-        unsigned NumElts = ObjectVT.getVectorNumElements();
-        assert(TLI->getNumRegisters(F->getContext(), ObjectVT) == NumElts &&
-               "Vector was not scalarized");
-        EVT EltVT = ObjectVT.getVectorElementType();
-
-        // V1 load
-        // f32 = load ...
-        if (NumElts == 1) {
-          // We only have one element, so just directly load it
-          Value *SrcValue = Constant::getNullValue(PointerType::get(
-              EltVT.getTypeForEVT(F->getContext()), ADDRESS_SPACE_PARAM));
-          SDValue P = DAG.getLoad(
-              EltVT, dl, Root, Arg, MachinePointerInfo(SrcValue),
-              DL.getABITypeAlignment(EltVT.getTypeForEVT(F->getContext())),
-              MachineMemOperand::MODereferenceable |
-                  MachineMemOperand::MOInvariant);
-          if (P.getNode())
-            P.getNode()->setIROrder(idx + 1);
-
-          if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
-            P = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, P);
-          InVals.push_back(P);
-          ++InsIdx;
-        } else if (NumElts == 2) {
-          // V2 load
-          // f32,f32 = load ...
-          EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, 2);
-          Value *SrcValue = Constant::getNullValue(PointerType::get(
-              VecVT.getTypeForEVT(F->getContext()), ADDRESS_SPACE_PARAM));
-          SDValue P = DAG.getLoad(
-              VecVT, dl, Root, Arg, MachinePointerInfo(SrcValue),
-              DL.getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())),
-              MachineMemOperand::MODereferenceable |
-                  MachineMemOperand::MOInvariant);
-          if (P.getNode())
-            P.getNode()->setIROrder(idx + 1);
-
-          SDValue Elt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
-                                     DAG.getIntPtrConstant(0, dl));
-          SDValue Elt1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
-                                     DAG.getIntPtrConstant(1, dl));
-
-          if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits()) {
-            Elt0 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt0);
-            Elt1 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt1);
-          }
+      VectorizePTXValueVTs(VTs, Offsets, DL.getABITypeAlignment(Ty),
+                           VectorInfo);
+
+      int VecIdx = -1; // Index of the first element of the current vector.
+      for (unsigned parti = 0, parte = VTs.size(); parti != parte;
+        if (VectorInfo[parti] & PTX_LDST_BEGIN) {
+          assert(VecIdx == -1 && "Orphaned vector.");
+          VecIdx = parti;
 
-          InVals.push_back(Elt0);
-          InVals.push_back(Elt1);
-          InsIdx += 2;
-        } else {
-          // V4 loads
-          // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
-          // the vector will be expanded to a power of 2 elements, so we know we
-          // can always round up to the next multiple of 4 when creating the
-          // vector loads.
-          // e.g.  4 elem => 1 ld.v4
-          //       6 elem => 2 ld.v4
-          //       8 elem => 2 ld.v4
-          //      11 elem => 3 ld.v4
-          unsigned VecSize = 4;
-          if (EltVT.getSizeInBits() == 64) {
-            VecSize = 2;
-          }
-          EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
-          unsigned Ofst = 0;
-          for (unsigned i = 0; i < NumElts; i += VecSize) {
-            Value *SrcValue = Constant::getNullValue(
-                PointerType::get(VecVT.getTypeForEVT(F->getContext()),
-                                 ADDRESS_SPACE_PARAM));
-            SDValue SrcAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Arg,
-                                          DAG.getConstant(Ofst, dl, PtrVT));
-            SDValue P = DAG.getLoad(
-                VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue),
-                DL.getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())),
+        // That's the last element of this store op.
+        if (VectorInfo[parti] & PTX_LDST_END) {
+          unsigned NumElts = parti - VecIdx + 1;
+          EVT EltVT = VTs[parti];
+          // i1 is loaded/stored as i8.
+          EVT LoadVT = EltVT == MVT::i1 ? MVT::i8 : EltVT;
+          EVT VecVT = EVT::getVectorVT(F->getContext(), LoadVT, NumElts);
+          SDValue VecAddr =
+              DAG.getNode(ISD::ADD, dl, PtrVT, Arg,
+                          DAG.getConstant(Offsets[VecIdx], dl, PtrVT));
+          Value *srcValue = Constant::getNullValue(PointerType::get(
+              EltVT.getTypeForEVT(F->getContext()), ADDRESS_SPACE_PARAM));
+          SDValue P =
+              DAG.getLoad(VecVT, dl, Root, VecAddr,
+                          MachinePointerInfo(srcValue), aggregateIsPacked,
                           MachineMemOperand::MODereferenceable |
                               MachineMemOperand::MOInvariant);
-            if (P.getNode())
-              P.getNode()->setIROrder(idx + 1);
-
-            for (unsigned j = 0; j < VecSize; ++j) {
-              if (i + j >= NumElts)
-                break;
-              SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
-                                        DAG.getIntPtrConstant(j, dl));
-              if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
-                Elt = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt);
-              InVals.push_back(Elt);
-            }
-            Ofst += DL.getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
-          }
-          InsIdx += NumElts;
-        }
-
-        if (NumElts > 0)
-          --InsIdx;
-        continue;
-      }
-      // A plain scalar.
-      EVT ObjectVT = getValueType(DL, Ty);
-      // If ABI, load from the param symbol
-      SDValue Arg = getParamSymbol(DAG, idx, PtrVT);
-      Value *srcValue = Constant::getNullValue(PointerType::get(
-          ObjectVT.getTypeForEVT(F->getContext()), ADDRESS_SPACE_PARAM));
-      SDValue p;
-       if (ObjectVT.getSizeInBits() < Ins[InsIdx].VT.getSizeInBits()) {
-        ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ? 
-                                       ISD::SEXTLOAD : ISD::ZEXTLOAD;
-        p = DAG.getExtLoad(
-            ExtOp, dl, Ins[InsIdx].VT, Root, Arg, MachinePointerInfo(srcValue),
-            ObjectVT,
-            DL.getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
-      } else {
-        p = DAG.getLoad(
-            Ins[InsIdx].VT, dl, Root, Arg, MachinePointerInfo(srcValue),
-            DL.getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
-      }
-      if (p.getNode())
-        p.getNode()->setIROrder(idx + 1);
-      InVals.push_back(p);
-      continue;
-    }
+          for (unsigned j = 0; j < NumElts; ++j) {
+            SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, LoadVT, P,
+            // We've loaded i1 as an i8 and now must truncate it back to i1
+            if (EltVT == MVT::i1)
+              Elt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Elt);
+            // Extend the element if necesary (e.g an i8 is loaded
+            // into an i16 register)
+            if (Ins[InsIdx].VT.getSizeInBits() > LoadVT.getSizeInBits()) {
+              unsigned Extend = Ins[InsIdx].Flags.isSExt() ? ISD::SIGN_EXTEND
+                                                           : ISD::ZERO_EXTEND;
+              Elt = DAG.getNode(Extend, dl, Ins[InsIdx].VT, Elt);
+            }
+          // Reset vector tracking state.
+          VecIdx = -1;
+        ++InsIdx;
+      if (VTs.size() > 0)
 
     // Param has ByVal attribute
     // Return MoveParam(param symbol).
     // Ideally, the param symbol can be returned directly,
     // but when SDNode builder decides to use it in a CopyToReg(),
     // machine instruction fails because TargetExternalSymbol
     // (not lowered) is target dependent, and CopyToReg assumes
     // the source is lowered.

 
 SDValue
 NVPTXTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
                                  bool isVarArg,
                                  const SmallVectorImpl<ISD::OutputArg> &Outs,
                                  const SmallVectorImpl<SDValue> &OutVals,
                                  const SDLoc &dl, SelectionDAG &DAG) const {
   MachineFunction &MF = DAG.getMachineFunction();
-  const Function *F = MF.getFunction();
-  Type *RetTy = F->getReturnType();
-  const DataLayout &TD = DAG.getDataLayout();
+  Type *RetTy = MF.getFunction()->getReturnType();
 
   bool isABI = (STI.getSmVersion() >= 20);
   assert(isABI && "Non-ABI compilation is not supported");
   if (!isABI)
     return Chain;
 
-  if (VectorType *VTy = dyn_cast<VectorType>(RetTy)) {
-    // If we have a vector type, the OutVals array will be the scalarized
-    // components and we have combine them into 1 or more vector stores.
-    unsigned NumElts = VTy->getNumElements();
-    assert(NumElts == Outs.size() && "Bad scalarization of return value");
-
-    // const_cast can be removed in later LLVM versions
-    EVT EltVT = getValueType(TD, RetTy).getVectorElementType();
-    bool NeedExtend = false;
-    if (EltVT.getSizeInBits() < 16)
-      NeedExtend = true;
-
-    // V1 store
-    if (NumElts == 1) {
-      SDValue StoreVal = OutVals[0];
-      // We only have one element, so just directly store it
-      if (NeedExtend)
-        StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
-      SDValue Ops[] = { Chain, DAG.getConstant(0, dl, MVT::i32), StoreVal };
-      Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
-                                      DAG.getVTList(MVT::Other), Ops,
-                                      EltVT, MachinePointerInfo());
-    } else if (NumElts == 2) {
-      // V2 store
-      SDValue StoreVal0 = OutVals[0];
-      SDValue StoreVal1 = OutVals[1];
-
-      if (NeedExtend) {
-        StoreVal0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal0);
-        StoreVal1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal1);
-      }
-
-      SDValue Ops[] = { Chain, DAG.getConstant(0, dl, MVT::i32), StoreVal0,
-                        StoreVal1 };
-      Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetvalV2, dl,
-                                      DAG.getVTList(MVT::Other), Ops,
-                                      EltVT, MachinePointerInfo());
-    } else {
-      // V4 stores
-      // We have at least 4 elements (<3 x Ty> expands to 4 elements) and the
-      // vector will be expanded to a power of 2 elements, so we know we can
-      // always round up to the next multiple of 4 when creating the vector
-      // stores.
-      // e.g.  4 elem => 1 st.v4
-      //       6 elem => 2 st.v4
-      //       8 elem => 2 st.v4
-      //      11 elem => 3 st.v4
-
-      unsigned VecSize = 4;
-      if (OutVals[0].getValueSizeInBits() == 64)
-        VecSize = 2;
-
-      unsigned Offset = 0;
-
-      EVT VecVT =
-          EVT::getVectorVT(F->getContext(), EltVT, VecSize);
-      unsigned PerStoreOffset =
-          TD.getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
-
-      for (unsigned i = 0; i < NumElts; i += VecSize) {
-        // Get values
-        SDValue StoreVal;
-        SmallVector<SDValue, 8> Ops;
-        Ops.push_back(Chain);
-        Ops.push_back(DAG.getConstant(Offset, dl, MVT::i32));
-        unsigned Opc = NVPTXISD::StoreRetvalV2;
-        EVT ExtendedVT = (NeedExtend) ? MVT::i16 : OutVals[0].getValueType();
-
-        StoreVal = OutVals[i];
-        if (NeedExtend)
-          StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
-        Ops.push_back(StoreVal);
-
-        if (i + 1 < NumElts) {
-          StoreVal = OutVals[i + 1];
-          if (NeedExtend)
-            StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
-        } else {
-          StoreVal = DAG.getUNDEF(ExtendedVT);
-        }
-        Ops.push_back(StoreVal);
-
-        if (VecSize == 4) {
-          Opc = NVPTXISD::StoreRetvalV4;
-          if (i + 2 < NumElts) {
-            StoreVal = OutVals[i + 2];
-            if (NeedExtend)
-              StoreVal =
-                  DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
-          } else {
-            StoreVal = DAG.getUNDEF(ExtendedVT);
-          }
-          Ops.push_back(StoreVal);
+  const DataLayout DL = DAG.getDataLayout();
+  SmallVector<EVT, 16> VTs;
+  SmallVector<uint64_t, 16> Offsets;
+  ComputePTXValueVTs(*this, DL, RetTy, VTs, &Offsets);
+  assert(VTs.size() == OutVals.size() && "Bad return value decomposition");
 
