[X86][regcall] Support passing / returning structures

[pengfei]

Summary

Currently, the regcall calling conversion in Clang doesn't match with
ICC when passing / returning structures. https://godbolt.org/z/axxKMKrW7

This patch tries to fix the problem to match with ICC.

[LuoYuanke]

clang/include/clang/CodeGen/CGFunctionInfo.h

744

Use "Log2_32()"?

[pengfei]

Address Yuanke's comment.

[craig.topper]

clang/include/clang/CodeGen/CGFunctionInfo.h

744

Are you assuming Width is a power of 2? Should we assert that?

clang/lib/CodeGen/CGCall.cpp

4673

static?

clang/lib/CodeGen/TargetInfo.cpp

3740

clang-format

[pengfei]

Address review comments. Thanks Craig!

[pengfei]

clang/include/clang/CodeGen/CGFunctionInfo.h

744

Good question! I assumed it, but I found it's not true for Clang, although ICC and GCC error for it.
Maybe we should diagnose it too? Anyway, I added an assert for it.

[pengfei]

clang-formatted.

[pengfei]

Ping? We have internal request for this.

[LuoYuanke]

clang/include/clang/CodeGen/CGFunctionInfo.h

590

I notice some code would indicate it is log 2 size with Log2 suffix in the variable name. Do you think it is more readable to add Log2 suffix?

clang/lib/CodeGen/CGCall.cpp

5238

Does this also affect other calling convention besides fastcall?

clang/lib/CodeGen/TargetInfo.cpp

2303

Update the comments for the new parameter?

3031

Would you add a test for non regcall? Pass 1024 bit vector parameter and check if it is well handled both with regcall and without regcall.
Would you add comments to depict why regcall accept the size which is more than 512?

clang/test/CodeGen/aarch64-neon-tbl.c

45

I'm curious why aarch64 test cases are affected by the patch.

clang/test/CodeGen/regcall2.c

3

Add test case for target that has no avx512 feature?

10

May add a test case to show what's the max register we can pass with regcall.

[pengfei]

Address Yuanke's comments.

clang/include/clang/CodeGen/CGFunctionInfo.h

590

I think it doesn't matter. We have getter and setter for the variable. People won't access it directly.

clang/lib/CodeGen/CGCall.cpp

5238

I don't think so. The change here adds for the missing cases like [2 x <4 x double>] or { <2 x double>, <4 x double> } which should also set min-legal-width-width to the maximum of the vector length.
There're several reasons why other calling convention won't be affected.

1. If a target has ability to pass arguments like [2 x <4 x double>], it must have the ability for <4 x double> and have set min-legal-width-width to 256 when passing it. So it makes more sense to set min-legal-width-width to 256 for [2 x <4 x double>] rather than keeping it as 0;
2. AFAIK, targets other than X86 simply ignore min-legal-width-width. So the change won't affect them;
3. On x86, calling conventions other than regcall don't allow arguments size larger than 512, see if (!IsRegCall && Size > 512). They will be turned into pointers, so they won't be affected by this change;
4. For arguments size no larger than 512 and only contain single vector element, we have already turned them into pure vectors. So they have already set min-legal-width-width to the correct value;
5. For arguments have more then one vector elements. Clang has bug which doesn't match with GCC and ICC. I have filed a bug here https://github.com/llvm/llvm-project/issues/54582
6. Thus, only regcall can generate arguments type like [2 x <4 x double>] on X86. So only it will be affected by this.

clang/lib/CodeGen/TargetInfo.cpp

3031

Added one to non regcall. regcall doesn't specify how to handle 1024 bit vector. I'd take it as UB, so we don't need such a test.
https://www.intel.com/content/www/us/en/develop/documentation/cpp-compiler-developer-guide-and-reference/top/compiler-reference/c-c-calling-conventions.html

clang/test/CodeGen/aarch64-neon-tbl.c

45

As I explained above, [2 x <16 x i8>] should have the same value of min-legal-vector-width as <16 x i8>. The difference between #0 and #1 is the value of min-legal-vector-width.

clang/test/CodeGen/regcall2.c

3

I'd take it as UB for Clang and GCC using 512 bit vector without avx512 feature. ICC always promotes to avx512 when it finds 512 bit vector. So we don't need such tests.

10

We have tests for it in clang/test/CodeGen/regcall.c. This patch doesn't affect the capability of regcall.

[LuoYuanke]

LGTM, thanks.

[pengfei]

clang/lib/CodeGen/CGCall.cpp

5238

Update for bullet 5. The handling for multi vector elements is correct. It's also passed / returned by memory. So it's still not affected by this change.

[erichkeane]

clang/lib/CodeGen/TargetInfo.cpp

3981

This line here accesses NeededSSE while uninitialized. Do you need to initalize it to something on 3963?

[erichkeane]

clang/lib/CodeGen/TargetInfo.cpp

3981

Fixed: https://github.com/llvm/llvm-project/commit/2267549296dabfed31ce194bb124f7bdb91f74c5

Please confirm the fix.

[aaron.ballman]

clang/lib/CodeGen/TargetInfo.cpp

3981

I've fixed this in 2267549296dabfed31ce194bb124f7bdb91f74c5 as it broke many tests on Windows (but I don't think one of the community build bots noticed this, which suggests we may want to have a Debug Windows builder).

