Vectorcall Calling Convention - Adding CodeGen Complete Support

[oren_ben_simhon]

Summary

The vectorcall calling convention specifies that arguments to functions are to be passed in registers, when possible.
vectorcall uses more registers for arguments than fastcall or the default x64 calling convention use.
The vectorcall calling convention is only supported in native code on x86 and x64 processors that include Streaming SIMD Extensions 2 (SSE2) and above.

The current implementation does not handle Homogeneous Vector Aggregates (HVAs) correctly and this review attempts to fix it.
The review also includes additional lit tests to cover better HVAs corner cases.

[majnemer]

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7741

Please use isa here.

8039

Ditto.

lib/Target/X86/X86CallingConv.cpp

95

Range-based for loop?

197–199

This comment seems garbled.

206–207

else after return is not in the LLVM style.

[oren_ben_simhon]

Implemented comments posted by David (Thank You)

[majnemer]

include/llvm/CodeGen/CallingConvLower.h

327

Comments should end with a period.

329

Variables start with an uppercase letter.

include/llvm/Target/TargetCallingConv.h

54–62

Can we align this code to match its neighbors?

lib/CodeGen/CallingConvLower.cpp

74

Range-based for loop?

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7740–7741

Formatting looks strange.

lib/Target/X86/X86CallingConv.cpp

95

Formatting looks strange.

104

I'd lose these parens.

132

I don't think these parens are adding anything.

145–181

I wonder if this can be refactored to reduce redundancy:

if (!ArgFlags.isHva() || ArgFlags.isHvaStart()) {
    // Assign shadow GPR register.
    (void)State.AllocateReg(CC_X86_64_VectorCallGetGPRs());

    // Assign XMM register.
    if (unsigned Reg = State.AllocateReg(CC_X86_VectorCallGetSSEs(ValVT))) {
      if (!ArgFlags.isHva())
        State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));

      // In Vectorcall Calling convention, additional shadow stack can be
      // created on top of the basic 32 bytes of win64.
      // It can happen if the fifth or sixth argument is vector type or HVA.
      // At that case for each argument a shadow stack of 8 bytes is allocated.
      if (Reg == X86::XMM4 || Reg == X86::XMM5)
        State.AllocateStack(8, 8);
    }

    return true;
}

162–163

Else after return doesn't conform to LLVM style.

189

Ditto.

201

I don't think these parens add anything.

206

Comments should end with a period.

208

Ditto.

214

Ditto.

lib/Target/X86/X86ISelLowering.cpp

3276–3277

I'd capitalize these variable names.

[rnk]

lib/Target/X86/X86CallingConv.cpp

65

ArrayRefs are implicitly constructable from C arrays. You should be able to just return RegListZMM here and throughout this file. If not, makeArrayRef takes C arrays and will do the right thing.

130

mojibake

lib/Target/X86/X86CallingConv.td

631

Ouch. I locally confirmed this is correct, but why design a new calling convention that doesn't handle the latest vector types... =P

test/CodeGen/X86/vectorcall.ll

75

This comment has been deleted.

75

ignore the comment above, I can't seem to delete it from phab. :(

75–208

This was testing that %r is passed indirectly in the first integer register parameter, but I guess that's incorrect because of the way __vectorcall pins arguments to registers based on their exact argument position. We should check for the load off the stack to capture the correct behavior. Something like this:

mov{{[l|q}} {{[0-9]+}}(%rsp), %[[r_reg:[^ ]*]]
movaps %[[r_reg]], %xmm0

[aaboud]

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7747

you already checked this case.

lib/Target/X86/X86ISelLowering.cpp

2807

You have a typo: "the the".
Also, can you explain why you are sorting the ArgLoc even for non-VectorCall calling convention?
Is that needed?

test/CodeGen/X86/vectorcall.ll

71

Do you really want to remove this test? Cannot you just fix it?

[oren_ben_simhon]

lib/Target/X86/X86CallingConv.cpp

65

You are right. Thank you.

95

I ran clang format before uploading the file. I added comments, hopefully it makes the formatting look better.

132

Consider a case in which ValVT is a vector type.
If i will remove these parens, i will enter the if statement block.
However i do not want to enter this statement in the case of vectors type .
That is why this parens are not redundant.

I will change a bit the condition look to make it clearer.

lib/Target/X86/X86CallingConv.td

631

If you refer to 512 bit vector types, Microsoft probably forgot to mention that it should be assigned to ZMM here.

lib/Target/X86/X86ISelLowering.cpp

2807

The next loop assumes that the locations are in the same order of the input arguments.
So the order should be kept for all calling conventions.

Currently AFAIK, vectorcall is the only one that changes the arguments order, still additional calling conventions might do the same.

Anyway, I will add a comment to clarify this.

test/CodeGen/X86/vectorcall.ll

71

AFAIK, This test checks something that cannot happen.
There is no case in which a function will return {double, double, double, double, double}.
If it wanted to return such a structure it will be returned by pointer as the first argument to the function:

define x86_vectorcallcc void @test_fp_4(%struct.five_doubles* sret %a)

[oren_ben_simhon]

Implemented comments submitted until 12/10 (Thank you, David/Reid/Amjad)

[majnemer]

lib/Target/X86/X86CallingConv.cpp

163–166

This is just return ArgFlags.isHva();

[oren_ben_simhon]

Implemented changes posted until 12/12 (Thank you David)

[rnk]

include/llvm/CodeGen/CallingConvLower.h

326

typo on "allocated"

lib/CodeGen/CallingConvLower.cpp

74

This creates a copy of a CCValAssign, which is unnecessary.

lib/Target/X86/X86CallingConv.cpp

130

I still see "type庸or example" with an Asian character in there.

lib/Target/X86/X86CallingConv.td

631

Actually, I think I just misunderstood. It looks like you allocate things to AVX registers elsewhere. I thought with this change you made it so that AVX registers would never be used with vectorcall.

lib/Target/X86/X86ISelLowering.cpp

2799

Unnecessary copy

2810

The stable_sort is only needed if the CC was vectorcall.

I think the code would actually be clearer if we used std::merge. After the first pass, ArgLocs should be all the non-HVA argument locations sorted by argument number. The second pass appends additional sorted HVA argument locations. Then you merge those two sorted lists. You'll need a temporary ArgLocs vector to make this work.

Please factor out this two pass vectorcall code into a template function that operates on a SmallVectorImpl<T>, where T can be InputArg or OutputArg.

test/CodeGen/X86/vectorcall.ll

71

This was really just a whitebox test to show that LLVM does not crash when you return 5 doubles. We should keep the test if we don't fatal error. Any behavior is fine. For example, it might trigger LLVM's logic to demote return by value to sret.