-          if (i + 3 < NumElts) {
-            StoreVal = OutVals[i + 3];
-            if (NeedExtend)
-              StoreVal =
-                  DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
-          } else {
-            StoreVal = DAG.getUNDEF(ExtendedVT);
-          }
-          Ops.push_back(StoreVal);
-        }
+  SmallVector<PtxVectorInfo, 16> VectorInfo;
+  VectorizePTXValueVTs(VTs, Offsets,
+                       RetTy->isSized() ? DL.getABITypeAlignment(RetTy) : 1,
+                       VectorInfo);
+
+  // PTX Interoperability Guide 3.3(A): [Integer] Values shorter than
+  // 32-bits are sign extended or zero extended, depending on whether
+  // they are signed or unsigned types.
+  bool ExtendIntegerRetVal =
+      RetTy->isIntegerTy() && DL.getTypeAllocSizeInBits(RetTy) < 32;
+  bool aggregateIsPacked = false;
 
-        // Chain = DAG.getNode(Opc, dl, MVT::Other, &Ops[0], Ops.size());
-        Chain =
-            DAG.getMemIntrinsicNode(Opc, dl, DAG.getVTList(MVT::Other), Ops,
-                                    EltVT, MachinePointerInfo());
-        Offset += PerStoreOffset;
-      }
-    }
-  } else {
-    SmallVector<EVT, 16> ValVTs;
-    SmallVector<uint64_t, 16> Offsets;
-    ComputePTXValueVTs(*this, DAG.getDataLayout(), RetTy, ValVTs, &Offsets, 0);
-    assert(ValVTs.size() == OutVals.size() && "Bad return value decomposition");
+  if (StructType *STy = dyn_cast<StructType>(RetTy))
+    aggregateIsPacked = STy->isPacked();
 
-    for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
-      SDValue theVal = OutVals[i];
-      EVT TheValType = theVal.getValueType();
-      unsigned numElems = 1;
-      if (TheValType.isVector())
-        numElems = TheValType.getVectorNumElements();
-      for (unsigned j = 0, je = numElems; j != je; ++j) {
-        SDValue TmpVal = theVal;
-        if (TheValType.isVector())
-          TmpVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
-                               TheValType.getVectorElementType(), TmpVal,
-                               DAG.getIntPtrConstant(j, dl));
-        EVT TheStoreType = ValVTs[i];
-        if (RetTy->isIntegerTy() && TD.getTypeAllocSizeInBits(RetTy) < 32) {
-          // The following zero-extension is for integer types only, and
-          // specifically not for aggregates.
-          TmpVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, TmpVal);
-          TheStoreType = MVT::i32;
-        } else if (RetTy->isHalfTy()) {
-          TheStoreType = MVT::f16;
-        } else if (TmpVal.getValueSizeInBits() < 16)
-          TmpVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, TmpVal);
-
-        SDValue Ops[] = {
-          Chain,
-          DAG.getConstant(Offsets[i], dl, MVT::i32),
-          TmpVal };
-        Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
-                                        DAG.getVTList(MVT::Other), Ops,
-                                        TheStoreType,
-                                        MachinePointerInfo());
+  SmallVector<SDValue, 6> LdStOps;
+    // New load/store. Record chain and offset operands.
+    if (VectorInfo[i] & PTX_LDST_BEGIN) {
+      assert(LdStOps.empty() && "Orphaned operand list.");
+      LdStOps.push_back(Chain);
+      LdStOps.push_back(DAG.getConstant(Offsets[i], dl, MVT::i32));
+    }
+
+    SDValue RetVal = OutVals[i];
+    if (ExtendIntegerRetVal) {
+      RetVal = DAG.getNode(Outs[i].Flags.isSExt() ? ISD::SIGN_EXTEND
+                                                  : ISD::ZERO_EXTEND,
+                           dl, MVT::i32, RetVal);
+    } else if (RetVal.getValueSizeInBits() < 16) {
+      // Use 16-bit registers for small load-stores as it's the
+      // smallest general purpose register size supported by NVPTX.
+      RetVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, RetVal);
+    }
+
+    // Record the value to return.
+    LdStOps.push_back(RetVal);
+
+    // That's the last element of this store op.
+    if (VectorInfo[i] & PTX_LDST_END) {
+      NVPTXISD::NodeType Op;
+      unsigned NumElts = LdStOps.size() - 2;
+      switch (NumElts) {
+      case 1:
+        Op = NVPTXISD::StoreRetval;
+        break;
+      case 2:
+        Op = NVPTXISD::StoreRetvalV2;
+        break;
+      case 4:
+        Op = NVPTXISD::StoreRetvalV4;
+        break;
+      default:
+        llvm_unreachable("Invalid vector info.");
       }
+
+      // Adjust type of load/store op if we've extended the scalar
+      // return value.
+      EVT TheStoreType = ExtendIntegerRetVal ? MVT::i32 : VTs[i];
+      Chain = DAG.getMemIntrinsicNode(Op, dl, DAG.getVTList(MVT::Other),
+                                      LdStOps, TheStoreType,
+                                      MachinePointerInfo(), 1);
+      // Cleanup vector state.
+      LdStOps.clear();
     }
   }
 
   return DAG.getNode(NVPTXISD::RET_FLAG, dl, MVT::Other, Chain);
 }
 
 void NVPTXTargetLowering::LowerAsmOperandForConstraint(
     SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,

test/CodeGen/NVPTX/aggregate-return.ll

 ; RUN: llc < %s -march=nvptx64 -mcpu=sm_35 | FileCheck %s
 
 declare <2 x float> @barv(<2 x float> %input)
+declare <3 x float> @barv3(<3 x float> %input)
 declare [2 x float] @bara([2 x float] %input)
 declare {float, float} @bars({float, float} %input)
 
-define void @foov(<2 x float> %input, <2 x float>* %output) {
-; CHECK-LABEL: @foov
+define void @test_v2f32(<2 x float> %input, <2 x float>* %output) {
+; CHECK-LABEL: @test_v2f32
   %call = tail call <2 x float> @barv(<2 x float> %input)
 ; CHECK: .param .align 8 .b8 retval0[8];
-; CHECK: ld.param.v2.f32 {[[ELEMV1:%f[0-9]+]], [[ELEMV2:%f[0-9]+]]}, [retval0+0];
+; CHECK: ld.param.v2.f32 {[[E0:%f[0-9]+]], [[E1:%f[0-9]+]]}, [retval0+0];
   store <2 x float> %call, <2 x float>* %output, align 8
-; CHECK: st.v2.f32 [{{%rd[0-9]+}}], {[[ELEMV1]], [[ELEMV2]]}
+; CHECK: st.v2.f32 [{{%rd[0-9]+}}], {[[E0]], [[E1]]}
   ret void
 }
 
-define void @fooa([2 x float] %input, [2 x float]* %output) {
-; CHECK-LABEL: @fooa
+define void @test_v3f32(<3 x float> %input, <3 x float>* %output) {
+; CHECK-LABEL: @test_v3f32
+;
+  %call = tail call <3 x float> @barv3(<3 x float> %input)
+; CHECK: .param .align 16 .b8 retval0[16];
+; CHECK-DAG: ld.param.v2.f32 {[[E0:%f[0-9]+]], [[E1:%f[0-9]+]]}, [retval0+0];
+; CHECK-DAG: ld.param.f32 [[E2:%f[0-9]+]], [retval0+8];
+; Make sure we don't load more values than than we need to.
+; CHECK-NOT: ld.param.f32 [[E3:%f[0-9]+]], [retval0+12];
+  store <3 x float> %call, <3 x float>* %output, align 8
+; CHECK-DAG: st.f32 [{{%rd[0-9]}}+8],
+; -- This is suboptimal. We should do st.v2.f32 instead
+;    of combining 2xf32 info i64.
+; CHECK-DAG: st.u64 [{{%rd[0-9]}}],
+; CHECK: ret;
+  ret void
+}
+
+define void @test_a2f32([2 x float] %input, [2 x float]* %output) {
+; CHECK-LABEL: @test_a2f32
   %call = tail call [2 x float] @bara([2 x float] %input)
 ; CHECK: .param .align 4 .b8 retval0[8];
 ; CHECK-DAG: ld.param.f32 [[ELEMA1:%f[0-9]+]], [retval0+0];
 ; CHECK-DAG: ld.param.f32 [[ELEMA2:%f[0-9]+]], [retval0+4];
   store [2 x float] %call, [2 x float]* %output, align 4
 ; CHECK: }
 ; CHECK-DAG: st.f32 [{{%rd[0-9]+}}], [[ELEMA1]]
 ; CHECK-DAG: st.f32 [{{%rd[0-9]+}}+4], [[ELEMA2]]
   ret void
 ; CHECK: ret
 }
 
-define void @foos({float, float} %input, {float, float}* %output) {
-; CHECK-LABEL: @foos
+define void @test_s2f32({float, float} %input, {float, float}* %output) {
+; CHECK-LABEL: @test_s2f32
   %call = tail call {float, float} @bars({float, float} %input)
 ; CHECK: .param .align 4 .b8 retval0[8];
 ; CHECK-DAG: ld.param.f32 [[ELEMS1:%f[0-9]+]], [retval0+0];
 ; CHECK-DAG: ld.param.f32 [[ELEMS2:%f[0-9]+]], [retval0+4];
   store {float, float} %call, {float, float}* %output, align 4
 ; CHECK: }
 ; CHECK-DAG: st.f32 [{{%rd[0-9]+}}], [[ELEMS1]]
 ; CHECK-DAG: st.f32 [{{%rd[0-9]+}}+4], [[ELEMS2]]
   ret void
 ; CHECK: ret
 }

test/CodeGen/NVPTX/f16-instructions.ll

 define half @test_tailcall_flipped(half %a, half %b) #0 {
   %r = tail call half @test_callee(half %b, half %a)
   ret half %r
 }
 
 ; CHECK-LABEL: test_select(
 ; CHECK-DAG:  ld.param.b16    [[A:%h[0-9]+]], [test_select_param_0];
 ; CHECK-DAG:  ld.param.b16    [[B:%h[0-9]+]], [test_select_param_1];
-; CHECK:      setp.eq.b16     [[PRED:%p[0-9]+]], %rs{{.*}}, 1;
+; CHECK-DAG:  setp.eq.b16     [[PRED:%p[0-9]+]], %rs{{.*}}, 1;
 ; CHECK-NEXT: selp.b16        [[R:%h[0-9]+]], [[A]], [[B]], [[PRED]];
 ; CHECK-NEXT: st.param.b16    [func_retval0+0], [[R]];
 ; CHECK-NEXT: ret;
 define half @test_select(half %a, half %b, i1 zeroext %c) #0 {
   %r = select i1 %c, half %a, half %b
   ret half %r
 }
 

test/CodeGen/NVPTX/ldparam-v4.ll

 ; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
 
 declare <4 x float> @bar()
 
+; CHECK-LABEL: .func foo(
 define void @foo(<4 x float>* %ptr) {
-; CHECK: ld.param.v4.f32
+; CHECK:     ld.param.u32 %[[PTR:r[0-9]+]], [foo_param_0];
+; CHECK:     ld.param.v4.f32 {[[E0:%f[0-9]+]], [[E1:%f[0-9]+]], [[E2:%f[0-9]+]], [[E3:%f[0-9]+]]}, [retval0+0];
+; CHECK:     st.v4.f32    [%[[PTR]]], {[[E0]], [[E1]], [[E2]], [[E3]]}
   %val = tail call <4 x float> @bar()
   store <4 x float> %val, <4 x float>* %ptr
   ret void
 }

test/CodeGen/NVPTX/lower-aggr-copies.ll

-; RUN: llc < %s -march=nvptx64 -mcpu=sm_35 | FileCheck %s --check-prefix PTX
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_35 -O0 | FileCheck %s --check-prefix PTX
 ; RUN: opt < %s -S -nvptx-lower-aggr-copies | FileCheck %s --check-prefix IR
 