[pengfei]

Thanks @erichkeane @aaron.ballman ! Yeah, I didn't receive buildbots notice about that.

clang/include/clang/CodeGen/CGFunctionInfo.h

class CGFunctionInfo final
 
   /// How many arguments to pass inreg.
   unsigned HasRegParm : 1;
   unsigned RegParm : 3;
 
   /// Whether this function has nocf_check attribute.
   unsigned NoCfCheck : 1;
 
+  /// Log 2 of the maximum vector width.
+  unsigned MaxVectorWidth : 4;
+
   RequiredArgs Required;
 
   /// The struct representing all arguments passed in memory.  Only used when
   /// passing non-trivial types with inalloca.  Not part of the profile.
   llvm::StructType *ArgStruct;
   unsigned ArgStructAlign : 31;
   unsigned HasExtParameterInfos : 1;
 
public:
   CharUnits getArgStructAlignment() const {
     return CharUnits::fromQuantity(ArgStructAlign);
   }
   void setArgStruct(llvm::StructType *Ty, CharUnits Align) {
     ArgStruct = Ty;
     ArgStructAlign = Align.getQuantity();
   }
 
+  /// Return the maximum vector width in the arguments.
+  unsigned getMaxVectorWidth() const {
+    return MaxVectorWidth ? 1U << (MaxVectorWidth - 1) : 0;
+  }
+
+  /// Set the maximum vector width in the arguments.
+  void setMaxVectorWidth(unsigned Width) {
+    MaxVectorWidth = llvm::Log2_32(Width) + 1;
+  }
+
   void Profile(llvm::FoldingSetNodeID &ID) {
     ID.AddInteger(getASTCallingConvention());
     ID.AddBoolean(InstanceMethod);
     ID.AddBoolean(ChainCall);
     ID.AddBoolean(NoReturn);
     ID.AddBoolean(ReturnsRetained);
     ID.AddBoolean(NoCallerSavedRegs);
     ID.AddBoolean(HasRegParm);

clang/lib/CodeGen/CGCall.cpp

CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
   FI->Required = required;
   FI->HasRegParm = info.getHasRegParm();
   FI->RegParm = info.getRegParm();
   FI->ArgStruct = nullptr;
   FI->ArgStructAlign = 0;
   FI->NumArgs = argTypes.size();
   FI->HasExtParameterInfos = !paramInfos.empty();
   FI->getArgsBuffer()[0].type = resultType;
+  FI->MaxVectorWidth = 0;
   for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
     FI->getArgsBuffer()[i + 1].type = argTypes[i];
   for (unsigned i = 0, e = paramInfos.size(); i != e; ++i)
     FI->getExtParameterInfosBuffer()[i] = paramInfos[i];
   return FI;
 }
 
 /***/
AllocAlignAttrEmitter(CodeGenFunction &CGF_, const Decl *FuncDecl,
     Alignment = CallArgs[AA->getParamIndex().getLLVMIndex()]
                     .getRValue(CGF)
                     .getScalarVal();
   }
 };
 
 } // namespace
 
+unsigned getMaxVectorWidth(const llvm::Type *Ty) {
+  if (auto *VT = dyn_cast<llvm::VectorType>(Ty))
+    return VT->getPrimitiveSizeInBits().getKnownMinSize();
+  if (auto *AT = dyn_cast<llvm::ArrayType>(Ty))
+    return getMaxVectorWidth(AT->getElementType());
+
+  unsigned MaxVectorWidth = 0;
+  if (auto *ST = dyn_cast<llvm::StructType>(Ty))
+    for (auto *I : ST->elements())
+      MaxVectorWidth = std::max(MaxVectorWidth, getMaxVectorWidth(I));
+  return MaxVectorWidth;
+}
+
 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
                                  const CGCallee &Callee,
                                  ReturnValueSlot ReturnValue,
                                  const CallArgList &CallArgs,
                                  llvm::CallBase **callOrInvoke, bool IsMustTail,
                                  SourceLocation Loc) {
   // FIXME: We no longer need the types from CallArgs; lift up and simplify.
 
if (IRFunctionArgs.hasInallocaArg() &&
         i == IRFunctionArgs.getInallocaArgNo())
       continue;
     if (i < IRFuncTy->getNumParams())
       assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
   }
 #endif
 
   // Update the largest vector width if any arguments have vector types.
-  for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
-    if (auto *VT = dyn_cast<llvm::VectorType>(IRCallArgs[i]->getType()))
-      LargestVectorWidth =
-          std::max((uint64_t)LargestVectorWidth,
-                   VT->getPrimitiveSizeInBits().getKnownMinSize());
-  }
+  for (unsigned i = 0; i < IRCallArgs.size(); ++i)
+    LargestVectorWidth = std::max(LargestVectorWidth,
+                                  getMaxVectorWidth(IRCallArgs[i]->getType()));
 
   // Compute the calling convention and attributes.
   unsigned CallingConv;
   llvm::AttributeList Attrs;
   CGM.ConstructAttributeList(CalleePtr->getName(), CallInfo,
                              Callee.getAbstractInfo(), Attrs, CallingConv,
                              /*AttrOnCallSite=*/true,
                              /*IsThunk=*/false);
#endif
   CI->setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
 
   // Apply various metadata.
 
   if (!CI->getType()->isVoidTy())
     CI->setName("call");
 
   // Update largest vector width from the return type.
-  if (auto *VT = dyn_cast<llvm::VectorType>(CI->getType()))
-    LargestVectorWidth =
-        std::max((uint64_t)LargestVectorWidth,
-                 VT->getPrimitiveSizeInBits().getKnownMinSize());
+      std::max(LargestVectorWidth, getMaxVectorWidth(CI->getType()));
 
   // Insert instrumentation or attach profile metadata at indirect call sites.
   // For more details, see the comment before the definition of
   // IPVK_IndirectCallTarget in InstrProfData.inc.
   if (!CI->getCalledFunction())
     PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
                      CI, CalleePtr);
 

clang/lib/CodeGen/CodeGenFunction.cpp

if (auto *VT = dyn_cast<llvm::VectorType>(A.getType()))
                    VT->getPrimitiveSizeInBits().getKnownMinSize());
 