[oren_ben_simhon]

lib/Target/X86/X86ISelLowering.cpp

2810

I created a template function as you suggested.

The next loop assumes that the locations are sorted. Today Vectorcall CC change the order, tomorrow, some other functionality/CC could change that.
Since we already do the sort, we might as well do it for all CC and avoid future issues.

I agree that merge algorithm could be faster than stable_sort. But the overhead of using it is big.
Not only i will need a temporary ArgLocs, I will need a temporary CCState. IMHO, the current solution is preferred.

test/CodeGen/X86/vectorcall.ll

71

There is no fatal error. I am fine with leaving the test.

[oren_ben_simhon]

Implemented comments posted until 12/13 (Thank you Reid)

[rnk]

lib/Target/X86/X86ISelLowering.cpp

2810

I had envisioned that the sort or merge logic would live in the factored out second pass logic that is specific to vectorcall. All conventions other than vectorcall already preserve the invariant that argument locations are ordered by their IR position, which is why I see this sort/merge as being specific to vectorcall. I'd expect readers to be surprised that this sort is necessary, so I'd like to bind it closely with the code that makes the list unsorted.

If you want to document the invariant of the loop below, we can do that with an assertion.

Also, the merge isn't that bad. You should be able to do this:

unsigned NumFirstPassLocs = ArgLocs.size();
CCState.AnalyzeFormalArguments(...);
decltype(ArgLocs) TmpArgLocs;
std::swap(TmpArgLocs, ArgLocs);
auto B = TmpArgLocs.begin(), E = TmpArgLocs.end();
std::merge(B, B + NumFirstPassLocs, B + NumFirstPassLocs, E, ArgLocs.begin());

[oren_ben_simhon]

Implemented comments submitted until 12/14 (Thank you Reid)

[rnk]

Looks good. Please fix the remaining style issues before committing. Thanks for the patch!

include/llvm/CodeGen/CallingConvLower.h

533

Thanks! I think this looks a lot better factored here.

lib/Target/X86/X86ISelLowering.cpp

2755

This should have a more descriptive name in LLVM style. isSortedByValNo or something?

2804

formatting

3336

formatting

[majnemer]

include/llvm/CodeGen/CallingConvLower.h

532–533

Please clang-format this.
Also, I think Args should be an ArrayRef<T>.

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7751

Please remove the extra parens.

8038

Ditto.

lib/Target/X86/X86ISelLowering.cpp

2756–2759

This doesn't look correctly formatted.

[oren_ben_simhon]

include/llvm/CodeGen/CallingConvLower.h

532–533

Thank you Reid and David.
I reran clang format on the latest patch.

Other functions that receive Args (Like, AnalyzeCallOperands, CheckReturn, etc.) use the SmallVector representation of the data.
So, I prefer to be consistent with them and leave it as SmallVectorImpl.

[oren_ben_simhon]

Implemented comments submitted until 12/16.

Thank you Reid, David and Amjad for a very constructive and fruitful code review.
I will leave the review open for a couple of days (in case you have additional comments).

include/llvm/CodeGen/CallingConvLower.h

void AnalyzeCallOperands(SmallVectorImpl<MVT> &ArgVTs,
                            SmallVectorImpl<ISD::ArgFlagsTy> &Flags,
                            CCAssignFn Fn);
 
   /// AnalyzeCallResult - Analyze the return values of a call,
   /// incorporating info about the passed values into this state.
   void AnalyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins,
                          CCAssignFn Fn);
 
+  /// A shadow allocated register is a register that was allcoated
+  /// but wasn't added to the location list (Locs)
+  /// \returns true if the register was allocated as shadow or false otherwise
+  bool IsShadowAllocatedReg(unsigned reg) const;
+
   /// AnalyzeCallResult - Same as above except it's specialized for calls which
   /// produce a single value.
   void AnalyzeCallResult(MVT VT, CCAssignFn Fn);
 
   /// getFirstUnallocated - Return the index of the first unallocated register
   /// in the set, or Regs.size() if they are all allocated.
   unsigned getFirstUnallocated(ArrayRef<MCPhysReg> Regs) const {
     for (unsigned i = 0; i < Regs.size(); ++i)
static bool resultsCompatible(CallingConv::ID CalleeCC,
                                 CallingConv::ID CallerCC, MachineFunction &MF,
                                 LLVMContext &C,
                                 const SmallVectorImpl<ISD::InputArg> &Ins,
                                 CCAssignFn CalleeFn, CCAssignFn CallerFn);
 
 private:
   /// MarkAllocated - Mark a register and all of its aliases as allocated.
   void MarkAllocated(unsigned Reg);
 };
 
 
 
 } // end namespace llvm
 
 #endif

include/llvm/Target/TargetCallingConv.h

private:
     static const uint64_t InAlloca       = 1ULL<<12; ///< Passed with inalloca
     static const uint64_t InAllocaOffs   = 12;
     static const uint64_t SplitEnd       = 1ULL<<13; ///< Last part of a split
     static const uint64_t SplitEndOffs   = 13;
     static const uint64_t SwiftSelf      = 1ULL<<14; ///< Swift self parameter
     static const uint64_t SwiftSelfOffs  = 14;
     static const uint64_t SwiftError     = 1ULL<<15; ///< Swift error parameter
     static const uint64_t SwiftErrorOffs = 15;
+    static const uint64_t Hva = 1ULL << 16; ///< HVA field for
+                                            ///< vectorcall
+    static const uint64_t HvaOffs = 16;
+    static const uint64_t HvaStart = 1ULL << 17; ///< HVA structure start
+                                                 ///<  for vectorcall
+    static const uint64_t HvaStartOffs = 17;
+    static const uint64_t SecArgPass = 1ULL << 18; ///< Second argument
+                                                   ///<  pass for vectorcall
+    static const uint64_t SecArgPassOffs = 18;
+
     static const uint64_t OrigAlign      = 0x1FULL<<27;
     static const uint64_t OrigAlignOffs  = 27;
     static const uint64_t ByValSize      = 0x3fffffffULL<<32; ///< Struct size
     static const uint64_t ByValSizeOffs  = 32;
     static const uint64_t InConsecutiveRegsLast      = 0x1ULL<<62; ///< Struct size
     static const uint64_t InConsecutiveRegsLastOffs  = 62;
     static const uint64_t InConsecutiveRegs      = 0x1ULL<<63; ///< Struct size
     static const uint64_t InConsecutiveRegsOffs  = 63;
public:
     void setInAlloca() { Flags |= One << InAllocaOffs; }
 
     bool isSwiftSelf() const { return Flags & SwiftSelf; }
     void setSwiftSelf() { Flags |= One << SwiftSelfOffs; }
 
     bool isSwiftError() const { return Flags & SwiftError; }
     void setSwiftError() { Flags |= One << SwiftErrorOffs; }
 