 ; Verify that the NVPTXLowerAggrCopies pass works as expected - calls to
 ; llvm.mem* intrinsics get lowered to loops.
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "nvptx64-unknown-unknown"
 

 ; IR-NEXT:    [[VAL:%[0-9]+]] = load i8, i8* [[LOADPTR]]
 ; IR-NEXT:    [[STOREPTR:%[0-9]+]] = getelementptr inbounds i8, i8* %dst, i64
 ; IR-NEXT:    store i8 [[VAL]], i8* [[STOREPTR]]
 
 ; PTX-LABEL:  .visible .func (.param .b64 func_retval0) memcpy_caller
 ; PTX:        LBB[[LABEL:[_0-9]+]]:
 ; PTX:        ld.u8 %rs[[REG:[0-9]+]]
 ; PTX:        st.u8 [%rd{{[0-9]+}}], %rs[[REG]]
-; PTX:        add.s64 %rd[[COUNTER:[0-9]+]], %rd[[COUNTER]], 1
-; PTX-NEXT:   setp.lt.u64 %p[[PRED:[0-9]+]], %rd[[COUNTER]], %rd
-; PTX-NEXT:   @%p[[PRED]] bra LBB[[LABEL]]
+; PTX:        add.s64 %rd[[COUNTER:[0-9]+]], %rd{{[0-9]+}}, 1
+; PTX:        setp.lt.u64 %p[[PRED:[0-9]+]], %rd[[COUNTER]], %rd
+; PTX:        @%p[[PRED]] bra LBB[[LABEL]]
 }
 
 define i8* @memcpy_volatile_caller(i8* %dst, i8* %src, i64 %n) #0 {
 entry:
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %dst, i8* %src, i64 %n, i32 1, i1 true)
   ret i8* %dst
 
 ; IR-LABEL:   @memcpy_volatile_caller
 ; IR:         load volatile
 ; IR:         store volatile
 
 ; PTX-LABEL:  .visible .func (.param .b64 func_retval0) memcpy_volatile_caller
 ; PTX:        LBB[[LABEL:[_0-9]+]]:
 ; PTX:        ld.volatile.u8 %rs[[REG:[0-9]+]]
 ; PTX:        st.volatile.u8 [%rd{{[0-9]+}}], %rs[[REG]]
-; PTX:        add.s64 %rd[[COUNTER:[0-9]+]], %rd[[COUNTER]], 1
-; PTX-NEXT:   setp.lt.u64 %p[[PRED:[0-9]+]], %rd[[COUNTER]], %rd
-; PTX-NEXT:   @%p[[PRED]] bra LBB[[LABEL]]
+; PTX:        add.s64 %rd[[COUNTER:[0-9]+]], %rd{{[0-9]+}}, 1
+; PTX:        setp.lt.u64 %p[[PRED:[0-9]+]], %rd[[COUNTER]], %rd
+; PTX:        @%p[[PRED]] bra LBB[[LABEL]]
 }
 
 define i8* @memcpy_casting_caller(i32* %dst, i32* %src, i64 %n) #0 {
 entry:
   %0 = bitcast i32* %dst to i8*
   %1 = bitcast i32* %src to i8*
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %0, i8* %1, i64 %n, i32 1, i1 false)
   ret i8* %0

 
 ; IR-LABEL:   @memset_caller
 ; IR:         [[VAL:%[0-9]+]] = trunc i32 %c to i8
 ; IR:         loadstoreloop:
 ; IR:         [[STOREPTR:%[0-9]+]] = getelementptr inbounds i8, i8* %dst, i64
 ; IR-NEXT:    store i8 [[VAL]], i8* [[STOREPTR]]
 
 ; PTX-LABEL:  .visible .func (.param .b64 func_retval0) memset_caller(
-; PTX:        ld.param.u8 %rs[[REG:[0-9]+]]
+; PTX:        ld.param.u32 %r[[C:[0-9]+]]
+; PTX:        cvt.u16.u32  %rs[[REG:[0-9]+]], %r[[C]];
 ; PTX:        LBB[[LABEL:[_0-9]+]]:
 ; PTX:        st.u8 [%rd{{[0-9]+}}], %rs[[REG]]
-; PTX:        add.s64 %rd[[COUNTER:[0-9]+]], %rd[[COUNTER]], 1
-; PTX-NEXT:   setp.lt.u64 %p[[PRED:[0-9]+]], %rd[[COUNTER]], %rd
-; PTX-NEXT:   @%p[[PRED]] bra LBB[[LABEL]]
+; PTX:        add.s64 %rd[[COUNTER:[0-9]+]], %rd{{[0-9]+}}, 1
+; PTX:        setp.lt.u64 %p[[PRED:[0-9]+]], %rd[[COUNTER]], %rd
+; PTX:        @%p[[PRED]] bra LBB[[LABEL]]
 }
 
 define i8* @volatile_memset_caller(i8* %dst, i32 %c, i64 %n) #0 {
 entry:
   %0 = trunc i32 %c to i8
   tail call void @llvm.memset.p0i8.i64(i8* %dst, i8 %0, i64 %n, i32 1, i1 true)
   ret i8* %dst
 

 
 ; PTX-LABEL:  .visible .func (.param .b64 func_retval0) memmove_caller(
 ; PTX:        ld.param.u64 %rd[[N:[0-9]+]]
 ; PTX:        setp.eq.s64 %p[[NEQ0:[0-9]+]], %rd[[N]], 0
 ; PTX:        setp.ge.u64 %p[[SRC_GT_THAN_DST:[0-9]+]], %rd{{[0-9]+}}, %rd{{[0-9]+}}
 ; PTX-NEXT:   @%p[[SRC_GT_THAN_DST]] bra LBB[[FORWARD_BB:[0-9_]+]]
 ; -- this is the backwards copying BB
 ; PTX:        @%p[[NEQ0]] bra LBB[[EXIT:[0-9_]+]]
-; PTX:        add.s64 %rd[[N]], %rd[[N]], -1
+; PTX:        add.s64 %rd{{[0-9]}}, %rd{{[0-9]}}, -1
 ; PTX:        ld.u8 %rs[[ELEMENT:[0-9]+]]
 ; PTX:        st.u8 [%rd{{[0-9]+}}], %rs[[ELEMENT]]
 ; -- this is the forwards copying BB
 ; PTX:        LBB[[FORWARD_BB]]:
 ; PTX:        @%p[[NEQ0]] bra LBB[[EXIT]]
 ; PTX:        ld.u8 %rs[[ELEMENT2:[0-9]+]]
 ; PTX:        st.u8 [%rd{{[0-9]+}}], %rs[[ELEMENT2]]
-; PTX:        add.s64 %rd[[INDEX:[0-9]+]], %rd[[INDEX]], 1
+; PTX:        add.s64 %rd{{[0-9]+}}, %rd{{[0-9]+}}, 1
 ; -- exit block
 ; PTX:        LBB[[EXIT]]:
 ; PTX-NEXT:   st.param.b64 [func_retval0
 ; PTX-NEXT:   ret
 }