   // Update vector width based on return type.
   if (auto *VT = dyn_cast<llvm::VectorType>(CurFn->getReturnType()))
     LargestVectorWidth =
         std::max((uint64_t)LargestVectorWidth,
                  VT->getPrimitiveSizeInBits().getKnownMinSize());
 
+  if (CurFnInfo->getMaxVectorWidth() > LargestVectorWidth)
+    LargestVectorWidth = CurFnInfo->getMaxVectorWidth();
+
   // Add the required-vector-width attribute. This contains the max width from:
   // 1. min-vector-width attribute used in the source program.
   // 2. Any builtins used that have a vector width specified.
   // 3. Values passed in and out of inline assembly.
   // 4. Width of vector arguments and return types for this function.
   // 5. Width of vector aguments and return types for functions called by this
   //    function.
   CurFn->addFnAttr("min-legal-vector-width", llvm::utostr(LargestVectorWidth));

clang/lib/CodeGen/TargetInfo.cpp

class X86_64ABIInfo : public SwiftABIInfo {
   /// be passed in Memory then at least the classification of \arg Lo
   /// will be Memory.
   ///
   /// The \arg Lo class will be NoClass iff the argument is ignored.
   ///
   /// If the \arg Lo class is ComplexX87, then the \arg Hi class will
   /// also be ComplexX87.
   void classify(QualType T, uint64_t OffsetBase, Class &Lo, Class &Hi,
-                bool isNamedArg) const;
+                bool isNamedArg, bool IsRegCall = false) const;
 
   llvm::Type *GetByteVectorType(QualType Ty) const;
   llvm::Type *GetSSETypeAtOffset(llvm::Type *IRType,
                                  unsigned IROffset, QualType SourceTy,
                                  unsigned SourceOffset) const;
   llvm::Type *GetINTEGERTypeAtOffset(llvm::Type *IRType,
                                      unsigned IROffset, QualType SourceTy,
                                      unsigned SourceOffset) const;
 
   /// getIndirectResult - Give a source type \arg Ty, return a suitable result
   /// such that the argument will be returned in memory.
   ABIArgInfo getIndirectReturnResult(QualType Ty) const;
 
   /// getIndirectResult - Give a source type \arg Ty, return a suitable result
   /// such that the argument will be passed in memory.
   ///
   /// \param freeIntRegs - The number of free integer registers remaining
   /// available.
   ABIArgInfo getIndirectResult(QualType Ty, unsigned freeIntRegs) const;
 
   ABIArgInfo classifyReturnType(QualType RetTy) const;
 
   ABIArgInfo classifyArgumentType(QualType Ty, unsigned freeIntRegs,
                                   unsigned &neededInt, unsigned &neededSSE,
-                                  bool isNamedArg) const;
+                                  bool isNamedArg,
+                                  bool IsRegCall = false) const;
 
   ABIArgInfo classifyRegCallStructType(QualType Ty, unsigned &NeededInt,
-                                       unsigned &NeededSSE) const;
+                                       unsigned &NeededSSE,
+                                       unsigned &MaxSSEWidth) const;
 
   ABIArgInfo classifyRegCallStructTypeImpl(QualType Ty, unsigned &NeededInt,
-                                           unsigned &NeededSSE) const;
+                                           unsigned &NeededSSE,
+                                           unsigned &MaxSSEWidth) const;
 
   bool IsIllegalVectorType(QualType Ty) const;
 
   /// The 0.98 ABI revision clarified a lot of ambiguities,
   /// unfortunately in ways that were not always consistent with
   /// certain previous compilers.  In particular, platforms which
   /// required strict binary compatibility with older versions of GCC
   /// may need to exempt themselves.
X86_64ABIInfo::Class X86_64ABIInfo::merge(Class Accum, Class Field) {
   if (Accum == Integer || Field == Integer)
     return Integer;
   if (Field == X87 || Field == X87Up || Field == ComplexX87 ||
       Accum == X87 || Accum == X87Up)
     return Memory;
   return SSE;
 }
 
-void X86_64ABIInfo::classify(QualType Ty, uint64_t OffsetBase,
-                             Class &Lo, Class &Hi, bool isNamedArg) const {
+void X86_64ABIInfo::classify(QualType Ty, uint64_t OffsetBase, Class &Lo,
+                             Class &Hi, bool isNamedArg, bool IsRegCall) const {
   // FIXME: This code can be simplified by introducing a simple value class for
   // Class pairs with appropriate constructor methods for the various
   // situations.
 
   // FIXME: Some of the split computations are wrong; unaligned vectors
   // shouldn't be passed in registers for example, so there is no chance they
   // can straddle an eightbyte. Verify & simplify.
 
void X86_64ABIInfo::classify(QualType Ty, uint64_t OffsetBase, Class &Lo,
 
   if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) {
     // Arrays are treated like structures.
 
     uint64_t Size = getContext().getTypeSize(Ty);
 
     // AMD64-ABI 3.2.3p2: Rule 1. If the size of an object is larger
     // than eight eightbytes, ..., it has class MEMORY.
-    if (Size > 512)
+    if (!IsRegCall && Size > 512)
       return;
 