+    bool isHva() const { return Flags & Hva; }
+    void setHva() { Flags |= One << HvaOffs; }
+
+    bool isHvaStart() const { return Flags & HvaStart; }
+    void setHvaStart() { Flags |= One << HvaStartOffs; }
+
+    bool isSecArgPass() const { return Flags & SecArgPass; }
+    void setSecArgPass() { Flags |= One << SecArgPassOffs; }
+
     bool isNest()      const { return Flags & Nest; }
     void setNest()     { Flags |= One << NestOffs; }
 
     bool isReturned()  const { return Flags & Returned; }
     void setReturned() { Flags |= One << ReturnedOffs; }
 
     bool isInConsecutiveRegs()  const { return Flags & InConsecutiveRegs; }
     void setInConsecutiveRegs() { Flags |= One << InConsecutiveRegsOffs; }

lib/CodeGen/CallingConvLower.cpp

 }
 
 /// Mark a register and all of its aliases as allocated.
 void CCState::MarkAllocated(unsigned Reg) {
   for (MCRegAliasIterator AI(Reg, &TRI, true); AI.isValid(); ++AI)
     UsedRegs[*AI/32] |= 1 << (*AI&31);
 }
 
+/// A shadow allocated register is a register that was allcoated
+/// but wasn't added to the location list (Locs)
+/// \return true if the register was allocated as shadow and false otherwise
+bool CCState::IsShadowAllocatedReg(unsigned reg) const {
+  if (!isAllocated(reg))
+    return false;
+
+  for (unsigned i = 0, e = Locs.size(); i != e; ++i) {
+    for (MCRegAliasIterator AI(Locs[i].getLocReg(), &TRI, true); AI.isValid();
+         ++AI) {
+      if (*AI == reg)
+        return false;
+    }
+  }
+  return true;
+}
+
 /// Analyze an array of argument values,
 /// incorporating info about the formals into this state.
 void
 CCState::AnalyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Ins,
                                 CCAssignFn Fn) {
   unsigned NumArgs = Ins.size();
 
   for (unsigned i = 0; i != NumArgs; ++i) {

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
                            Args[i].Node.getResNo() + Value);
       ISD::ArgFlagsTy Flags;
       unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
 
       if (Args[i].isZExt)
         Flags.setZExt();
       if (Args[i].isSExt)
         Flags.setSExt();
+      if (Args[i].isInReg) {
+        Flags.setInReg();
+        // If we are using vectorcall calling convention, a structure that is
+        // passed InReg - is surely an HVA
+        if (CLI.CallConv == CallingConv::X86_VectorCall &&
+            (dyn_cast<StructType>(FinalType))) {
+          Flags.setHva();
+          if (0 == Value)
+            Flags.setHvaStart();
+        }
+      }
       if (Args[i].isInReg)
         Flags.setInReg();
       if (Args[i].isSRet)
         Flags.setSRet();
       if (Args[i].isSwiftSelf)
         Flags.setSwiftSelf();
       if (Args[i].isSwiftError)
         Flags.setSwiftError();
       if (Args[i].isByVal)
         Flags.setByVal();
       if (Args[i].isInAlloca) {
         Flags.setInAlloca();
         // Set the byval flag for CCAssignFn callbacks that don't know about
for (unsigned Value = 0, NumValues = ValueVTs.size();
       Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
       ISD::ArgFlagsTy Flags;
       unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
 
       if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
         Flags.setZExt();
       if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
         Flags.setSExt();
-      if (F.getAttributes().hasAttribute(Idx, Attribute::InReg))
+      if (F.getAttributes().hasAttribute(Idx, Attribute::InReg)) {
         Flags.setInReg();
+        // If we are using vectorcall calling convention, a structure that is
+        // passed InReg - is surely an HVA
+        if (F.getCallingConv() == CallingConv::X86_VectorCall &&
+            (dyn_cast<StructType>(I->getType()))) {
+          Flags.setHva();
+          if (0 == Value)
+            Flags.setHvaStart();
+        }
+      }
       if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet))
         Flags.setSRet();
       if (F.getAttributes().hasAttribute(Idx, Attribute::SwiftSelf))
         Flags.setSwiftSelf();
       if (F.getAttributes().hasAttribute(Idx, Attribute::SwiftError))
         Flags.setSwiftError();
       if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
         Flags.setByVal();

lib/Target/X86/X86CallingConv.h

 #include "llvm/CodeGen/CallingConvLower.h"
 #include "llvm/IR/CallingConv.h"
 
 namespace llvm {
 
 /// When regcall calling convention compiled to 32 bit arch, special treatment
 /// is required for 64 bit masks.
 /// The value should be assigned to two GPRs.
-/// @return true if registers were allocated and false otherwise
+/// \return true if registers were allocated and false otherwise.
 bool CC_X86_32_RegCall_Assign2Regs(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
                                    CCValAssign::LocInfo &LocInfo,
                                    ISD::ArgFlagsTy &ArgFlags, CCState &State);
 
-inline bool CC_X86_32_VectorCallIndirect(unsigned &ValNo, MVT &ValVT,
-                                         MVT &LocVT,
-                                         CCValAssign::LocInfo &LocInfo,
-                                         ISD::ArgFlagsTy &ArgFlags,
-                                         CCState &State) {
-  // Similar to CCPassIndirect, with the addition of inreg.
-  LocVT = MVT::i32;
-  LocInfo = CCValAssign::Indirect;
-  ArgFlags.setInReg();
-  return false; // Continue the search, but now for i32.
-}
+/// Vectorcall calling convention has special handling for vector types or
+/// HVA for 64 bit arch.
+/// For HVAs shadow registers might be allocated on the first pass
+/// and actual XMM registers are allocated on the second pass.
+/// For vector types, actual XMM registers are allocated on the first pass.
+/// \return true if registers were allocated and false otherwise.
+bool CC_X86_64_VectorCall(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
+                          ISD::ArgFlagsTy &ArgFlags, CCState &State);
 
+/// Vectorcall calling convention has special handling for vector types or
+/// HVA for 32 bit arch.
+/// For HVAs actual XMM registers are allocated on the second pass.
+/// For vector types, actual XMM registers are allocated on the first pass.
+/// \return true if registers were allocated and false otherwise.
+bool CC_X86_32_VectorCall(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
+                          CCValAssign::LocInfo &LocInfo,
+                          ISD::ArgFlagsTy &ArgFlags, CCState &State);
 
 inline bool CC_X86_AnyReg_Error(unsigned &, MVT &, MVT &,
                                 CCValAssign::LocInfo &, ISD::ArgFlagsTy &,
                                 CCState &) {
   llvm_unreachable("The AnyReg calling convention is only supported by the " \
                    "stackmap and patchpoint intrinsics.");
   // gracefully fallback to X86 C calling convention on Release builds.
   return false;

lib/Target/X86/X86CallingConv.cpp

 //=== X86CallingConv.cpp - X86 Custom Calling Convention Impl   -*- C++ -*-===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file contains the implementation of custom routines for the X86
 // Calling Convention that aren't done by tablegen.
 //
 //===----------------------------------------------------------------------===//
 