test/CodeGen/NVPTX/param-load-store.ll

+; Verifies correctness of load/store of parameters and return values.
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_35 -O0 | FileCheck %s
+
+%s_i1 = type { i1 }
+%s_i8 = type { i8 }
+%s_i16 = type { i16 }
+%s_half = type { half }
+%s_i32 = type { i32 }
+%s_float = type { float }
+%s_i64 = type { i64 }
+%s_f64 = type { double }
+
+; More complicated types. i64 is used to increase natural alignment
+; requirement for the type.
+%s_i32x4 = type { i32, i32, i32, i32, i64}
+%s_i32f32 = type { i32, float, i32, float, i64}
+%s_i8i32x4 = type { i32, i32, i8, i32, i32, i64}
+%s_i8i32x4p = type <{ i32, i32, i8, i32, i32, i64}>
+%s_crossfield = type { i32, [2 x i32], <4 x i32>, [3 x {i32, i32, i32}]}
+; All scalar parameters must be at least 32 bits in size.
+; i1 is loaded/stored as i8.
+
+; CHECK: .func  (.param .b32 func_retval0)
+; CHECK-LABEL: test_i1(
+; CHECK-NEXT: .param .b32 test_i1_param_0
+; CHECK:      ld.param.u8 [[A8:%r[0-9]+]], [test_i1_param_0];
+; CHECK:      and.b32 [[A:%r[0-9]+]], [[A8]], 1;
+; CHECK:      .param .b32 param0;
+; CHECK:      st.param.b32    [param0+0], [[A]]
+; CHECK:      .param .b32 retval0;
+; CHECK:      call.uni
+; CHECK-NEXT: test_i1,
+; CHECK:      ld.param.b32    [[R8:%r[0-9]+]], [retval0+0];
+; CHECK:      and.b32         [[R:%r[0-9]+]], [[R8]], 1;
+; CHECK:      st.param.b32    [func_retval0+0], [[R]];
+; CHECK:      ret;
+define i1 @test_i1(i1 %a) {
+  %r = tail call i1 @test_i1(i1 %a);
+  ret i1 %r;
+}
+
+; Signed i1 is a somewhat special case. We only care about one bit and
+; then us neg.s32 to convert it to 32-bit -1 if it's set.
+; CHECK: .func  (.param .b32 func_retval0)
+; CHECK-LABEL: test_i1s(
+; CHECK-NEXT: .param .b32 test_i1s_param_0
+; CHECK:      ld.param.u8 [[A8:%rs[0-9]+]], [test_i1s_param_0];
+; CHECK:      cvt.u32.u16     [[A32:%r[0-9]+]], [[A8]];
+; CHECK:      and.b32         [[A1:%r[0-9]+]], [[A32]], 1;
+; CHECK:      neg.s32         [[A:%r[0-9]+]], [[A1]];
+; CHECK:      .param .b32 param0;
+; CHECK:      st.param.b32    [param0+0], [[A]];
+; CHECK:      .param .b32 retval0;
+; CHECK:      call.uni
+; CHECK:      ld.param.b32    [[R8:%r[0-9]+]], [retval0+0];
+; CHECK:      and.b32         [[R1:%r[0-9]+]], [[R8]], 1;
+; CHECK:      neg.s32         [[R:%r[0-9]+]], [[R1]];
+; CHECK:      st.param.b32    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define signext i1 @test_i1s(i1 signext %a) {
+       %r = tail call signext i1 @test_i1s(i1 signext %a);
+       ret i1 %r;
+}
+
+; Make sure that i1 loads are vectorized as i8 loads, respecting each element alignment.
+; CHECK: .func  (.param .align 4 .b8 func_retval0[4])
+; CHECK-LABEL: test_v3i1(
+; CHECK-NEXT: .param .align 4 .b8 test_v3i1_param_0[4]
+; CHECK-DAG:  ld.param.u8     [[E2:%rs[0-9]+]], [test_v3i1_param_0+2];
+; CHECK-DAG:  ld.param.v2.u8  {[[E0:%rs[0-9]+]], [[E1:%rs[0-9]+]]}, [test_v3i1_param_0]
+; CHECK:      .param .align 4 .b8 param0[4];
+; CHECK-DAG:  st.param.v2.b8  [param0+0], {[[E0]], [[E1]]};
+; CHECK-DAG:  st.param.b8     [param0+2], [[E2]];
+; CHECK:      .param .align 4 .b8 retval0[4];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v3i1,
+; CHECK-DAG:  ld.param.v2.b8  {[[RE0:%rs[0-9]+]], [[RE1:%rs[0-9]+]]}, [retval0+0];
+; CHECK-DAG:  ld.param.b8     [[RE2:%rs[0-9]+]], [retval0+2];
+; CHECK-DAG:  st.param.v2.b8  [func_retval0+0], {[[RE0]], [[RE1]]}
+; CHECK-DAG:  st.param.b8     [func_retval0+2], [[RE2]];
+; CHECK-NEXT: ret;
+define <3 x i1> @test_v3i1(<3 x i1> %a) {
+       %r = tail call <3 x i1> @test_v3i1(<3 x i1> %a);
+       ret <3 x i1> %r;
+}
+
+; CHECK: .func  (.param .align 4 .b8 func_retval0[4])
+; CHECK-LABEL: test_v4i1(
+; CHECK-NEXT: .param .align 4 .b8 test_v4i1_param_0[4]
+; CHECK:      ld.param.v4.u8 {[[E0:%rs[0-9]+]], [[E1:%rs[0-9]+]], [[E2:%rs[0-9]+]], [[E3:%rs[0-9]+]]}, [test_v4i1_param_0]
+; CHECK:      .param .align 4 .b8 param0[4];
+; CHECK:      st.param.v4.b8  [param0+0], {[[E0]], [[E1]], [[E2]], [[E3]]};
+; CHECK:      .param .align 4 .b8 retval0[4];
+; CHECK:      call.uni (retval0),
+; CHECK:      test_v4i1,
+; CHECK:      ld.param.v4.b8  {[[RE0:%rs[0-9]+]], [[RE1:%rs[0-9]+]], [[RE2:%rs[0-9]+]], [[RE3:%rs[0-9]+]]}, [retval0+0];
+; CHECK:      st.param.v4.b8 [func_retval0+0], {[[RE0]], [[RE1]], [[RE2]], [[RE3]]};
+; CHECK-NEXT: ret;
+define <4 x i1> @test_v4i1(<4 x i1> %a) {
+       %r = tail call <4 x i1> @test_v4i1(<4 x i1> %a);
+       ret <4 x i1> %r;
+}
+
+; CHECK: .func  (.param .align 8 .b8 func_retval0[8])
+; CHECK-LABEL: test_v5i1(
+; CHECK-NEXT: .param .align 8 .b8 test_v5i1_param_0[8]
+; CHECK-DAG:  ld.param.u8     [[E4:%rs[0-9]+]], [test_v5i1_param_0+4];
+; CHECK-DAG:  ld.param.v4.u8  {[[E0:%rs[0-9]+]], [[E1:%rs[0-9]+]], [[E2:%rs[0-9]+]], [[E3:%rs[0-9]+]]}, [test_v5i1_param_0]
+; CHECK:      .param .align 8 .b8 param0[8];
+; CHECK-DAG:  st.param.v4.b8  [param0+0], {[[E0]], [[E1]], [[E2]], [[E3]]};
+; CHECK-DAG:  st.param.b8     [param0+4], [[E4]];
+; CHECK:      .param .align 8 .b8 retval0[8];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v5i1,
+; CHECK-DAG:  ld.param.v4.b8  {[[RE0:%rs[0-9]+]], [[RE1:%rs[0-9]+]], [[RE2:%rs[0-9]+]], [[RE3:%rs[0-9]+]]}, [retval0+0];
+; CHECK-DAG:  ld.param.b8     [[RE4:%rs[0-9]+]], [retval0+4];
+; CHECK-DAG:  st.param.v4.b8  [func_retval0+0], {[[RE0]], [[RE1]], [[RE2]], [[RE3]]}
+; CHECK-DAG:  st.param.b8     [func_retval0+4], [[RE4]];
+; CHECK-NEXT: ret;
+define <5 x i1> @test_v5i1(<5 x i1> %a) {
+       %r = tail call <5 x i1> @test_v5i1(<5 x i1> %a);
+       ret <5 x i1> %r;
+}
+
+; Unsigned i8 is loaded directly into 32-bit register.
+; CHECK: .func  (.param .b32 func_retval0)
+; CHECK-LABEL: test_i8(
+; CHECK-NEXT: .param .b32 test_i8_param_0
+; CHECK:      ld.param.u8 [[A8:%rs[0-9]+]], [test_i8_param_0];
+; CHECK:      cvt.u32.u16     [[A32:%r[0-9]+]], [[A8]];
+; CHECK:      and.b32         [[A:%r[0-9]+]], [[A32]], 255;
+; CHECK:      .param .b32 param0;
+; CHECK:      st.param.b32    [param0+0], [[A]];
+; CHECK:      .param .b32 retval0;
+; CHECK:      call.uni (retval0),
+; CHECK:      test_i8,
+; CHECK:      ld.param.b32    [[R32:%r[0-9]+]], [retval0+0];
+; CHECK:      and.b32         [[R:%r[0-9]+]], [[R32]], 255;
+; CHECK:      st.param.b32    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define i8 @test_i8(i8 %a) {
+       %r = tail call i8 @test_i8(i8 %a);
+       ret i8 %r;
+}
+
+; signed i8 is loaded into 16-bit register which is then sign-extended to i32.
+; CHECK: .func  (.param .b32 func_retval0)
+; CHECK-LABEL: test_i8s(
+; CHECK-NEXT: .param .b32 test_i8s_param_0
+; CHECK:      ld.param.s8 [[A8:%rs[0-9]+]], [test_i8s_param_0];
+; CHECK:      cvt.s32.s16     [[A:%r[0-9]+]], [[A8]];
+; CHECK:      .param .b32 param0;
+; CHECK:      st.param.b32    [param0+0], [[A]];
+; CHECK:      .param .b32 retval0;
+; CHECK:      call.uni (retval0),
+; CHECK:      test_i8s,
+; CHECK:      ld.param.b32    [[R32:%r[0-9]+]], [retval0+0];
+; -- This is suspicious (though correct) -- why not cvt.u8.u32, cvt.s8.s32 ?
+; CHECK:      cvt.u16.u32     [[R16:%rs[0-9]+]], [[R32]];
+; CHECK:      cvt.s32.s16     [[R:%r[0-9]+]], [[R16]];
+; CHECK:      st.param.b32    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define signext i8 @test_i8s(i8 signext %a) {
+       %r = tail call signext i8 @test_i8s(i8 signext %a);
+       ret i8 %r;
+}
+
+; CHECK: .func  (.param .align 4 .b8 func_retval0[4])
+; CHECK-LABEL: test_v3i8(
+; CHECK-NEXT: .param .align 4 .b8 test_v3i8_param_0[4]
+; CHECK-DAG:  ld.param.u8     [[E2:%rs[0-9]+]], [test_v3i8_param_0+2];
+; CHECK-DAG:  ld.param.v2.u8  {[[E0:%rs[0-9]+]], [[E1:%rs[0-9]+]]}, [test_v3i8_param_0];
+; CHECK:      .param .align 4 .b8 param0[4];
+; CHECK:      st.param.v2.b8  [param0+0], {[[E0]], [[E1]]};
+; CHECK:      st.param.b8     [param0+2], [[E2]];
+; CHECK:      .param .align 4 .b8 retval0[4];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v3i8,
+; CHECK-DAG:  ld.param.v2.b8  {[[RE0:%rs[0-9]+]], [[RE1:%rs[0-9]+]]}, [retval0+0];
+; CHECK-DAG:  ld.param.b8     [[RE2:%rs[0-9]+]], [retval0+2];
+; CHECK-DAG:  st.param.v2.b8  [func_retval0+0], {[[RE0]], [[RE1]]};
+; CHECK-DAG:  st.param.b8     [func_retval0+2], [[RE2]];
+; CHECK-NEXT: ret;
+define <3 x i8> @test_v3i8(<3 x i8> %a) {
+       %r = tail call <3 x i8> @test_v3i8(<3 x i8> %a);
+       ret <3 x i8> %r;
+}
+
+; CHECK: .func  (.param .align 4 .b8 func_retval0[4])
+; CHECK-LABEL: test_v4i8(
+; CHECK-NEXT: .param .align 4 .b8 test_v4i8_param_0[4]
+; CHECK:      ld.param.v4.u8 {[[E0:%rs[0-9]+]], [[E1:%rs[0-9]+]], [[E2:%rs[0-9]+]], [[E3:%rs[0-9]+]]}, [test_v4i8_param_0]
+; CHECK:      .param .align 4 .b8 param0[4];
+; CHECK:      st.param.v4.b8  [param0+0], {[[E0]], [[E1]], [[E2]], [[E3]]};
+; CHECK:      .param .align 4 .b8 retval0[4];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v4i8,
+; CHECK:      ld.param.v4.b8  {[[RE0:%rs[0-9]+]], [[RE1:%rs[0-9]+]], [[RE2:%rs[0-9]+]], [[RE3:%rs[0-9]+]]}, [retval0+0];
+; CHECK:      st.param.v4.b8 [func_retval0+0], {[[RE0]], [[RE1]], [[RE2]], [[RE3]]}
+; CHECK-NEXT: ret;
+define <4 x i8> @test_v4i8(<4 x i8> %a) {
+       %r = tail call <4 x i8> @test_v4i8(<4 x i8> %a);
+       ret <4 x i8> %r;
+}
+
+; CHECK: .func  (.param .align 8 .b8 func_retval0[8])
+; CHECK-LABEL: test_v5i8(
+; CHECK-NEXT: .param .align 8 .b8 test_v5i8_param_0[8]
+; CHECK-DAG:  ld.param.u8     [[E4:%rs[0-9]+]], [test_v5i8_param_0+4];
+; CHECK-DAG   ld.param.v4.u8  {[[E0:%rs[0-9]+]], [[E1:%rs[0-9]+]], [[E2:%rs[0-9]+]], [[E3:%rs[0-9]+]]}, [test_v5i8_param_0]
+; CHECK:      .param .align 8 .b8 param0[8];