     // AMD64-ABI 3.2.3p2: Rule 1. If ..., or it contains unaligned
     // fields, it has class MEMORY.
     //
     // Only need to check alignment of array base.
     if (OffsetBase % getContext().getTypeAlign(AT->getElementType()))
       return;
classifyReturnType(QualType RetTy) const {
   // known to pass in the high eightbyte of the result.  We do this by forming a
   // first class struct aggregate with the high and low part: {low, high}
   if (HighPart)
     ResType = GetX86_64ByValArgumentPair(ResType, HighPart, getDataLayout());
 
   return ABIArgInfo::getDirect(ResType);
 }
 
 ABIArgInfo X86_64ABIInfo::classifyArgumentType(
   QualType Ty, unsigned freeIntRegs, unsigned &neededInt, unsigned &neededSSE,
-  bool isNamedArg)
+  bool isNamedArg, bool IsRegCall)
   const
 {
   Ty = useFirstFieldIfTransparentUnion(Ty);
 
   X86_64ABIInfo::Class Lo, Hi;
-  classify(Ty, 0, Lo, Hi, isNamedArg);
+  classify(Ty, 0, Lo, Hi, isNamedArg, IsRegCall);
 
   // Check some invariants.
   // FIXME: Enforce these by construction.
   assert((Hi != Memory || Lo == Memory) && "Invalid memory classification.");
   assert((Hi != SSEUp || Lo == SSE) && "Invalid SSEUp classification.");
 
   neededInt = 0;
   neededSSE = 0;
ABIArgInfo X86_64ABIInfo::classifyArgumentType(
   if (HighPart)
     ResType = GetX86_64ByValArgumentPair(ResType, HighPart, getDataLayout());
 
   return ABIArgInfo::getDirect(ResType);
 }
 
 ABIArgInfo
 X86_64ABIInfo::classifyRegCallStructTypeImpl(QualType Ty, unsigned &NeededInt,
-                                             unsigned &NeededSSE) const {
+                                             unsigned &NeededSSE,
+                                             unsigned &MaxSSEWidth) const {
   auto RT = Ty->getAs<RecordType>();
   assert(RT && "classifyRegCallStructType only valid with struct types");
 
   if (RT->getDecl()->hasFlexibleArrayMember())
     return getIndirectReturnResult(Ty);
 
   // Sum up bases
   if (auto CXXRD = dyn_cast<CXXRecordDecl>(RT->getDecl())) {
     if (CXXRD->isDynamicClass()) {
       NeededInt = NeededSSE = 0;
       return getIndirectReturnResult(Ty);
     }
 
     for (const auto &I : CXXRD->bases())
-      if (classifyRegCallStructTypeImpl(I.getType(), NeededInt, NeededSSE)
+      if (classifyRegCallStructTypeImpl(I.getType(), NeededInt, NeededSSE,
+                                        MaxSSEWidth)
               .isIndirect()) {
         NeededInt = NeededSSE = 0;
         return getIndirectReturnResult(Ty);
       }
   }
 
   // Sum up members
   for (const auto *FD : RT->getDecl()->fields()) {
-    if (FD->getType()->isRecordType() && !FD->getType()->isUnionType()) {
-      if (classifyRegCallStructTypeImpl(FD->getType(), NeededInt, NeededSSE)
+    QualType MTy = FD->getType();
+    if (MTy->isRecordType() && !MTy->isUnionType()) {
+      if (classifyRegCallStructTypeImpl(MTy, NeededInt, NeededSSE, MaxSSEWidth)
               .isIndirect()) {
         NeededInt = NeededSSE = 0;
         return getIndirectReturnResult(Ty);
       }
     } else {
       unsigned LocalNeededInt, LocalNeededSSE;
-      if (classifyArgumentType(FD->getType(), UINT_MAX, LocalNeededInt,
-                               LocalNeededSSE, true)
+      if (classifyArgumentType(MTy, UINT_MAX, LocalNeededInt, LocalNeededSSE,
+                               true, true)
               .isIndirect()) {
         NeededInt = NeededSSE = 0;
         return getIndirectReturnResult(Ty);
       }
+      if (const auto *AT = getContext().getAsConstantArrayType(MTy))
+        MTy = AT->getElementType();
+      if (const auto *VT = MTy->getAs<VectorType>())
+        if (getContext().getTypeSize(VT) > MaxSSEWidth)
+          MaxSSEWidth = getContext().getTypeSize(VT);
       NeededInt += LocalNeededInt;
       NeededSSE += LocalNeededSSE;
     }
   }
 
   return ABIArgInfo::getDirect();
 }
 
-ABIArgInfo X86_64ABIInfo::classifyRegCallStructType(QualType Ty,
-                                                    unsigned &NeededInt,
-                                                    unsigned &NeededSSE) const {
+ABIArgInfo
+X86_64ABIInfo::classifyRegCallStructType(QualType Ty, unsigned &NeededInt,
+                                         unsigned &NeededSSE,
+                                         unsigned &MaxSSEWidth) const {
 
   NeededInt = 0;
   NeededSSE = 0;
+  MaxSSEWidth = 0;
 
-  return classifyRegCallStructTypeImpl(Ty, NeededInt, NeededSSE);
+  return classifyRegCallStructTypeImpl(Ty, NeededInt, NeededSSE, MaxSSEWidth);
 }
 
 void X86_64ABIInfo::computeInfo(CGFunctionInfo &FI) const {
 
   const unsigned CallingConv = FI.getCallingConvention();
   // It is possible to force Win64 calling convention on any x86_64 target by
   // using __attribute__((ms_abi)). In such case to correctly emit Win64
   // compatible code delegate this call to WinX86_64ABIInfo::computeInfo.
   if (CallingConv == llvm::CallingConv::Win64) {
     WinX86_64ABIInfo Win64ABIInfo(CGT, AVXLevel);
     Win64ABIInfo.computeInfo(FI);
     return;
   }
 
   bool IsRegCall = CallingConv == llvm::CallingConv::X86_RegCall;
 