 #include "MCTargetDesc/X86MCTargetDesc.h"
+#include "X86Subtarget.h"
 #include "llvm/CodeGen/CallingConvLower.h"
 #include "llvm/IR/CallingConv.h"
 
 namespace llvm {
 
 bool CC_X86_32_RegCall_Assign2Regs(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
                                    CCValAssign::LocInfo &LocInfo,
                                    ISD::ArgFlagsTy &ArgFlags, CCState &State) {
for (unsigned I = 0; I < RequiredGprsUponSplit; I++) {
     // Assign the value to the allocated register
     State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
   }
 
   // Successful in allocating regsiters - stop scanning next rules.
   return true;
 }
 
+static ArrayRef<MCPhysReg> CC_X86_VectorCallGetSSEs(const MVT &ValVT) {
+  if (ValVT.is512BitVector()) {
+    static const MCPhysReg RegListZMM[] = {X86::ZMM0, X86::ZMM1, X86::ZMM2,
+                                           X86::ZMM3, X86::ZMM4, X86::ZMM5};
+    return makeArrayRef(std::begin(RegListZMM), std::end(RegListZMM));
+  }
+
+  if (ValVT.is256BitVector()) {
+    static const MCPhysReg RegListYMM[] = {X86::YMM0, X86::YMM1, X86::YMM2,
+                                           X86::YMM3, X86::YMM4, X86::YMM5};
+    return makeArrayRef(std::begin(RegListYMM), std::end(RegListYMM));
+  }
+
+  static const MCPhysReg RegListXMM[] = {X86::XMM0, X86::XMM1, X86::XMM2,
+                                         X86::XMM3, X86::XMM4, X86::XMM5};
+  return makeArrayRef(std::begin(RegListXMM), std::end(RegListXMM));
+}
+
+static ArrayRef<MCPhysReg> CC_X86_64_VectorCallGetGPRs() {
+  static const MCPhysReg RegListGPR[] = {X86::RCX, X86::RDX, X86::R8, X86::R9};
+  return makeArrayRef(std::begin(RegListGPR), std::end(RegListGPR));
+}
+
+static bool CC_X86_VectorCallAssignRegister(unsigned &ValNo, MVT &ValVT,
+                                            MVT &LocVT,
+                                            CCValAssign::LocInfo &LocInfo,
+                                            ISD::ArgFlagsTy &ArgFlags,
+                                            CCState &State) {
+
+  ArrayRef<MCPhysReg> RegList = CC_X86_VectorCallGetSSEs(ValVT);
+  bool Is64bit = static_cast<const X86Subtarget &>(
+                     State.getMachineFunction().getSubtarget())
+                     .is64Bit();
+
+  for (unsigned I = 0; I < RegList.size(); ++I) {
+    if (!State.isAllocated(RegList[I])) {
+      unsigned AssigedReg = State.AllocateReg(RegList[I]);
+      assert(AssigedReg == RegList[I] &&
+             "Expecting a valid register allocation");
+      State.addLoc(
+          CCValAssign::getReg(ValNo, ValVT, AssigedReg, LocVT, LocInfo));
+      return true;
+    }
+    if (Is64bit && (State.IsShadowAllocatedReg(RegList[I]))) {
+      State.addLoc(
+          CCValAssign::getReg(ValNo, ValVT, RegList[I], LocVT, LocInfo));
+      return true;
+    }
+  }
+
+  assert(!Is64bit && "Clang should ensure that hva marked vectors will have "
+                     "an available register on 64 bit arch.");
+
+  return false;
+}
+
+bool CC_X86_64_VectorCall(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
+                          CCValAssign::LocInfo &LocInfo,
+                          ISD::ArgFlagsTy &ArgFlags, CCState &State) {
+  // On the second pass, go through the HVAs only
+  if (ArgFlags.isSecArgPass()) {
+    if (ArgFlags.isHva())
+      return CC_X86_VectorCallAssignRegister(ValNo, ValVT, LocVT, LocInfo,
+                                             ArgFlags, State);
+    return true;
+  }
+
+  // Process only vector types as defined by vectorcall spec:
+  // "A vector type is either a floating-point type庸or example,
+  //  a float or double熔r an SIMD vector type庸or example, __m128 or __m256."
+  if (!ValVT.isFloatingPoint() &&
+      !(ValVT.isVector() && ValVT.getSizeInBits() >= 128)) {
+
+    // If R9 was already assigned it means that we are after the fourth element
+    // and since that this is not an HVA / Vector type, we need to alloctae
+    // shadow XMM register
+    if (State.isAllocated(X86::R9)) {
+      // Assign shadow XMM register
+      (void)State.AllocateReg(CC_X86_VectorCallGetSSEs(ValVT));
+    }
+
+    return false;
+  }
+
+  if (ArgFlags.isHva()) {
+    if (ArgFlags.isHvaStart()) {
+      // Assign shadow GPR register
+      (void)State.AllocateReg(CC_X86_64_VectorCallGetGPRs());
+
+      // Assign shadow XMM register
+      unsigned Reg = State.AllocateReg(CC_X86_VectorCallGetSSEs(ValVT));
+
+      // In Vectorcall Calling convention, additional shadow stack can be
+      // created on top of the basic 32 bytes of win64.
+      // It can happen if the fifth or sixth argument is vector type or HVA.
+      // At that case for each argument a shadow stack of 8 bytes is allocated.
+      if (Reg == X86::XMM4 || Reg == X86::XMM5)
+        State.AllocateStack(8, 8);
+    }
+
+    // If this is an HVA - Stop the search
+    return true;
+  } else {
+
+    // Assign shadow GPR register
+    (void)State.AllocateReg(CC_X86_64_VectorCallGetGPRs());
+
+    // Assign XMM register
+    if (unsigned Reg = State.AllocateReg(CC_X86_VectorCallGetSSEs(ValVT))) {
+      State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
+
+      // In Vectorcall Calling convention, additional shadow stack can be
+      // created on top of the basic 32 bytes of win64.
+      // It can happen if the fifth or sixth argument is vector type or HVA.
+      // At that case for each argument a shadow stack of 8 bytes is allocated.
+      if (Reg == X86::XMM4 || Reg == X86::XMM5)
+        State.AllocateStack(8, 8);
+
+      return true; // Allocated a register - Stop the search
+    }
+  }
+
+  return false; // Continue the search
+}
+
+bool CC_X86_32_VectorCall(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
+                          CCValAssign::LocInfo &LocInfo,
+                          ISD::ArgFlagsTy &ArgFlags, CCState &State) {
+  // On the second pass, go through the HVAs only
+  if (ArgFlags.isSecArgPass()) {
+    if (ArgFlags.isHva())
+      return CC_X86_VectorCallAssignRegister(ValNo, ValVT, LocVT, LocInfo,
+                                             ArgFlags, State);
+    return true;
+  }
+
+  // Process only vector types as defined by vectorcall spec:
+  // "A vector type is either a floating-point type庸or example,
+  //  a float or double熔r an SIMD vector type庸or example, __m128 or __m256."
+  if (!ValVT.isFloatingPoint() &&
+      !(ValVT.isVector() && ValVT.getSizeInBits() >= 128)) {
+    return false;
+  }
+
+  if (ArgFlags.isHva()) {
+    return true; // If this is an HVA - Stop the search
+  } else {
+    // Assign XMM register
+    if (unsigned Reg = State.AllocateReg(CC_X86_VectorCallGetSSEs(ValVT))) {
+      State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
+      return true;
+    }
+  }
+
+  return false; // No register was assigned - Continue the search
+}
+
 } // End llvm namespace