+; CHECK-DAG:  st.param.v4.b8  [param0+0], {[[E0]], [[E1]], [[E2]], [[E3]]};
+; CHECK-DAG:  st.param.b8     [param0+4], [[E4]];
+; CHECK:      .param .align 8 .b8 retval0[8];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v5i8,
+; CHECK-DAG:  ld.param.v4.b8  {[[RE0:%rs[0-9]+]], [[RE1:%rs[0-9]+]], [[RE2:%rs[0-9]+]], [[RE3:%rs[0-9]+]]}, [retval0+0];
+; CHECK-DAG:  ld.param.b8     [[RE4:%rs[0-9]+]], [retval0+4];
+; CHECK-DAG:  st.param.v4.b8  [func_retval0+0], {[[RE0]], [[RE1]], [[RE2]], [[RE3]]}
+; CHECK-DAG:  st.param.b8     [func_retval0+4], [[RE4]];
+; CHECK-NEXT: ret;
+define <5 x i8> @test_v5i8(<5 x i8> %a) {
+       %r = tail call <5 x i8> @test_v5i8(<5 x i8> %a);
+       ret <5 x i8> %r;
+}
+
+; CHECK: .func  (.param .b32 func_retval0)
+; CHECK-LABEL: test_i16(
+; CHECK-NEXT: .param .b32 test_i16_param_0
+; CHECK:      ld.param.u16    [[E16:%rs[0-9]+]], [test_i16_param_0];
+; CHECK:      cvt.u32.u16     [[E32:%r[0-9]+]], [[E16]];
+; CHECK:      .param .b32 param0;
+; CHECK:      st.param.b32    [param0+0], [[E32]];
+; CHECK:      .param .b32 retval0;
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_i16,
+; CHECK:      ld.param.b32    [[RE32:%r[0-9]+]], [retval0+0];
+; CHECK:      and.b32         [[R:%r[0-9]+]], [[RE32]], 65535;
+; CHECK:      st.param.b32    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define i16 @test_i16(i16 %a) {
+       %r = tail call i16 @test_i16(i16 %a);
+       ret i16 %r;
+}
+
+; CHECK: .func  (.param .b32 func_retval0)
+; CHECK-LABEL: test_i16s(
+; CHECK-NEXT: .param .b32 test_i16s_param_0
+; CHECK:      ld.param.u16    [[E16:%rs[0-9]+]], [test_i16s_param_0];
+; CHECK:      cvt.s32.s16     [[E32:%r[0-9]+]], [[E16]];
+; CHECK:      .param .b32 param0;
+; CHECK:      st.param.b32    [param0+0], [[E32]];
+; CHECK:      .param .b32 retval0;
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_i16s,
+; CHECK:      ld.param.b32    [[RE32:%r[0-9]+]], [retval0+0];
+; CHECK:      cvt.s32.s16     [[R:%r[0-9]+]], [[RE32]];
+; CHECK:      st.param.b32    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define signext i16 @test_i16s(i16 signext %a) {
+       %r = tail call signext i16 @test_i16s(i16 signext %a);
+       ret i16 %r;
+}
+
+; CHECK: .func  (.param .align 8 .b8 func_retval0[8])
+; CHECK-LABEL: test_v3i16(
+; CHECK-NEXT: .param .align 8 .b8 test_v3i16_param_0[8]
+; CHECK-DAG:  ld.param.u16    [[E2:%rs[0-9]+]], [test_v3i16_param_0+4];
+; CHECK-DAG:  ld.param.v2.u16 {[[E0:%rs[0-9]+]], [[E1:%rs[0-9]+]]}, [test_v3i16_param_0];
+; CHECK:      .param .align 8 .b8 param0[8];
+; CHECK:      st.param.v2.b16 [param0+0], {[[E0]], [[E1]]};
+; CHECK:      st.param.b16    [param0+4], [[E2]];
+; CHECK:      .param .align 8 .b8 retval0[8];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v3i16,
+; CHECK:      ld.param.v2.b16 {[[RE0:%rs[0-9]+]], [[RE1:%rs[0-9]+]]}, [retval0+0];
+; CHECK:      ld.param.b16    [[RE2:%rs[0-9]+]], [retval0+4];
+; CHECK-DAG:  st.param.v2.b16 [func_retval0+0], {[[RE0]], [[RE1]]};
+; CHECK-DAG:  st.param.b16    [func_retval0+4], [[RE2]];
+; CHECK-NEXT: ret;
+define <3 x i16> @test_v3i16(<3 x i16> %a) {
+       %r = tail call <3 x i16> @test_v3i16(<3 x i16> %a);
+       ret <3 x i16> %r;
+}
+
+; CHECK: .func  (.param .align 8 .b8 func_retval0[8])
+; CHECK-LABEL: test_v4i16(
+; CHECK-NEXT: .param .align 8 .b8 test_v4i16_param_0[8]
+; CHECK:      ld.param.v4.u16 {[[E0:%rs[0-9]+]], [[E1:%rs[0-9]+]], [[E2:%rs[0-9]+]], [[E3:%rs[0-9]+]]}, [test_v4i16_param_0]
+; CHECK:      .param .align 8 .b8 param0[8];
+; CHECK:      st.param.v4.b16 [param0+0], {[[E0]], [[E1]], [[E2]], [[E3]]};
+; CHECK:      .param .align 8 .b8 retval0[8];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v4i16,
+; CHECK:      ld.param.v4.b16 {[[RE0:%rs[0-9]+]], [[RE1:%rs[0-9]+]], [[RE2:%rs[0-9]+]], [[RE3:%rs[0-9]+]]}, [retval0+0];
+; CHECK:      st.param.v4.b16 [func_retval0+0], {[[RE0]], [[RE1]], [[RE2]], [[RE3]]}
+; CHECK-NEXT: ret;
+define <4 x i16> @test_v4i16(<4 x i16> %a) {
+       %r = tail call <4 x i16> @test_v4i16(<4 x i16> %a);
+       ret <4 x i16> %r;
+}
+
+; CHECK: .func  (.param .align 16 .b8 func_retval0[16])
+; CHECK-LABEL: test_v5i16(
+; CHECK-NEXT: .param .align 16 .b8 test_v5i16_param_0[16]
+; CHECK-DAG:  ld.param.u16    [[E4:%rs[0-9]+]], [test_v5i16_param_0+8];
+; CHECK-DAG   ld.param.v4.u16 {[[E0:%rs[0-9]+]], [[E1:%rs[0-9]+]], [[E2:%rs[0-9]+]], [[E3:%rs[0-9]+]]}, [test_v5i16_param_0]
+; CHECK:      .param .align 16 .b8 param0[16];
+; CHECK-DAG:  st.param.v4.b16 [param0+0], {[[E0]], [[E1]], [[E2]], [[E3]]};
+; CHECK-DAG:  st.param.b16    [param0+8], [[E4]];
+; CHECK:      .param .align 16 .b8 retval0[16];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v5i16,
+; CHECK-DAG:  ld.param.v4.b16 {[[RE0:%rs[0-9]+]], [[RE1:%rs[0-9]+]], [[RE2:%rs[0-9]+]], [[RE3:%rs[0-9]+]]}, [retval0+0];
+; CHECK-DAG:  ld.param.b16    [[RE4:%rs[0-9]+]], [retval0+8];
+; CHECK-DAG:  st.param.v4.b16 [func_retval0+0], {[[RE0]], [[RE1]], [[RE2]], [[RE3]]}
+; CHECK-DAG:  st.param.b16    [func_retval0+8], [[RE4]];
+; CHECK-NEXT: ret;
+define <5 x i16> @test_v5i16(<5 x i16> %a) {
+       %r = tail call <5 x i16> @test_v5i16(<5 x i16> %a);
+       ret <5 x i16> %r;
+}
+
+; CHECK: .func  (.param .b32 func_retval0)
+; CHECK-LABEL: test_half(
+; CHECK-NEXT: .param .b32 test_half_param_0
+; CHECK:      ld.param.b16    [[E:%h[0-9]+]], [test_half_param_0];
+; CHECK:      .param .b32 param0;
+; CHECK:      st.param.b16    [param0+0], [[E]];
+; CHECK:      .param .b32 retval0;
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_half,
+; CHECK:      ld.param.b16    [[R:%h[0-9]+]], [retval0+0];
+; CHECK:      st.param.b16    [func_retval0+0], [[R]]
+; CHECK-NEXT: ret;
+define half @test_half(half %a) {
+       %r = tail call half @test_half(half %a);
+       ret half %r;
+}
+
+; CHECK: .func  (.param .b32 func_retval0)
+; CHECK-LABEL: test_i32(
+; CHECK-NEXT: .param .b32 test_i32_param_0
+; CHECK:      ld.param.u32    [[E:%r[0-9]+]], [test_i32_param_0];
+; CHECK:      .param .b32 param0;
+; CHECK:      st.param.b32    [param0+0], [[E]];
+; CHECK:      .param .b32 retval0;
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_i32,
+; CHECK:      ld.param.b32    [[R:%r[0-9]+]], [retval0+0];
+; CHECK:      st.param.b32    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define i32 @test_i32(i32 %a) {
+       %r = tail call i32 @test_i32(i32 %a);
+       ret i32 %r;
+}
+
+; CHECK: .func  (.param .align 16 .b8 func_retval0[16])
+; CHECK-LABEL: test_v3i32(
+; CHECK-NEXT: .param .align 16 .b8 test_v3i32_param_0[16]
+; CHECK-DAG:  ld.param.u32     [[E2:%r[0-9]+]], [test_v3i32_param_0+8];
+; CHECK-DAG:  ld.param.v2.u32  {[[E0:%r[0-9]+]], [[E1:%r[0-9]+]]}, [test_v3i32_param_0];
+; CHECK:      .param .align 16 .b8 param0[16];
+; CHECK:      st.param.v2.b32  [param0+0], {[[E0]], [[E1]]};
+; CHECK:      st.param.b32     [param0+8], [[E2]];
+; CHECK:      .param .align 16 .b8 retval0[16];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v3i32,
+; CHECK:      ld.param.v2.b32  {[[RE0:%r[0-9]+]], [[RE1:%r[0-9]+]]}, [retval0+0];
+; CHECK:      ld.param.b32     [[RE2:%r[0-9]+]], [retval0+8];
+; CHECK-DAG:  st.param.v2.b32  [func_retval0+0], {[[RE0]], [[RE1]]};
+; CHECK-DAG:  st.param.b32     [func_retval0+8], [[RE2]];
+; CHECK-NEXT: ret;
+define <3 x i32> @test_v3i32(<3 x i32> %a) {
+       %r = tail call <3 x i32> @test_v3i32(<3 x i32> %a);
+       ret <3 x i32> %r;
+}
+
+; CHECK: .func  (.param .align 16 .b8 func_retval0[16])
+; CHECK-LABEL: test_v4i32(
+; CHECK-NEXT: .param .align 16 .b8 test_v4i32_param_0[16]
+; CHECK:      ld.param.v4.u32  {[[E0:%r[0-9]+]], [[E1:%r[0-9]+]], [[E2:%r[0-9]+]], [[E3:%r[0-9]+]]}, [test_v4i32_param_0]
+; CHECK:      .param .align 16 .b8 param0[16];
+; CHECK:      st.param.v4.b32  [param0+0], {[[E0]], [[E1]], [[E2]], [[E3]]};
+; CHECK:      .param .align 16 .b8 retval0[16];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v4i32,
+; CHECK:      ld.param.v4.b32  {[[RE0:%r[0-9]+]], [[RE1:%r[0-9]+]], [[RE2:%r[0-9]+]], [[RE3:%r[0-9]+]]}, [retval0+0];
+; CHECK:      st.param.v4.b32  [func_retval0+0], {[[RE0]], [[RE1]], [[RE2]], [[RE3]]}
+; CHCK-NEXT: ret;
+define <4 x i32> @test_v4i32(<4 x i32> %a) {
+       %r = tail call <4 x i32> @test_v4i32(<4 x i32> %a);
+       ret <4 x i32> %r;
+}
+
+; CHECK: .func  (.param .align 32 .b8 func_retval0[32])
+; CHECK-LABEL: test_v5i32(
+; CHECK-NEXT: .param .align 32 .b8 test_v5i32_param_0[32]
+; CHECK-DAG:  ld.param.u32     [[E4:%r[0-9]+]], [test_v5i32_param_0+16];
+; CHECK-DAG   ld.param.v4.u32  {[[E0:%r[0-9]+]], [[E1:%r[0-9]+]], [[E2:%r[0-9]+]], [[E3:%r[0-9]+]]}, [test_v5i32_param_0]
+; CHECK:      .param .align 32 .b8 param0[32];
+; CHECK-DAG:  st.param.v4.b32  [param0+0], {[[E0]], [[E1]], [[E2]], [[E3]]};
+; CHECK-DAG:  st.param.b32     [param0+16], [[E4]];
+; CHECK:      .param .align 32 .b8 retval0[32];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v5i32,
+; CHECK-DAG:  ld.param.v4.b32  {[[RE0:%r[0-9]+]], [[RE1:%r[0-9]+]], [[RE2:%r[0-9]+]], [[RE3:%r[0-9]+]]}, [retval0+0];
+; CHECK-DAG:  ld.param.b32     [[RE4:%r[0-9]+]], [retval0+16];
+; CHECK-DAG:  st.param.v4.b32  [func_retval0+0], {[[RE0]], [[RE1]], [[RE2]], [[RE3]]}
+; CHECK-DAG:  st.param.b32     [func_retval0+16], [[RE4]];
+; CHECK-NEXT: ret;
+define <5 x i32> @test_v5i32(<5 x i32> %a) {
+       %r = tail call <5 x i32> @test_v5i32(<5 x i32> %a);
+       ret <5 x i32> %r;
+}
+
+; CHECK: .func  (.param .b32 func_retval0)
+; CHECK-LABEL: test_float(
+; CHECK-NEXT: .param .b32 test_float_param_0
+; CHECK:      ld.param.f32    [[E:%f[0-9]+]], [test_float_param_0];