   // Keep track of the number of assigned registers.
   unsigned FreeIntRegs = IsRegCall ? 11 : 6;
   unsigned FreeSSERegs = IsRegCall ? 16 : 8;
-  unsigned NeededInt, NeededSSE;
+  unsigned NeededInt, NeededSSE, MaxSSEWidth = 0;
 
   if (!::classifyReturnType(getCXXABI(), FI, *this)) {
     if (IsRegCall && FI.getReturnType()->getTypePtr()->isRecordType() &&
         !FI.getReturnType()->getTypePtr()->isUnionType()) {
-      FI.getReturnInfo() =
-          classifyRegCallStructType(FI.getReturnType(), NeededInt, NeededSSE);
+      FI.getReturnInfo() = classifyRegCallStructType(
+          FI.getReturnType(), NeededInt, NeededSSE, MaxSSEWidth);
       if (FreeIntRegs >= NeededInt && FreeSSERegs >= NeededSSE) {
         FreeIntRegs -= NeededInt;
         FreeSSERegs -= NeededSSE;
       } else {
         FI.getReturnInfo() = getIndirectReturnResult(FI.getReturnType());
       }
     } else if (IsRegCall && FI.getReturnType()->getAs<ComplexType>() &&
                getContext().getCanonicalType(FI.getReturnType()
                                                  ->getAs<ComplexType>()
                                                  ->getElementType()) ==
                    getContext().LongDoubleTy)
       // Complex Long Double Type is passed in Memory when Regcall
       // calling convention is used.
       FI.getReturnInfo() = getIndirectReturnResult(FI.getReturnType());
     else
       FI.getReturnInfo() = classifyReturnType(FI.getReturnType());
   }
 
   // If the return value is indirect, then the hidden argument is consuming one
   // integer register.
   if (FI.getReturnInfo().isIndirect())
     --FreeIntRegs;
+  else if (NeededSSE)
+    FI.setMaxVectorWidth(MaxSSEWidth);
 
   // The chain argument effectively gives us another free register.
   if (FI.isChainCall())
     ++FreeIntRegs;
 
   unsigned NumRequiredArgs = FI.getNumRequiredArgs();
   // AMD64-ABI 3.2.3p3: Once arguments are classified, the registers
   // get assigned (in left-to-right order) for passing as follows...
   unsigned ArgNo = 0;
   for (CGFunctionInfo::arg_iterator it = FI.arg_begin(), ie = FI.arg_end();
        it != ie; ++it, ++ArgNo) {
     bool IsNamedArg = ArgNo < NumRequiredArgs;
 
     if (IsRegCall && it->type->isStructureOrClassType())
-      it->info = classifyRegCallStructType(it->type, NeededInt, NeededSSE);
+      it->info = classifyRegCallStructType(it->type, NeededInt, NeededSSE,
+                                           MaxSSEWidth);
     else
       it->info = classifyArgumentType(it->type, FreeIntRegs, NeededInt,
                                       NeededSSE, IsNamedArg);
 
     // AMD64-ABI 3.2.3p3: If there are no registers available for any
     // eightbyte of an argument, the whole argument is passed on the
     // stack. If registers have already been assigned for some
     // eightbytes of such an argument, the assignments get reverted.
     if (FreeIntRegs >= NeededInt && FreeSSERegs >= NeededSSE) {
       FreeIntRegs -= NeededInt;
       FreeSSERegs -= NeededSSE;
+      if (MaxSSEWidth > FI.getMaxVectorWidth())
+        FI.setMaxVectorWidth(MaxSSEWidth);
     } else {
       it->info = getIndirectResult(it->type, FreeIntRegs);
     }
   }
 }
 
 static Address EmitX86_64VAArgFromMemory(CodeGenFunction &CGF,
                                          Address VAListAddr, QualType Ty) {