lib/Target/X86/X86CallingConv.td

 def RetCC_X86_32_HiPE : CallingConv<[
   // Promote all types to i32
   CCIfType<[i8, i16], CCPromoteToType<i32>>,
 
   // Return: HP, P, VAL1, VAL2
   CCIfType<[i32], CCAssignToReg<[ESI, EBP, EAX, EDX]>>
 ]>;
 
-// X86-32 HiPE return-value convention.
+// X86-32 Vectorcall return-value convention.
 def RetCC_X86_32_VectorCall : CallingConv<[
-  // Vector types are returned in XMM0,XMM1,XMMM2 and XMM3.
-  CCIfType<[f32, f64, v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
+  // Floating Point types are returned in XMM0,XMM1,XMMM2 and XMM3.
+  CCIfType<[f32, f64, f128],
             CCAssignToReg<[XMM0,XMM1,XMM2,XMM3]>>,
 
-  // 256-bit FP vectors
-  CCIfType<[v32i8, v16i16, v8i32, v4i64, v8f32, v4f64],
-            CCAssignToReg<[YMM0,YMM1,YMM2,YMM3]>>,
-
-  // 512-bit FP vectors
-  CCIfType<[v64i8, v32i16, v16i32, v8i64, v16f32, v8f64],
-            CCAssignToReg<[ZMM0,ZMM1,ZMM2,ZMM3]>>,
-
   // Return integers in the standard way.
   CCDelegateTo<RetCC_X86Common>
 ]>;
 
 // X86-64 C return-value convention.
 def RetCC_X86_64_C : CallingConv<[
   // The X86-64 calling convention always returns FP values in XMM0.
   CCIfType<[f32], CCAssignToReg<[XMM0, XMM1]>>,

 def RetCC_X86_Win64_C : CallingConv<[
   // The X86-Win64 calling convention always returns __m64 values in RAX.
   CCIfType<[x86mmx], CCBitConvertToType<i64>>,
 
   // Otherwise, everything is the same as 'normal' X86-64 C CC.
   CCDelegateTo<RetCC_X86_64_C>
 ]>;
 
+// X86-64 vectorcall return-value convention.
+def RetCC_X86_64_Vectorcall : CallingConv<[
+  // Vectorcall calling convention always returns FP values in XMMs.
+  CCIfType<[f32, f64, f128], 
+    CCAssignToReg<[XMM0, XMM1, XMM2, XMM3]>>,
+
+  // Otherwise, everything is the same as Windows X86-64 C CC.
+  CCDelegateTo<RetCC_X86_Win64_C>
+]>;
+
 // X86-64 HiPE return-value convention.
 def RetCC_X86_64_HiPE : CallingConv<[
   // Promote all types to i64
   CCIfType<[i8, i16, i32], CCPromoteToType<i64>>,
 
   // Return: HP, P, VAL1, VAL2
   CCIfType<[i64], CCAssignToReg<[R15, RBP, RAX, RDX]>>
 ]>;
def RetCC_X86_64 : CallingConv<[
 
   // Handle Swift calls.
   CCIfCC<"CallingConv::Swift", CCDelegateTo<RetCC_X86_64_Swift>>,
 
   // Handle explicit CC selection
   CCIfCC<"CallingConv::X86_64_Win64", CCDelegateTo<RetCC_X86_Win64_C>>,
   CCIfCC<"CallingConv::X86_64_SysV", CCDelegateTo<RetCC_X86_64_C>>,
 
+  // Handle Vectorcall CC
+  CCIfCC<"CallingConv::X86_VectorCall", CCDelegateTo<RetCC_X86_64_Vectorcall>>,
+
   // Handle HHVM calls.
   CCIfCC<"CallingConv::HHVM", CCDelegateTo<RetCC_X86_64_HHVM>>,
 
   CCIfCC<"CallingConv::X86_RegCall",
           CCIfSubtarget<"isTargetWin64()",
                         CCDelegateTo<RetCC_X86_Win64_RegCall>>>,
   CCIfCC<"CallingConv::X86_RegCall", CCDelegateTo<RetCC_X86_SysV64_RegCall>>,
           
def CC_X86_Win64_C : CallingConv<[
   CCIfType<[i32, i64, f32, f64], CCAssignToStack<8, 8>>,
 
   // Long doubles get stack slots whose size and alignment depends on the
   // subtarget.
   CCIfType<[f80], CCAssignToStack<0, 0>>
 ]>;
 
 def CC_X86_Win64_VectorCall : CallingConv<[
-  // The first 6 floating point and vector types of 128 bits or less use
-  // XMM0-XMM5.
-  CCIfType<[f32, f64, v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
-           CCAssignToReg<[XMM0, XMM1, XMM2, XMM3, XMM4, XMM5]>>,
-
-  // 256-bit vectors use YMM registers.
-  CCIfType<[v32i8, v16i16, v8i32, v4i64, v8f32, v4f64],
-           CCAssignToReg<[YMM0, YMM1, YMM2, YMM3, YMM4, YMM5]>>,
-
-  // 512-bit vectors use ZMM registers.
-  CCIfType<[v64i8, v32i16, v16i32, v8i64, v16f32, v8f64],
-           CCAssignToReg<[ZMM0, ZMM1, ZMM2, ZMM3, ZMM4, ZMM5]>>,
+  CCCustom<"CC_X86_64_VectorCall">,
 