+; CHECK:      .param .b32 param0;
+; CHECK:      st.param.f32    [param0+0], [[E]];
+; CHECK:      .param .b32 retval0;
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_float,
+; CHECK:      ld.param.f32    [[R:%f[0-9]+]], [retval0+0];
+; CHECK:      st.param.f32    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define float @test_float(float %a) {
+       %r = tail call float @test_float(float %a);
+       ret float %r;
+}
+
+; CHECK: .func  (.param .b64 func_retval0)
+; CHECK-LABEL: test_i64(
+; CHECK-NEXT: .param .b64 test_i64_param_0
+; CHECK:      ld.param.u64    [[E:%rd[0-9]+]], [test_i64_param_0];
+; CHECK:      .param .b64 param0;
+; CHECK:      st.param.b64    [param0+0], [[E]];
+; CHECK:      .param .b64 retval0;
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_i64,
+; CHECK:      ld.param.b64    [[R:%rd[0-9]+]], [retval0+0];
+; CHECK:      st.param.b64    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define i64 @test_i64(i64 %a) {
+       %r = tail call i64 @test_i64(i64 %a);
+       ret i64 %r;
+}
+
+; CHECK: .func  (.param .align 32 .b8 func_retval0[32])
+; CHECK-LABEL: test_v3i64(
+; CHECK-NEXT: .param .align 32 .b8 test_v3i64_param_0[32]
+; CHECK-DAG:  ld.param.u64     [[E2:%rd[0-9]+]], [test_v3i64_param_0+16];
+; CHECK-DAG:  ld.param.v2.u64  {[[E0:%rd[0-9]+]], [[E1:%rd[0-9]+]]}, [test_v3i64_param_0];
+; CHECK:      .param .align 32 .b8 param0[32];
+; CHECK:      st.param.v2.b64  [param0+0], {[[E0]], [[E1]]};
+; CHECK:      st.param.b64     [param0+16], [[E2]];
+; CHECK:      .param .align 32 .b8 retval0[32];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v3i64,
+; CHECK:      ld.param.v2.b64  {[[RE0:%rd[0-9]+]], [[RE1:%rd[0-9]+]]}, [retval0+0];
+; CHECK:      ld.param.b64     [[RE2:%rd[0-9]+]], [retval0+16];
+; CHECK-DAG:  st.param.v2.b64  [func_retval0+0], {[[RE0]], [[RE1]]};
+; CHECK-DAG:  st.param.b64     [func_retval0+16], [[RE2]];
+; CHECK-DAG:  st.param.v2.b64  [func_retval0+0], {[[RE0]], [[RE1]]};
+; CHECK-DAG:  st.param.b64     [func_retval0+16], [[RE2]];
+; CHECK-NEXT: ret;
+define <3 x i64> @test_v3i64(<3 x i64> %a) {
+       %r = tail call <3 x i64> @test_v3i64(<3 x i64> %a);
+       ret <3 x i64> %r;
+}
+
+; For i64 vector loads are limited by PTX to 2 elements.
+; CHECK: .func  (.param .align 32 .b8 func_retval0[32])
+; CHECK-LABEL: test_v4i64(
+; CHECK-NEXT: .param .align 32 .b8 test_v4i64_param_0[32]
+; CHECK-DAG:  ld.param.v2.u64  {[[E2:%rd[0-9]+]], [[E3:%rd[0-9]+]]}, [test_v4i64_param_0+16];
+; CHECK-DAG:  ld.param.v2.u64  {[[E0:%rd[0-9]+]], [[E1:%rd[0-9]+]]}, [test_v4i64_param_0];
+; CHECK:      .param .align 32 .b8 param0[32];
+; CHECK:      st.param.v2.b64  [param0+0], {[[E0]], [[E1]]};
+; CHECK:      st.param.v2.b64  [param0+16], {[[E2]], [[E3]]};
+; CHECK:      .param .align 32 .b8 retval0[32];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_v4i64,
+; CHECK:      ld.param.v2.b64  {[[RE0:%rd[0-9]+]], [[RE1:%rd[0-9]+]]}, [retval0+0];
+; CHECK:      ld.param.v2.b64  {[[RE2:%rd[0-9]+]], [[RE3:%rd[0-9]+]]}, [retval0+16];
+; CHECK-DAG:  st.param.v2.b64  [func_retval0+16], {[[RE2]], [[RE3]]};
+; CHECK-DAG:  st.param.v2.b64  [func_retval0+0], {[[RE0]], [[RE1]]};
+; CHECK-NEXT: ret;
+define <4 x i64> @test_v4i64(<4 x i64> %a) {
+       %r = tail call <4 x i64> @test_v4i64(<4 x i64> %a);
+       ret <4 x i64> %r;
+}
+
+; Aggregates, on the other hand, do not get extended.
+
+; CHECK: .func  (.param .align 1 .b8 func_retval0[1])
+; CHECK-LABEL: test_s_i1(
+; CHECK-NEXT: .align 1 .b8 test_s_i1_param_0[1]
+; CHECK:      ld.param.u8 [[A:%rs[0-9]+]], [test_s_i1_param_0];
+; CHECK:      .param .align 1 .b8 param0[1];
+; CHECK:      st.param.b8    [param0+0], [[A]]
+; CHECK:      .param .align 1 .b8 retval0[1];
+; CHECK:      call.uni
+; CHECK-NEXT: test_s_i1,
+; CHECK:      ld.param.b8    [[R:%rs[0-9]+]], [retval0+0];
+; CHECK:      st.param.b8    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define %s_i1 @test_s_i1(%s_i1 %a) {
+       %r = tail call %s_i1 @test_s_i1(%s_i1 %a);
+       ret %s_i1 %r;
+}
+
+; CHECK: .func  (.param .align 1 .b8 func_retval0[1])
+; CHECK-LABEL: test_s_i8(
+; CHECK-NEXT: .param .align 1 .b8 test_s_i8_param_0[1]
+; CHECK:      ld.param.u8 [[A:%rs[0-9]+]], [test_s_i8_param_0];
+; CHECK:      .param .align 1 .b8 param0[1];
+; CHECK:      st.param.b8    [param0+0], [[A]]
+; CHECK:      .param .align 1 .b8 retval0[1];
+; CHECK:      call.uni
+; CHECK-NEXT: test_s_i8,
+; CHECK:      ld.param.b8    [[R:%rs[0-9]+]], [retval0+0];
+; CHECK:      st.param.b8    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define %s_i8 @test_s_i8(%s_i8 %a) {
+       %r = tail call %s_i8 @test_s_i8(%s_i8 %a);
+       ret %s_i8 %r;
+}
+
+; CHECK: .func  (.param .align 2 .b8 func_retval0[2])
+; CHECK-LABEL: test_s_i16(
+; CHECK-NEXT: .param .align 2 .b8 test_s_i16_param_0[2]
+; CHECK:      ld.param.u16 [[A:%rs[0-9]+]], [test_s_i16_param_0];
+; CHECK:      .param .align 2 .b8 param0[2];
+; CHECK:      st.param.b16    [param0+0], [[A]]
+; CHECK:      .param .align 2 .b8 retval0[2];
+; CHECK:      call.uni
+; CHECK-NEXT: test_s_i16,
+; CHECK:      ld.param.b16    [[R:%rs[0-9]+]], [retval0+0];
+; CHECK:      st.param.b16    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define %s_i16 @test_s_i16(%s_i16 %a) {
+       %r = tail call %s_i16 @test_s_i16(%s_i16 %a);
+       ret %s_i16 %r;
+}
+
+; CHECK: .func  (.param .align 2 .b8 func_retval0[2])
+; CHECK-LABEL: test_s_half(
+; CHECK-NEXT: .param .align 2 .b8 test_s_half_param_0[2]
+; CHECK:      ld.param.b16 [[A:%h[0-9]+]], [test_s_half_param_0];
+; CHECK:      .param .align 2 .b8 param0[2];
+; CHECK:      st.param.b16    [param0+0], [[A]]
+; CHECK:      .param .align 2 .b8 retval0[2];
+; CHECK:      call.uni
+; CHECK-NEXT: test_s_half,
+; CHECK:      ld.param.b16    [[R:%h[0-9]+]], [retval0+0];
+; CHECK:      st.param.b16    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define %s_half @test_s_half(%s_half %a) {
+       %r = tail call %s_half @test_s_half(%s_half %a);
+       ret %s_half %r;
+}
+
+; CHECK: .func  (.param .align 4 .b8 func_retval0[4])
+; CHECK-LABEL: test_s_i32(
+; CHECK-NEXT: .param .align 4 .b8 test_s_i32_param_0[4]
+; CHECK:      ld.param.u32    [[E:%r[0-9]+]], [test_s_i32_param_0];
+; CHECK:      .param .align 4 .b8 param0[4]
+; CHECK:      st.param.b32    [param0+0], [[E]];
+; CHECK:      .param .align 4 .b8 retval0[4];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_s_i32,
+; CHECK:      ld.param.b32    [[R:%r[0-9]+]], [retval0+0];
+; CHECK:      st.param.b32    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define %s_i32 @test_s_i32(%s_i32 %a) {
+       %r = tail call %s_i32 @test_s_i32(%s_i32 %a);
+       ret %s_i32 %r;
+}
+
+; CHECK: .func  (.param .align 4 .b8 func_retval0[4])
+; CHECK-LABEL: test_s_float(
+; CHECK-NEXT: .param .align 4 .b8 test_s_float_param_0[4]
+; CHECK:      ld.param.f32    [[E:%f[0-9]+]], [test_s_float_param_0];
+; CHECK:      .param .align 4 .b8 param0[4]
+; CHECK:      st.param.f32    [param0+0], [[E]];
+; CHECK:      .param .align 4 .b8 retval0[4];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_s_float,
+; CHECK:      ld.param.f32    [[R:%f[0-9]+]], [retval0+0];
+; CHECK:      st.param.f32    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define %s_float @test_s_float(%s_float %a) {
+       %r = tail call %s_float @test_s_float(%s_float %a);
+       ret %s_float %r;
+}
+
+; CHECK: .func  (.param .align 8 .b8 func_retval0[8])
+; CHECK-LABEL: test_s_i64(
+; CHECK-NEXT: .param .align 8 .b8 test_s_i64_param_0[8]
+; CHECK:      ld.param.u64    [[E:%rd[0-9]+]], [test_s_i64_param_0];
+; CHECK:      .param .align 8 .b8 param0[8];
+; CHECK:      st.param.b64    [param0+0], [[E]];
+; CHECK:      .param .align 8 .b8 retval0[8];
+; CHECK:      call.uni (retval0),
+; CHECK-NEXT: test_s_i64,
+; CHECK:      ld.param.b64    [[R:%rd[0-9]+]], [retval0+0];
+; CHECK:      st.param.b64    [func_retval0+0], [[R]];
+; CHECK-NEXT: ret;
+define %s_i64 @test_s_i64(%s_i64 %a) {
+       %r = tail call %s_i64 @test_s_i64(%s_i64 %a);
+       ret %s_i64 %r;
+}
+
+; Fields that have different types, but identical sizes are not vectorized.
+; CHECK: .func  (.param .align 8 .b8 func_retval0[24])
+; CHECK-LABEL: test_s_i32f32(
+; CHECK:        .param .align 8 .b8 test_s_i32f32_param_0[24]
+; CHECK-DAG:    ld.param.u64    [[E4:%rd[0-9]+]], [test_s_i32f32_param_0+16];
+; CHECK-DAG:    ld.param.f32    [[E3:%f[0-9]+]], [test_s_i32f32_param_0+12];
+; CHECK-DAG:    ld.param.u32    [[E2:%r[0-9]+]], [test_s_i32f32_param_0+8];
+; CHECK-DAG:    ld.param.f32    [[E1:%f[0-9]+]], [test_s_i32f32_param_0+4];
+; CHECK-DAG:    ld.param.u32    [[E0:%r[0-9]+]], [test_s_i32f32_param_0];
+; CHECK:        .param .align 8 .b8 param0[24];
+; CHECK-DAG:    st.param.b32    [param0+0], [[E0]];
+; CHECK-DAG:    st.param.f32    [param0+4], [[E1]];
+; CHECK-DAG:    st.param.b32    [param0+8], [[E2]];
+; CHECK-DAG:    st.param.f32    [param0+12], [[E3]];
+; CHECK-DAG:    st.param.b64    [param0+16], [[E4]];
+; CHECK:        .param .align 8 .b8 retval0[24];
+; CHECK:        call.uni (retval0),
+; CHECK-NEXT:   test_s_i32f32,
+; CHECK-DAG:    ld.param.b32    [[RE0:%r[0-9]+]], [retval0+0];
+; CHECK-DAG:    ld.param.f32    [[RE1:%f[0-9]+]], [retval0+4];
+; CHECK-DAG:    ld.param.b32    [[RE2:%r[0-9]+]], [retval0+8];
+; CHECK-DAG:    ld.param.f32    [[RE3:%f[0-9]+]], [retval0+12];
+; CHECK-DAG:    ld.param.b64    [[RE4:%rd[0-9]+]], [retval0+16];
+; CHECK-DAG:    st.param.b32    [func_retval0+0], [[RE0]];
+; CHECK-DAG:    st.param.f32    [func_retval0+4], [[RE1]];