clang/test/CodeGen/aarch64-neon-tbl.c

 // CHECK:   [[TMP2:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX2_I]], align 8
 // CHECK:   [[VTBL1_I:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> [[TMP2]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL13_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbl1.v8i8(<16 x i8> [[VTBL1_I]], <8 x i8> %b) #3
 // CHECK:   ret <8 x i8> [[VTBL13_I]]
 int8x8_t test_vtbl2_s8(int8x8x2_t a, int8x8_t b) {
   return vtbl2_s8(a, b);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl2_s8([2 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl2_s8([2 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #1 {
 // CHECK:   [[__P0_I:%.*]] = alloca %struct.int8x16x2_t, align 16
 // CHECK:   [[A:%.*]] = alloca %struct.int8x16x2_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.int8x16x2_t, %struct.int8x16x2_t* [[A]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[A]].coerce, [2 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.int8x16x2_t, %struct.int8x16x2_t* [[A]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [2 x <16 x i8>], [2 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.int8x16x2_t, %struct.int8x16x2_t* [[__P0_I]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[TMP0]], [2 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 // CHECK:   [[VTBL2_I:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> [[TMP2]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL25_I:%.*]] = shufflevector <8 x i8> [[TMP3]], <8 x i8> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL26_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbl2.v8i8(<16 x i8> [[VTBL2_I]], <16 x i8> [[VTBL25_I]], <8 x i8> %b) #3
 // CHECK:   ret <8 x i8> [[VTBL26_I]]
 int8x8_t test_vtbl3_s8(int8x8x3_t a, int8x8_t b) {
   return vtbl3_s8(a, b);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl3_s8([3 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl3_s8([3 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #1 {
 // CHECK:   [[__P0_I:%.*]] = alloca %struct.int8x16x3_t, align 16
 // CHECK:   [[A:%.*]] = alloca %struct.int8x16x3_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.int8x16x3_t, %struct.int8x16x3_t* [[A]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[A]].coerce, [3 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.int8x16x3_t, %struct.int8x16x3_t* [[A]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [3 x <16 x i8>], [3 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.int8x16x3_t, %struct.int8x16x3_t* [[__P0_I]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[TMP0]], [3 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 // CHECK:   [[VTBL2_I:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> [[TMP2]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL27_I:%.*]] = shufflevector <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL28_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbl2.v8i8(<16 x i8> [[VTBL2_I]], <16 x i8> [[VTBL27_I]], <8 x i8> %b) #3
 // CHECK:   ret <8 x i8> [[VTBL28_I]]
 int8x8_t test_vtbl4_s8(int8x8x4_t a, int8x8_t b) {
   return vtbl4_s8(a, b);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl4_s8([4 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl4_s8([4 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #1 {
 // CHECK:   [[__P0_I:%.*]] = alloca %struct.int8x16x4_t, align 16
 // CHECK:   [[A:%.*]] = alloca %struct.int8x16x4_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.int8x16x4_t, %struct.int8x16x4_t* [[A]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[A]].coerce, [4 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.int8x16x4_t, %struct.int8x16x4_t* [[A]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [4 x <16 x i8>], [4 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.int8x16x4_t, %struct.int8x16x4_t* [[__P0_I]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[TMP0]], [4 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 
 // CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx1_s8(<8 x i8> noundef %a, <16 x i8> noundef %b, <8 x i8> noundef %c) #1 {
 // CHECK:   [[VTBX1_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbx1.v8i8(<8 x i8> %a, <16 x i8> %b, <8 x i8> %c) #3
 // CHECK:   ret <8 x i8> [[VTBX1_I]]
 int8x8_t test_vqtbx1_s8(int8x8_t a, int8x16_t b, uint8x8_t c) {
   return vqtbx1_s8(a, b, c);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx2_s8(<8 x i8> noundef %a, [2 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx2_s8(<8 x i8> noundef %a, [2 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #1 {
 // CHECK:   [[__P1_I:%.*]] = alloca %struct.int8x16x2_t, align 16
 // CHECK:   [[B:%.*]] = alloca %struct.int8x16x2_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.int8x16x2_t, %struct.int8x16x2_t* [[B]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[B]].coerce, [2 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.int8x16x2_t, %struct.int8x16x2_t* [[B]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [2 x <16 x i8>], [2 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.int8x16x2_t, %struct.int8x16x2_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[TMP0]], [2 x <16 x i8>]* [[COERCE_DIVE_I]], align 16
 // CHECK:   [[VAL_I:%.*]] = getelementptr inbounds %struct.int8x16x2_t, %struct.int8x16x2_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x <16 x i8>], [2 x <16 x i8>]* [[VAL_I]], i64 0, i64 0
 // CHECK:   [[TMP1:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX_I]], align 16
 // CHECK:   [[VAL1_I:%.*]] = getelementptr inbounds %struct.int8x16x2_t, %struct.int8x16x2_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2_I:%.*]] = getelementptr inbounds [2 x <16 x i8>], [2 x <16 x i8>]* [[VAL1_I]], i64 0, i64 1
 // CHECK:   [[TMP2:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX2_I]], align 16
 // CHECK:   [[VTBX2_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbx2.v8i8(<8 x i8> %a, <16 x i8> [[TMP1]], <16 x i8> [[TMP2]], <8 x i8> %c) #3
 // CHECK:   ret <8 x i8> [[VTBX2_I]]
 int8x8_t test_vqtbx2_s8(int8x8_t a, int8x16x2_t b, uint8x8_t c) {
   return vqtbx2_s8(a, b, c);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx3_s8(<8 x i8> noundef %a, [3 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx3_s8(<8 x i8> noundef %a, [3 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #1 {
 // CHECK:   [[__P1_I:%.*]] = alloca %struct.int8x16x3_t, align 16
 // CHECK:   [[B:%.*]] = alloca %struct.int8x16x3_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.int8x16x3_t, %struct.int8x16x3_t* [[B]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[B]].coerce, [3 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.int8x16x3_t, %struct.int8x16x3_t* [[B]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [3 x <16 x i8>], [3 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.int8x16x3_t, %struct.int8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[TMP0]], [3 x <16 x i8>]* [[COERCE_DIVE_I]], align 16
 // CHECK:   [[VAL_I:%.*]] = getelementptr inbounds %struct.int8x16x3_t, %struct.int8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX_I:%.*]] = getelementptr inbounds [3 x <16 x i8>], [3 x <16 x i8>]* [[VAL_I]], i64 0, i64 0