   // Delegate to fastcall to handle integer types.
   CCDelegateTo<CC_X86_Win64_C>
 ]>;
 
 
 def CC_X86_64_GHC : CallingConv<[
   // Promote i8/i16/i32 arguments to i64.
def CC_X86_32_FastCall : CallingConv<[
 
   // The first 2 integer arguments are passed in ECX/EDX
   CCIfInReg<CCIfType<[i32], CCAssignToReg<[ECX, EDX]>>>,
 
   // Otherwise, same as everything else.
   CCDelegateTo<CC_X86_32_Common>
 ]>;
 
-def CC_X86_32_VectorCall : CallingConv<[
-  // The first 6 floating point and vector types of 128 bits or less use
-  // XMM0-XMM5.
-  CCIfType<[f32, f64, v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
-           CCAssignToReg<[XMM0, XMM1, XMM2, XMM3, XMM4, XMM5]>>,
-
-  // 256-bit vectors use YMM registers.
-  CCIfType<[v32i8, v16i16, v8i32, v4i64, v8f32, v4f64],
-           CCAssignToReg<[YMM0, YMM1, YMM2, YMM3, YMM4, YMM5]>>,
-
-  // 512-bit vectors use ZMM registers.
-  CCIfType<[v64i8, v32i16, v16i32, v8i64, v16f32, v8f64],
-           CCAssignToReg<[ZMM0, ZMM1, ZMM2, ZMM3, ZMM4, ZMM5]>>,
-
-  // Otherwise, pass it indirectly.
-  CCIfType<[v16i8, v8i16, v4i32, v2i64, v4f32, v2f64,
-            v32i8, v16i16, v8i32, v4i64, v8f32, v4f64,
-            v64i8, v32i16, v16i32, v8i64, v16f32, v8f64],
-           CCCustom<"CC_X86_32_VectorCallIndirect">>,
+def CC_X86_Win32_VectorCall : CallingConv<[
+  // Pass floating point in XMMs
+  CCCustom<"CC_X86_32_VectorCall">,
 
   // Delegate to fastcall to handle integer types.
   CCDelegateTo<CC_X86_32_FastCall>
 ]>;
 
 def CC_X86_32_ThisCall_Common : CallingConv<[
   // The first integer argument is passed in ECX
   CCIfType<[i32], CCAssignToReg<[ECX]>>,

 
 // This is the root argument convention for the X86-32 backend.
 def CC_X86_32 : CallingConv<[
   // X86_INTR calling convention is valid in MCU target and should override the
   // MCU calling convention. Thus, this should be checked before isTargetMCU().
   CCIfCC<"CallingConv::X86_INTR", CCDelegateTo<CC_X86_32_Intr>>,
   CCIfSubtarget<"isTargetMCU()", CCDelegateTo<CC_X86_32_MCU>>,
   CCIfCC<"CallingConv::X86_FastCall", CCDelegateTo<CC_X86_32_FastCall>>,
-  CCIfCC<"CallingConv::X86_VectorCall", CCDelegateTo<CC_X86_32_VectorCall>>,
+  CCIfCC<"CallingConv::X86_VectorCall", CCDelegateTo<CC_X86_Win32_VectorCall>>,
   CCIfCC<"CallingConv::X86_ThisCall", CCDelegateTo<CC_X86_32_ThisCall>>,
   CCIfCC<"CallingConv::Fast", CCDelegateTo<CC_X86_32_FastCC>>,
   CCIfCC<"CallingConv::GHC", CCDelegateTo<CC_X86_32_GHC>>,
   CCIfCC<"CallingConv::HiPE", CCDelegateTo<CC_X86_32_HiPE>>,
   CCIfCC<"CallingConv::X86_RegCall", CCDelegateTo<CC_X86_32_RegCall>>,
 
   // Otherwise, drop to normal X86-32 CC
   CCDelegateTo<CC_X86_32_C>

lib/Target/X86/X86ISelLowering.cpp

 //
 //===----------------------------------------------------------------------===//
 
 #include "X86ISelLowering.h"
 #include "Utils/X86ShuffleDecode.h"
 #include "X86CallingConv.h"
 #include "X86FrameLowering.h"
 #include "X86InstrBuilder.h"
+#include "X86IntrinsicsInfo.h"
 #include "X86MachineFunctionInfo.h"
 #include "X86ShuffleDecodeConstantPool.h"
 #include "X86TargetMachine.h"
 #include "X86TargetObjectFile.h"
 #include "llvm/ADT/SmallBitVector.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/ADT/StringExtras.h"

 #include "llvm/MC/MCContext.h"
 #include "llvm/MC/MCExpr.h"
 #include "llvm/MC/MCSymbol.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/MathExtras.h"
 #include "llvm/Target/TargetOptions.h"
-#include "X86IntrinsicsInfo.h"
+#include <algorithm>
 #include <bitset>
-#include <numeric>
 #include <cctype>
+#include <numeric>
 using namespace llvm;
 
 #define DEBUG_TYPE "x86-isel"
 
 STATISTIC(NumTailCalls, "Number of tail calls");
 
 static cl::opt<bool> ExperimentalVectorWideningLegalization(
     "x86-experimental-vector-widening-legalization", cl::init(false),
static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
 
   static const MCPhysReg XMMArgRegs64Bit[] = {
     X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
     X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
   };
   return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
 }
 
 SDValue X86TargetLowering::LowerFormalArguments(
     SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
   MachineFunction &MF = DAG.getMachineFunction();
   X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
   const TargetFrameLowering &TFI = *Subtarget.getFrameLowering();
 
   const Function *Fn = MF.getFunction();
   if (Fn->hasExternalLinkage() &&
       Subtarget.isTargetCygMing() &&
       Fn->getName() == "main")
     FuncInfo->setForceFramePointer(true);
SDValue X86TargetLowering::LowerFormalArguments(
   CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
 
   // Allocate shadow area for Win64
   if (IsWin64)
     CCInfo.AllocateStack(32, 8);
 
   CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
 
+  // In vectorcall calling convention a second pass is required for the HVA
+  // types
+  if (CallingConv::X86_VectorCall == CallConv) {
+    SmallVector<ISD::InputArg, 16> InsSec;
+    for (auto InArg : Ins) {
+      ISD::InputArg InArgSec(InArg);
+      InArgSec.Flags.setSecArgPass();
+      InsSec.push_back(InArgSec);
+    }
+    CCInfo.AnalyzeFormalArguments(InsSec, CC_X86);
+  }
+
+  // Sort the the locations of the arguments according to their original
+  // position
+  std::stable_sort(ArgLocs.begin(), ArgLocs.end(),
+                   [](const CCValAssign &a, const CCValAssign &b) -> bool {
+                     return a.getValNo() < b.getValNo();
+                   });
+
   SDValue ArgValue;
   for (unsigned I = 0, InsIndex = 0, E = ArgLocs.size(); I != E;
        ++I, ++InsIndex) {
     assert(InsIndex < Ins.size() && "Invalid Ins index");
     CCValAssign &VA = ArgLocs[I];
 
     if (VA.isRegLoc()) {
       EVT RegVT = VA.getLocVT();
X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
   CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
 