+; CHECK-DAG:    st.param.b32    [func_retval0+8], [[RE2]];
+; CHECK-DAG:    st.param.f32    [func_retval0+12], [[RE3]];
+; CHECK-DAG:    st.param.b64    [func_retval0+16], [[RE4]];
+; CHECK:        ret;
+define %s_i32f32 @test_s_i32f32(%s_i32f32 %a) {
+       %r = tail call %s_i32f32 @test_s_i32f32(%s_i32f32 %a);
+       ret %s_i32f32 %r;
+}
+
+; We do vectorize consecutive fields with matching types.
+; CHECK:.visible .func  (.param .align 8 .b8 func_retval0[24])
+; CHECK-LABEL: test_s_i32x4(
+; CHECK:        .param .align 8 .b8 test_s_i32x4_param_0[24]
+; CHECK-DAG:    ld.param.u64    [[RD1:%rd[0-9]+]], [test_s_i32x4_param_0+16];
+; CHECK-DAG:    ld.param.v2.u32 {[[E2:%r[0-9]+]], [[E3:%r[0-9]+]]}, [test_s_i32x4_param_0+8];
+; CHECK-DAG:    ld.param.v2.u32 {[[E0:%r[0-9]+]], [[E1:%r[0-9]+]]}, [test_s_i32x4_param_0];
+; CHECK:        .param .align 8 .b8 param0[24];
+; CHECK:        st.param.v2.b32 [param0+0], {[[E0]], [[E1]]};
+; CHECK:        st.param.v2.b32 [param0+8], {[[E2]], [[E3]]};
+; CHECK:        st.param.b64    [param0+16], [[E4]];
+; CHECK:        .param .align 8 .b8 retval0[24];
+; CHECK:        call.uni (retval0),
+; CHECK-NEXT:   test_s_i32x4,
+; CHECK:        ld.param.v2.b32 {[[RE0:%r[0-9]+]], [[RE1:%r[0-9]+]]}, [retval0+0];
+; CHECK:        ld.param.v2.b32 {[[RE2:%r[0-9]+]], [[RE3:%r[0-9]+]]}, [retval0+8];
+; CHECK:        ld.param.b64    [[RE4:%rd[0-9]+]], [retval0+16];
+; CHECK-DAG:    st.param.v2.b32 [func_retval0+0], {[[RE0]], [[RE1]]};
+; CHECK-DAG:    st.param.v2.b32 [func_retval0+8], {[[RE2]], [[RE3]]};
+; CHECK-DAG:    st.param.b64    [func_retval0+16], [[RE4]];
+; CHECK:        ret;
+
+define %s_i32x4 @test_s_i32x4(%s_i32x4 %a) {
+       %r = tail call %s_i32x4 @test_s_i32x4(%s_i32x4 %a);
+       ret %s_i32x4 %r;
+}
+
+; CHECK:.visible .func  (.param .align 8 .b8 func_retval0[32])
+; CHECK-LABEL: test_s_i1i32x4(
+; CHECK:        .param .align 8 .b8 test_s_i1i32x4_param_0[32]
+; CHECK:        ld.param.u64    [[E5:%rd[0-9]+]], [test_s_i1i32x4_param_0+24];
+; CHECK:        ld.param.u32    [[E4:%r[0-9]+]], [test_s_i1i32x4_param_0+16];
+; CHECK:        ld.param.u32    [[E3:%r[0-9]+]], [test_s_i1i32x4_param_0+12];
+; CHECK:        ld.param.u8     [[E2:%rs[0-9]+]], [test_s_i1i32x4_param_0+8];
+; CHECK:        ld.param.v2.u32         {[[E0:%r[0-9]+]], [[E1:%r[0-9]+]]}, [test_s_i1i32x4_param_0];
+; CHECK:        .param .align 8 .b8 param0[32];
+; CHECK:        st.param.v2.b32 [param0+0], {[[E0]], [[E1]]};
+; CHECK:        st.param.b8     [param0+8], [[E2]];
+; CHECK:        st.param.b32    [param0+12], [[E3]];
+; CHECK:        st.param.b32    [param0+16], [[E4]];
+; CHECK:        st.param.b64    [param0+24], [[E5]];
+; CHECK:        .param .align 8 .b8 retval0[32];
+; CHECK:        call.uni (retval0),
+; CHECK:        test_s_i1i32x4,
+; CHECK:        (
+; CHECK:        param0
+; CHECK:        );
+; CHECK:        ld.param.v2.b32 {[[RE0:%r[0-9]+]], [[RE1:%r[0-9]+]]}, [retval0+0];
+; CHECK:        ld.param.b8     [[RE2:%rs[0-9]+]], [retval0+8];
+; CHECK:        ld.param.b32    [[RE3:%r[0-9]+]], [retval0+12];
+; CHECK:        ld.param.b32    [[RE4:%r[0-9]+]], [retval0+16];
+; CHECK:        ld.param.b64    [[RE5:%rd[0-9]+]], [retval0+24];
+; CHECK:        st.param.v2.b32 [func_retval0+0], {[[RE0]], [[RE1]]};
+; CHECK:        st.param.b8     [func_retval0+8], [[RE2]];
+; CHECK:        st.param.b32    [func_retval0+12], [[RE3]];
+; CHECK:        st.param.b32    [func_retval0+16], [[RE4]];
+; CHECK:        st.param.b64    [func_retval0+24], [[RE5]];
+; CHECK:        ret;
+
+define %s_i8i32x4 @test_s_i1i32x4(%s_i8i32x4 %a) {
+       %r = tail call %s_i8i32x4 @test_s_i1i32x4(%s_i8i32x4 %a);
+       ret %s_i8i32x4 %r;
+}
+
+; -- All loads/stores from parameters aligned by one must be done one
+; -- byte at a time.
+; CHECK:.visible .func  (.param .align 1 .b8 func_retval0[25])
+; CHECK-LABEL: test_s_i1i32x4p(
+; CHECK-DAG:        .param .align 1 .b8 test_s_i1i32x4p_param_0[25]
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+24];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+23];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+22];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+21];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+20];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+19];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+18];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+17];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+16];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+15];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+14];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+13];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+12];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+11];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+10];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+9];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+8];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+7];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+6];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+5];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+4];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+3];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+2];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0+1];
+; CHECK-DAG:        ld.param.u8     %r{{.*}}, [test_s_i1i32x4p_param_0];
+; --- TODO
+; --- Unaligned parameter store/ return value load is broken in both nvcc
+; --- and llvm and needs to be fixed.
+; CHECK:        .param .align 1 .b8 param0[25];
+; CHECK-DAG:        st.param.b32    [param0+0],
+; CHECK-DAG:        st.param.b32    [param0+4],
+; CHECK-DAG:        st.param.b8     [param0+8],
+; CHECK-DAG:        st.param.b32    [param0+9],
+; CHECK-DAG:        st.param.b32    [param0+13],
+; CHECK-DAG:        st.param.b64    [param0+17],
+; CHECK:            .param .align 1 .b8 retval0[25];
+; CHECK:            call.uni (retval0),
+; CHECK-NEXT:       test_s_i1i32x4p,
+; CHECK-DAG:        ld.param.b32    %r41, [retval0+0];
+; CHECK-DAG:        ld.param.b32    %r42, [retval0+4];
+; CHECK-DAG:        ld.param.b8     %rs2, [retval0+8];
+; CHECK-DAG:        ld.param.b32    %r43, [retval0+9];
+; CHECK-DAG:        ld.param.b32    %r44, [retval0+13];
+; CHECK-DAG:        ld.param.b64    %rd23, [retval0+17];
+; CHECK-DAG:        st.param.b32    [func_retval0+0],
+; CHECK-DAG:        st.param.b32    [func_retval0+4],
+; CHECK-DAG:        st.param.b8     [func_retval0+8],
+; CHECK-DAG:        st.param.b32    [func_retval0+9],
+; CHECK-DAG:        st.param.b32    [func_retval0+13],
+; CHECK-DAG:        st.param.b64    [func_retval0+17],
+
+define %s_i8i32x4p @test_s_i1i32x4p(%s_i8i32x4p %a) {
+       %r = tail call %s_i8i32x4p @test_s_i1i32x4p(%s_i8i32x4p %a);
+       ret %s_i8i32x4p %r;
+}
+
+; Check that we can vectorize loads that span multiple aggregate fields.
+; CHECK:.visible .func  (.param .align 16 .b8 func_retval0[80])
+; CHECK-LABEL: test_s_crossfield(
+; CHECK:        .param .align 16 .b8 test_s_crossfield_param_0[80]
+; CHECK:        ld.param.u32    [[E15:%r[0-9]+]], [test_s_crossfield_param_0+64];
+; CHECK:        ld.param.v4.u32 {[[E11:%r[0-9]+]], [[E12:%r[0-9]+]], [[E13:%r[0-9]+]], [[E14:%r[0-9]+]]}, [test_s_crossfield_param_0+48];
+; CHECK:        ld.param.v4.u32 {[[E7:%r[0-9]+]], [[E8:%r[0-9]+]], [[E9:%r[0-9]+]], [[E10:%r[0-9]+]]}, [test_s_crossfield_param_0+32];
+; CHECK:        ld.param.v4.u32 {[[E3:%r[0-9]+]], [[E4:%r[0-9]+]], [[E5:%r[0-9]+]], [[E6:%r[0-9]+]]}, [test_s_crossfield_param_0+16];
+; CHECK:        ld.param.u32    [[E2:%r[0-9]+]], [test_s_crossfield_param_0+8];
+; CHECK:        ld.param.v2.u32 {[[E0:%r[0-9]+]], [[E1:%r[0-9]+]]}, [test_s_crossfield_param_0];
+; CHECK:        .param .align 16 .b8 param0[80];
+; CHECK:        st.param.v2.b32 [param0+0], {[[E0]], [[E1]]};
+; CHECK:        st.param.b32    [param0+8], [[E2]];
+; CHECK:        st.param.v4.b32 [param0+16], {[[E3]], [[E4]], [[E5]], [[E6]]};
+; CHECK:        st.param.v4.b32 [param0+32], {[[E7]], [[E8]], [[E9]], [[E10]]};
+; CHECK:        st.param.v4.b32 [param0+48], {[[E11]], [[E12]], [[E13]], [[E14]]};
+; CHECK:        st.param.b32    [param0+64], [[E15]];
+; CHECK:        .param .align 16 .b8 retval0[80];
+; CHECK:        call.uni (retval0),
+; CHECK:        test_s_crossfield,
+; CHECK:        ld.param.v2.b32 {[[RE0:%r[0-9]+]], [[RE1:%r[0-9]+]]}, [retval0+0];
+; CHECK:        ld.param.b32    [[RE2:%r[0-9]+]], [retval0+8];
+; CHECK:        ld.param.v4.b32 {[[RE3:%r[0-9]+]], [[RE4:%r[0-9]+]], [[RE5:%r[0-9]+]], [[RE6:%r[0-9]+]]}, [retval0+16];
+; CHECK:        ld.param.v4.b32 {[[RE7:%r[0-9]+]], [[RE8:%r[0-9]+]], [[RE9:%r[0-9]+]], [[RE10:%r[0-9]+]]}, [retval0+32];
+; CHECK:        ld.param.v4.b32 {[[RE11:%r[0-9]+]], [[RE12:%r[0-9]+]], [[RE13:%r[0-9]+]], [[RE14:%r[0-9]+]]}, [retval0+48];
+; CHECK:        ld.param.b32    [[RE15:%r[0-9]+]], [retval0+64];
+; CHECK:        st.param.v2.b32 [func_retval0+0], {[[RE0]], [[RE1]]};
+; CHECK:        st.param.b32    [func_retval0+8], [[RE2]];
+; CHECK:        st.param.v4.b32 [func_retval0+16], {[[RE3]], [[RE4]], [[RE5]], [[RE6]]};
+; CHECK:        st.param.v4.b32 [func_retval0+32], {[[RE7]], [[RE8]], [[RE9]], [[RE10]]};
+; CHECK:        st.param.v4.b32 [func_retval0+48], {[[RE11]], [[RE12]], [[RE13]], [[RE14]]};
+; CHECK:        st.param.b32    [func_retval0+64], [[RE15]];
+; CHECK:        ret;
+
+define %s_crossfield @test_s_crossfield(%s_crossfield %a) {
+       %r = tail call %s_crossfield @test_s_crossfield(%s_crossfield %a);
+       ret %s_crossfield %r;
+}