 // CHECK:   [[TMP1:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX_I]], align 16
 // CHECK:   [[VAL1_I:%.*]] = getelementptr inbounds %struct.int8x16x3_t, %struct.int8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2_I:%.*]] = getelementptr inbounds [3 x <16 x i8>], [3 x <16 x i8>]* [[VAL1_I]], i64 0, i64 1
 // CHECK:   [[TMP2:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX2_I]], align 16
 // CHECK:   [[VAL3_I:%.*]] = getelementptr inbounds %struct.int8x16x3_t, %struct.int8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4_I:%.*]] = getelementptr inbounds [3 x <16 x i8>], [3 x <16 x i8>]* [[VAL3_I]], i64 0, i64 2
 // CHECK:   [[TMP3:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4_I]], align 16
 // CHECK:   [[VTBX3_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbx3.v8i8(<8 x i8> %a, <16 x i8> [[TMP1]], <16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <8 x i8> %c) #3
 // CHECK:   ret <8 x i8> [[VTBX3_I]]
 int8x8_t test_vqtbx3_s8(int8x8_t a, int8x16x3_t b, uint8x8_t c) {
   return vqtbx3_s8(a, b, c);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx4_s8(<8 x i8> noundef %a, [4 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx4_s8(<8 x i8> noundef %a, [4 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #1 {
 // CHECK:   [[__P1_I:%.*]] = alloca %struct.int8x16x4_t, align 16
 // CHECK:   [[B:%.*]] = alloca %struct.int8x16x4_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.int8x16x4_t, %struct.int8x16x4_t* [[B]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[B]].coerce, [4 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.int8x16x4_t, %struct.int8x16x4_t* [[B]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [4 x <16 x i8>], [4 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.int8x16x4_t, %struct.int8x16x4_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[TMP0]], [4 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 // CHECK:   [[TMP2:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX2_I]], align 8
 // CHECK:   [[VTBL1_I:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> [[TMP2]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL13_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbl1.v8i8(<16 x i8> [[VTBL1_I]], <8 x i8> %b) #3
 // CHECK:   ret <8 x i8> [[VTBL13_I]]
 uint8x8_t test_vtbl2_u8(uint8x8x2_t a, uint8x8_t b) {
   return vtbl2_u8(a, b);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl2_u8([2 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl2_u8([2 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #1 {
 // CHECK:   [[__P0_I:%.*]] = alloca %struct.uint8x16x2_t, align 16
 // CHECK:   [[A:%.*]] = alloca %struct.uint8x16x2_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.uint8x16x2_t, %struct.uint8x16x2_t* [[A]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[A]].coerce, [2 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.uint8x16x2_t, %struct.uint8x16x2_t* [[A]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [2 x <16 x i8>], [2 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.uint8x16x2_t, %struct.uint8x16x2_t* [[__P0_I]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[TMP0]], [2 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 // CHECK:   [[VTBL2_I:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> [[TMP2]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL25_I:%.*]] = shufflevector <8 x i8> [[TMP3]], <8 x i8> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL26_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbl2.v8i8(<16 x i8> [[VTBL2_I]], <16 x i8> [[VTBL25_I]], <8 x i8> %b) #3
 // CHECK:   ret <8 x i8> [[VTBL26_I]]
 uint8x8_t test_vtbl3_u8(uint8x8x3_t a, uint8x8_t b) {
   return vtbl3_u8(a, b);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl3_u8([3 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl3_u8([3 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #1 {
 // CHECK:   [[__P0_I:%.*]] = alloca %struct.uint8x16x3_t, align 16
 // CHECK:   [[A:%.*]] = alloca %struct.uint8x16x3_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.uint8x16x3_t, %struct.uint8x16x3_t* [[A]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[A]].coerce, [3 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.uint8x16x3_t, %struct.uint8x16x3_t* [[A]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [3 x <16 x i8>], [3 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.uint8x16x3_t, %struct.uint8x16x3_t* [[__P0_I]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[TMP0]], [3 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 // CHECK:   [[VTBL2_I:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> [[TMP2]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL27_I:%.*]] = shufflevector <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL28_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbl2.v8i8(<16 x i8> [[VTBL2_I]], <16 x i8> [[VTBL27_I]], <8 x i8> %b) #3
 // CHECK:   ret <8 x i8> [[VTBL28_I]]
 uint8x8_t test_vtbl4_u8(uint8x8x4_t a, uint8x8_t b) {
   return vtbl4_u8(a, b);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl4_u8([4 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl4_u8([4 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #1 {
 // CHECK:   [[__P0_I:%.*]] = alloca %struct.uint8x16x4_t, align 16
 // CHECK:   [[A:%.*]] = alloca %struct.uint8x16x4_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.uint8x16x4_t, %struct.uint8x16x4_t* [[A]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[A]].coerce, [4 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.uint8x16x4_t, %struct.uint8x16x4_t* [[A]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [4 x <16 x i8>], [4 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.uint8x16x4_t, %struct.uint8x16x4_t* [[__P0_I]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[TMP0]], [4 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 
 // CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx1_u8(<8 x i8> noundef %a, <16 x i8> noundef %b, <8 x i8> noundef %c) #1 {
 // CHECK:   [[VTBX1_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbx1.v8i8(<8 x i8> %a, <16 x i8> %b, <8 x i8> %c) #3
 // CHECK:   ret <8 x i8> [[VTBX1_I]]
 uint8x8_t test_vqtbx1_u8(uint8x8_t a, uint8x16_t b, uint8x8_t c) {
   return vqtbx1_u8(a, b, c);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx2_u8(<8 x i8> noundef %a, [2 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx2_u8(<8 x i8> noundef %a, [2 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #1 {
 // CHECK:   [[__P1_I:%.*]] = alloca %struct.uint8x16x2_t, align 16
 // CHECK:   [[B:%.*]] = alloca %struct.uint8x16x2_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.uint8x16x2_t, %struct.uint8x16x2_t* [[B]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[B]].coerce, [2 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.uint8x16x2_t, %struct.uint8x16x2_t* [[B]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [2 x <16 x i8>], [2 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.uint8x16x2_t, %struct.uint8x16x2_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[TMP0]], [2 x <16 x i8>]* [[COERCE_DIVE_I]], align 16