   // Allocate shadow area for Win64
   if (IsWin64)
     CCInfo.AllocateStack(32, 8);
 
   CCInfo.AnalyzeCallOperands(Outs, CC_X86);
 
+  // In vectorcall calling convention a second pass is required for the HVA
+  // types
+  if (CallingConv::X86_VectorCall == CallConv) {
+    SmallVector<ISD::OutputArg, 4> OutsSec;
+    for (auto OutArg : Outs) {
+      ISD::OutputArg OutArgSec(OutArg);
+      OutArgSec.Flags.setSecArgPass();
+      OutsSec.push_back(OutArgSec);
+    }
+    CCInfo.AnalyzeCallOperands(OutsSec, CC_X86);
+  }
+
+  // Sort the the locations of the arguments according to their original
+  // position
+  std::stable_sort(ArgLocs.begin(), ArgLocs.end(),
+                   [](const CCValAssign &a, const CCValAssign &b) -> bool {
+                     return a.getValNo() < b.getValNo();
+                   });
+
   // Get a count of how many bytes are to be pushed on the stack.
   unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
   if (IsSibcall)
     // This is a sibcall. The memory operands are available in caller's
     // own caller's stack.
     NumBytes = 0;
   else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
            canGuaranteeTCO(CallConv))
X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
 
   SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
   SmallVector<SDValue, 8> MemOpChains;
   SDValue StackPtr;
 
   // Walk the register/memloc assignments, inserting copies/loads.  In the case
   // of tail call optimization arguments are handle later.
   const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo();
   for (unsigned I = 0, OutIndex = 0, E = ArgLocs.size(); I != E;
        ++I, ++OutIndex) {
     assert(OutIndex < Outs.size() && "Invalid Out index");
     // Skip inalloca arguments, they have already been written.
     ISD::ArgFlagsTy Flags = Outs[OutIndex].Flags;
     if (Flags.isInAlloca())
       continue;
 
     CCValAssign &VA = ArgLocs[I];

test/CodeGen/X86/vectorcall.ll

 ; RUN: llc -mtriple=i686-pc-win32 -mattr=+sse2 < %s | FileCheck %s --check-prefix=CHECK --check-prefix=X86
 ; RUN: llc -mtriple=x86_64-pc-win32 < %s | FileCheck %s --check-prefix=CHECK --check-prefix=X64
 
 ; Test integer arguments.
 
 define x86_vectorcallcc i32 @test_int_1() {
   ret i32 0
 }
-
 ; CHECK-LABEL: {{^}}test_int_1@@0:
 ; CHECK: xorl %eax, %eax
 
 define x86_vectorcallcc i32 @test_int_2(i32 inreg %a) {
   ret i32 %a
 }
-
 ; X86-LABEL: {{^}}test_int_2@@4:
 ; X64-LABEL: {{^}}test_int_2@@8:
 ; CHECK: movl %ecx, %eax
 
 define x86_vectorcallcc i32 @test_int_3(i64 inreg %a) {
   %at = trunc i64 %a to i32
   ret i32 %at
 }
-
 ; X86-LABEL: {{^}}test_int_3@@8:
 ; X64-LABEL: {{^}}test_int_3@@8:
 ; CHECK: movl %ecx, %eax
 
 define x86_vectorcallcc i32 @test_int_4(i32 inreg %a, i32 inreg %b) {
   %s = add i32 %a, %b
   ret i32 %s
 }
-
 ; X86-LABEL: {{^}}test_int_4@@8:
 ; X86: leal (%ecx,%edx), %eax
-
 ; X64-LABEL: {{^}}test_int_4@@16:
 ; X64: leal (%rcx,%rdx), %eax
 
 define x86_vectorcallcc i32 @"\01test_int_5"(i32, i32) {
   ret i32 0
 }
 ; CHECK-LABEL: {{^}}test_int_5:
 
ret {double, double, double, double}
         { double 0.0, double 0.0, double 0.0, double 0.0 }
 }
 ; CHECK-LABEL: {{^}}test_fp_3@@0:
 ; CHECK: xorps %xmm0
 ; CHECK: xorps %xmm1
 ; CHECK: xorps %xmm2
 ; CHECK: xorps %xmm3
 
-; FIXME: Returning via x87 isn't compatible, but its hard to structure the
-; tablegen any other way.
-define x86_vectorcallcc {double, double, double, double, double} @test_fp_4() {
-  ret {double, double, double, double, double}
-        { double 0.0, double 0.0, double 0.0, double 0.0, double 0.0 }
-}
-; CHECK-LABEL: {{^}}test_fp_4@@0:
-; CHECK: fldz
-; CHECK: xorps %xmm0
-; CHECK: xorps %xmm1
-; CHECK: xorps %xmm2
-; CHECK: xorps %xmm3
-
 define x86_vectorcallcc <16 x i8> @test_vec_1(<16 x i8> %a, <16 x i8> %b) {
   ret <16 x i8> %b
 }
 ; CHECK-LABEL: {{^}}test_vec_1@@32:
 ; CHECK: movaps %xmm1, %xmm0
 