test/CodeGen/NVPTX/vec-param-load.ll

 ; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
 
 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
 
-
-define <16 x float> @foo(<16 x float> %a) {
-; Make sure we index into vectors properly
-; CHECK: ld.param.v4.f32         {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}, [foo_param_0+48];
-; CHECK: ld.param.v4.f32         {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}, [foo_param_0+32];
-; CHECK: ld.param.v4.f32         {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}, [foo_param_0+16];
-; CHECK: ld.param.v4.f32         {%f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}, %f{{[0-9]+}}}, [foo_param_0];
+define <16 x float> @test_v16f32(<16 x float> %a) {
+; CHECK-LABEL: test_v16f32(
+; CHECK-DAG: ld.param.v4.f32     {[[V_12_15:(%f[0-9]+[, ]*){4}]]}, [test_v16f32_param_0+48];
+; CHECK-DAG: ld.param.v4.f32     {[[V_8_11:(%f[0-9]+[, ]*){4}]]}, [test_v16f32_param_0+32];
+; CHECK-DAG: ld.param.v4.f32     {[[V_4_7:(%f[0-9]+[, ]*){4}]]}, [test_v16f32_param_0+16];
+; CHECK-DAG: ld.param.v4.f32     {[[V_0_3:(%f[0-9]+[, ]*){4}]]}, [test_v16f32_param_0];
+; CHECK-DAG: st.param.v4.f32     [func_retval0+0],  {[[V_0_3]]}
+; CHECK-DAG: st.param.v4.f32     [func_retval0+16], {[[V_4_7]]}
+; CHECK-DAG: st.param.v4.f32     [func_retval0+32], {[[V_8_11]]}
+; CHECK-DAG: st.param.v4.f32     [func_retval0+48], {[[V_12_15]]}
+; CHECK: ret;
   ret <16 x float> %a
 }
+
+define <8 x float> @test_v8f32(<8 x float> %a) {
+; CHECK-LABEL: test_v8f32(
+; CHECK-DAG: ld.param.v4.f32     {[[V_4_7:(%f[0-9]+[, ]*){4}]]}, [test_v8f32_param_0+16];
+; CHECK-DAG: ld.param.v4.f32     {[[V_0_3:(%f[0-9]+[, ]*){4}]]}, [test_v8f32_param_0];
+; CHECK-DAG: st.param.v4.f32     [func_retval0+0],  {[[V_0_3]]}
+; CHECK-DAG: st.param.v4.f32     [func_retval0+16], {[[V_4_7]]}
+; CHECK: ret;
+  ret <8 x float> %a
+}
+
+define <4 x float> @test_v4f32(<4 x float> %a) {
+; CHECK-LABEL: test_v4f32(
+; CHECK-DAG: ld.param.v4.f32     {[[V_0_3:(%f[0-9]+[, ]*){4}]]}, [test_v4f32_param_0];
+; CHECK-DAG: st.param.v4.f32     [func_retval0+0],  {[[V_0_3]]}
+; CHECK: ret;
+  ret <4 x float> %a
+}
+
+define <2 x float> @test_v2f32(<2 x float> %a) {
+; CHECK-LABEL: test_v2f32(
+; CHECK-DAG: ld.param.v2.f32     {[[V_0_3:(%f[0-9]+[, ]*){2}]]}, [test_v2f32_param_0];
+; CHECK-DAG: st.param.v2.f32     [func_retval0+0],  {[[V_0_3]]}
+; CHECK: ret;
+  ret <2 x float> %a
+}
+
+; Oddly shaped vectors should not load any extra elements.
+define <3 x float> @test_v3f32(<3 x float> %a) {
+; CHECK-LABEL: test_v3f32(
+; CHECK-DAG: ld.param.f32        [[V_2:%f[0-9]+]], [test_v3f32_param_0+8];
+; CHECK-DAG: ld.param.v2.f32     {[[V_0_1:(%f[0-9]+[, ]*){2}]]}, [test_v3f32_param_0];
+; CHECK-DAG: st.param.v2.f32     [func_retval0+0], {[[V_0_1]]}
+; CHECK-DAG: st.param.f32        [func_retval0+8], [[V_2]]
+; CHECK: ret;
+  ret <3 x float> %a
+}
+
+define <8 x i64> @test_v8i64(<8 x i64> %a) {
+; CHECK-LABEL: test_v8i64(
+; CHECK-DAG: ld.param.v2.u64     {[[V_6_7:(%rd[0-9]+[, ]*){2}]]}, [test_v8i64_param_0+48];
+; CHECK-DAG: ld.param.v2.u64     {[[V_4_5:(%rd[0-9]+[, ]*){2}]]}, [test_v8i64_param_0+32];
+; CHECK-DAG: ld.param.v2.u64     {[[V_2_3:(%rd[0-9]+[, ]*){2}]]}, [test_v8i64_param_0+16];
+; CHECK-DAG: ld.param.v2.u64     {[[V_0_1:(%rd[0-9]+[, ]*){2}]]}, [test_v8i64_param_0];
+; CHECK-DAG: st.param.v2.b64     [func_retval0+0],  {[[V_0_1]]}
+; CHECK-DAG: st.param.v2.b64     [func_retval0+16], {[[V_2_3]]}
+; CHECK-DAG: st.param.v2.b64     [func_retval0+32], {[[V_4_5]]}
+; CHECK-DAG: st.param.v2.b64     [func_retval0+48], {[[V_6_7]]}
+; CHECK: ret;
+  ret <8 x i64> %a
+}
+
+define <16 x i16> @test_v16i16(<16 x i16> %a) {
+; CHECK-LABEL: test_v16i16(
+; CHECK-DAG: ld.param.v4.u16     {[[V_12_15:(%rs[0-9]+[, ]*){4}]]}, [test_v16i16_param_0+24];
+; CHECK-DAG: ld.param.v4.u16     {[[V_8_11:(%rs[0-9]+[, ]*){4}]]}, [test_v16i16_param_0+16];
+; CHECK-DAG: ld.param.v4.u16     {[[V_4_7:(%rs[0-9]+[, ]*){4}]]}, [test_v16i16_param_0+8];
+; CHECK-DAG: ld.param.v4.u16     {[[V_0_3:(%rs[0-9]+[, ]*){4}]]}, [test_v16i16_param_0];
+; CHECK-DAG: st.param.v4.b16     [func_retval0+0], {[[V_0_3]]}
+; CHECK-DAG: st.param.v4.b16     [func_retval0+8], {[[V_4_7]]}
+; CHECK-DAG: st.param.v4.b16     [func_retval0+16], {[[V_8_11]]}
+; CHECK-DAG: st.param.v4.b16     [func_retval0+24], {[[V_12_15]]}
+; CHECK: ret;
+  ret <16 x i16> %a
+}

test/CodeGen/NVPTX/vec8.ll

 ; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
 
 target triple = "nvptx-unknown-cuda"
 
 ; CHECK: .visible .func foo
 define void @foo(<8 x i8> %a, i8* %b) {
-  %t0 = extractelement <8 x i8> %a, i32 0
-; CHECK-DAG: ld.param.v4.u8
-; CHECK-DAG: ld.param.u32
-  store i8 %t0, i8* %b
+; CHECK-DAG: ld.param.v4.u8 {[[E0:%rs[0-9]+]], [[E1:%rs[0-9]+]], [[E2:%rs[0-9]+]], [[E3:%rs[0-9]+]]}, [foo_param_0]
+; CHECK-DAG: ld.param.v4.u8 {[[E4:%rs[0-9]+]], [[E5:%rs[0-9]+]], [[E6:%rs[0-9]+]], [[E7:%rs[0-9]+]]}, [foo_param_0+4]
+; CHECK-DAG: ld.param.u32   %[[B:r[0-9+]]], [foo_param_1]
+; CHECK:     add.s16        [[T:%rs[0-9+]]], [[E1]], [[E6]];
+; CHECK:     st.u8          [%[[B]]], [[T]];
+  %t0 = extractelement <8 x i8> %a, i32 1
+  %t1 = extractelement <8 x i8> %a, i32 6
+  %t  = add i8 %t0, %t1
+  store i8 %t, i8* %b
   ret void
 }
 

test/CodeGen/NVPTX/vector-call.ll

 ; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
 
 target triple = "nvptx-unknown-cuda"
 
 declare void @bar(<4 x i32>)
 
-; CHECK-LABEL: @foo
+; CHECK-LABEL: .func foo(
+; CHECK-DAG: ld.param.v4.u32 {[[E0:%r[0-9]+]], [[E1:%r[0-9]+]], [[E2:%r[0-9]+]], [[E3:%r[0-9]+]]}, [foo_param_0];
+; CHECK: .param .align 16 .b8 param0[16];
+; CHECK-DAG: st.param.v4.b32  [param0+0],  {[[E0]], [[E1]], [[E2]], [[E3]]};
+; CHECK:     call.uni
+; CHECK:     ret;
 define void @foo(<4 x i32> %a) {
-; CHECK: st.param.v4.b32
   tail call void @bar(<4 x i32> %a)
   ret void
 }
+
+; CHECK-LABEL: .func foo3(
+; CHECK-DAG: ld.param.v2.u32 {[[E0:%r[0-9]+]], [[E1:%r[0-9]+]]}, [foo3_param_0];
+; CHECK-DAG: ld.param.u32 [[E2:%r[0-9]+]], [foo3_param_0+8];
+; CHECK: .param .align 16 .b8 param0[16];
+; CHECK-DAG: st.param.v2.b32  [param0+0],  {[[E0]], [[E1]]};
+; CHECK-DAG: st.param.b32     [param0+8],  [[E2]];
+; CHECK:     call.uni
+; CHECK:     ret;
+declare void @bar3(<3 x i32>)
+define void @foo3(<3 x i32> %a) {
+  tail call void @bar3(<3 x i32> %a)
+  ret void
+}