 define x86_vectorcallcc <16 x i8> @test_vec_2(
     double, <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8> %r) {
   ret <16 x i8> %r
 }
 ; CHECK-LABEL: {{^}}test_vec_2@@104:
-; CHECK: movaps (%{{[re]}}cx), %xmm0
+; CHECK: movaps {{[0-9]*}}(%{{[re](bp|ax|cx)}}), %xmm0
+
+
+%struct.HVA5 = type { <4 x float>, <4 x float>, <4 x float>, <4 x float>, <4 x float> }
+%struct.HVA4 = type { <4 x float>, <4 x float>, <4 x float>, <4 x float> }
+%struct.HVA3 = type { <4 x float>, <4 x float>, <4 x float> }
+%struct.HVA2 = type { <4 x float>, <4 x float> }
+
+define x86_vectorcallcc <4 x float> @test_mixed_1(i32 %a, %struct.HVA4 inreg %bb, i32 %c) {
+entry:
+  %b = alloca %struct.HVA4, align 16
+  store %struct.HVA4 %bb, %struct.HVA4* %b, align 16
+  %w1 = getelementptr inbounds %struct.HVA4, %struct.HVA4* %b, i32 0, i32 1
+  %0 = load <4 x float>, <4 x float>* %w1, align 16
+  ret <4 x float> %0
+}
+; CHECK-LABEL: test_mixed_1
+; CHECK:       movaps %xmm1, %xmm0
+; CHECK:       ret{{q|l}}
+
+define x86_vectorcallcc <4 x float> @test_mixed_2(%struct.HVA4 inreg %a, %struct.HVA4* %b, <4 x float> %c) {
+entry:
+  %c.addr = alloca <4 x float>, align 16
+  store <4 x float> %c, <4 x float>* %c.addr, align 16
+  %0 = load <4 x float>, <4 x float>* %c.addr, align 16
+  ret <4 x float> %0
+}
+; CHECK-LABEL: test_mixed_2
+; X86:         movaps  %xmm0, (%esp)
+; X64:         movaps  %xmm2, %xmm0
+; CHECK:       ret{{[ql]}}
+
+define x86_vectorcallcc <4 x float> @test_mixed_3(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> %e, %struct.HVA2* %f) {
+entry:
+  %x = getelementptr inbounds %struct.HVA2, %struct.HVA2* %f, i32 0, i32 0
+  %0 = load <4 x float>, <4 x float>* %x, align 16
+  ret <4 x float> %0
+}
+; CHECK-LABEL: test_mixed_3
+; CHECK:       movaps (%{{[re][ac]}}x), %xmm0
+; CHECK:       ret{{[ql]}}
+
+define x86_vectorcallcc <4 x float> @test_mixed_4(%struct.HVA4 inreg %a, %struct.HVA2* %bb, <4 x float> %c) {
+entry:
+  %y4 = getelementptr inbounds %struct.HVA2, %struct.HVA2* %bb, i32 0, i32 1
+  %0 = load <4 x float>, <4 x float>* %y4, align 16
+  ret <4 x float> %0
+}
+; CHECK-LABEL: test_mixed_4
+; X86:         movaps 16(%eax), %xmm0
+; X64:         movaps 16(%rdx), %xmm0
+; CHECK:       ret{{[ql]}}
+
+define x86_vectorcallcc <4 x float> @test_mixed_5(%struct.HVA3 inreg %a, %struct.HVA3* %b, <4 x float> %c, %struct.HVA2 inreg %dd) {
+entry:
+  %d = alloca %struct.HVA2, align 16
+  store %struct.HVA2 %dd, %struct.HVA2* %d, align 16
+  %y5 = getelementptr inbounds %struct.HVA2, %struct.HVA2* %d, i32 0, i32 1
+  %0 = load <4 x float>, <4 x float>* %y5, align 16
+  ret <4 x float> %0
+}
+; CHECK-LABEL: test_mixed_5
+; CHECK:       movaps %xmm5, %xmm0
+; CHECK:       ret{{[ql]}}
+
+define x86_vectorcallcc %struct.HVA4 @test_mixed_6(%struct.HVA4 inreg %a, %struct.HVA4* %b) {
+entry:
+  %retval = alloca %struct.HVA4, align 16
+  %0 = bitcast %struct.HVA4* %retval to i8*
+  %1 = bitcast %struct.HVA4* %b to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %0, i8* %1, i32 64, i32 16, i1 false)
+  %2 = load %struct.HVA4, %struct.HVA4* %retval, align 16
+  ret %struct.HVA4 %2
+}
+; CHECK-LABEL: test_mixed_6
+; CHECK:       movaps (%{{[re]}}sp), %xmm0
+; CHECK:       movaps 16(%{{[re]}}sp), %xmm1
+; CHECK:       movaps 32(%{{[re]}}sp), %xmm2
+; CHECK:       movaps 48(%{{[re]}}sp), %xmm3
+; CHECK:       ret{{[ql]}}
+
+declare void @llvm.memset.p0i8.i64(i8* nocapture writeonly, i8, i64, i32, i1)
+declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture writeonly, i8* nocapture readonly, i64, i32, i1)
+declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture writeonly, i8* nocapture readonly, i32, i32, i1)
+
+define x86_vectorcallcc void @test_mixed_7(%struct.HVA5* noalias sret %agg.result) {
+entry:
+  %a = alloca %struct.HVA5, align 16
+  %0 = bitcast %struct.HVA5* %a to i8*
+  call void @llvm.memset.p0i8.i64(i8* %0, i8 0, i64 80, i32 16, i1 false)
+  %1 = bitcast %struct.HVA5* %agg.result to i8*
+  %2 = bitcast %struct.HVA5* %a to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %1, i8* %2, i64 80, i32 16, i1 false)
+  ret void
+}
+; CHECK-LABEL: test_mixed_7
+; CHECK:       movaps %xmm0, 64(%{{rcx|eax}})
+; CHECK:       movaps %xmm0, 48(%{{rcx|eax}})
+; CHECK:       movaps %xmm0, 32(%{{rcx|eax}})
+; CHECK:       movaps %xmm0, 16(%{{rcx|eax}})
+; CHECK:       movaps %xmm0, (%{{rcx|eax}})
+; X64:         mov{{[ql]}}  %rcx, %rax
+; CHECK:       ret{{[ql]}}
+
+define x86_vectorcallcc <4 x float> @test_mixed_8(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, i32 %e, <4 x float> %f) {
+entry:
+  %f.addr = alloca <4 x float>, align 16
+  store <4 x float> %f, <4 x float>* %f.addr, align 16
+  %0 = load <4 x float>, <4 x float>* %f.addr, align 16
+  ret <4 x float> %0
+}
+; CHECK-LABEL: test_mixed_8
+; X86:         movaps %xmm4, %xmm0
+; X64:         movaps %xmm5, %xmm0
+; CHECK:       ret{{[ql]}}
+
+%struct.HFA4 = type { double, double, double, double }
+declare x86_vectorcallcc double @test_mixed_9_callee(%struct.HFA4 %x, double %y)
+
+define x86_vectorcallcc double @test_mixed_9_caller(%struct.HFA4 inreg %b) {
+entry:
+  %call = call x86_vectorcallcc double @test_mixed_9_callee(%struct.HFA4 inreg %b, double 3.000000e+00)
+  %add = fadd double 1.000000e+00, %call
+  ret double %add
+}
+; CHECK-LABEL: test_mixed_9_caller
+; CHECK:       movaps  %xmm3, %xmm4
+; CHECK:       movaps  %xmm2, %xmm3
+; CHECK:       movaps  %xmm1, %xmm2
+; X32:         movasd  %xmm0, %xmm1
+; X64:         movapd  %xmm5, %xmm1
+; CHECK:       call{{l|q}}   test_mixed_9_callee@@40
+; CHECK:       addsd   {{.*}}, %xmm0
+; CHECK:       ret{{l|q}}