 // CHECK:   [[VAL_I:%.*]] = getelementptr inbounds %struct.uint8x16x2_t, %struct.uint8x16x2_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x <16 x i8>], [2 x <16 x i8>]* [[VAL_I]], i64 0, i64 0
 // CHECK:   [[TMP1:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX_I]], align 16
 // CHECK:   [[VAL1_I:%.*]] = getelementptr inbounds %struct.uint8x16x2_t, %struct.uint8x16x2_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2_I:%.*]] = getelementptr inbounds [2 x <16 x i8>], [2 x <16 x i8>]* [[VAL1_I]], i64 0, i64 1
 // CHECK:   [[TMP2:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX2_I]], align 16
 // CHECK:   [[VTBX2_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbx2.v8i8(<8 x i8> %a, <16 x i8> [[TMP1]], <16 x i8> [[TMP2]], <8 x i8> %c) #3
 // CHECK:   ret <8 x i8> [[VTBX2_I]]
 uint8x8_t test_vqtbx2_u8(uint8x8_t a, uint8x16x2_t b, uint8x8_t c) {
   return vqtbx2_u8(a, b, c);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx3_u8(<8 x i8> noundef %a, [3 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx3_u8(<8 x i8> noundef %a, [3 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #1 {
 // CHECK:   [[__P1_I:%.*]] = alloca %struct.uint8x16x3_t, align 16
 // CHECK:   [[B:%.*]] = alloca %struct.uint8x16x3_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.uint8x16x3_t, %struct.uint8x16x3_t* [[B]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[B]].coerce, [3 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.uint8x16x3_t, %struct.uint8x16x3_t* [[B]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [3 x <16 x i8>], [3 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.uint8x16x3_t, %struct.uint8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[TMP0]], [3 x <16 x i8>]* [[COERCE_DIVE_I]], align 16
 // CHECK:   [[VAL_I:%.*]] = getelementptr inbounds %struct.uint8x16x3_t, %struct.uint8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX_I:%.*]] = getelementptr inbounds [3 x <16 x i8>], [3 x <16 x i8>]* [[VAL_I]], i64 0, i64 0
 // CHECK:   [[TMP1:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX_I]], align 16
 // CHECK:   [[VAL1_I:%.*]] = getelementptr inbounds %struct.uint8x16x3_t, %struct.uint8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2_I:%.*]] = getelementptr inbounds [3 x <16 x i8>], [3 x <16 x i8>]* [[VAL1_I]], i64 0, i64 1
 // CHECK:   [[TMP2:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX2_I]], align 16
 // CHECK:   [[VAL3_I:%.*]] = getelementptr inbounds %struct.uint8x16x3_t, %struct.uint8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4_I:%.*]] = getelementptr inbounds [3 x <16 x i8>], [3 x <16 x i8>]* [[VAL3_I]], i64 0, i64 2
 // CHECK:   [[TMP3:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4_I]], align 16
 // CHECK:   [[VTBX3_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbx3.v8i8(<8 x i8> %a, <16 x i8> [[TMP1]], <16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <8 x i8> %c) #3
 // CHECK:   ret <8 x i8> [[VTBX3_I]]
 uint8x8_t test_vqtbx3_u8(uint8x8_t a, uint8x16x3_t b, uint8x8_t c) {
   return vqtbx3_u8(a, b, c);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx4_u8(<8 x i8> noundef %a, [4 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx4_u8(<8 x i8> noundef %a, [4 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #1 {
 // CHECK:   [[__P1_I:%.*]] = alloca %struct.uint8x16x4_t, align 16
 // CHECK:   [[B:%.*]] = alloca %struct.uint8x16x4_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.uint8x16x4_t, %struct.uint8x16x4_t* [[B]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[B]].coerce, [4 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.uint8x16x4_t, %struct.uint8x16x4_t* [[B]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [4 x <16 x i8>], [4 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.uint8x16x4_t, %struct.uint8x16x4_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[TMP0]], [4 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 // CHECK:   [[TMP2:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX2_I]], align 8
 // CHECK:   [[VTBL1_I:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> [[TMP2]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL13_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbl1.v8i8(<16 x i8> [[VTBL1_I]], <8 x i8> %b) #3
 // CHECK:   ret <8 x i8> [[VTBL13_I]]
 poly8x8_t test_vtbl2_p8(poly8x8x2_t a, uint8x8_t b) {
   return vtbl2_p8(a, b);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl2_p8([2 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl2_p8([2 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #1 {
 // CHECK:   [[__P0_I:%.*]] = alloca %struct.poly8x16x2_t, align 16
 // CHECK:   [[A:%.*]] = alloca %struct.poly8x16x2_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.poly8x16x2_t, %struct.poly8x16x2_t* [[A]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[A]].coerce, [2 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.poly8x16x2_t, %struct.poly8x16x2_t* [[A]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [2 x <16 x i8>], [2 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.poly8x16x2_t, %struct.poly8x16x2_t* [[__P0_I]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[TMP0]], [2 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 // CHECK:   [[VTBL2_I:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> [[TMP2]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL25_I:%.*]] = shufflevector <8 x i8> [[TMP3]], <8 x i8> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL26_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbl2.v8i8(<16 x i8> [[VTBL2_I]], <16 x i8> [[VTBL25_I]], <8 x i8> %b) #3
 // CHECK:   ret <8 x i8> [[VTBL26_I]]
 poly8x8_t test_vtbl3_p8(poly8x8x3_t a, uint8x8_t b) {
   return vtbl3_p8(a, b);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl3_p8([3 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl3_p8([3 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #1 {
 // CHECK:   [[__P0_I:%.*]] = alloca %struct.poly8x16x3_t, align 16
 // CHECK:   [[A:%.*]] = alloca %struct.poly8x16x3_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.poly8x16x3_t, %struct.poly8x16x3_t* [[A]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[A]].coerce, [3 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.poly8x16x3_t, %struct.poly8x16x3_t* [[A]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [3 x <16 x i8>], [3 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.poly8x16x3_t, %struct.poly8x16x3_t* [[__P0_I]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[TMP0]], [3 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 // CHECK:   [[VTBL2_I:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> [[TMP2]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL27_I:%.*]] = shufflevector <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
 // CHECK:   [[VTBL28_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbl2.v8i8(<16 x i8> [[VTBL2_I]], <16 x i8> [[VTBL27_I]], <8 x i8> %b) #3
 // CHECK:   ret <8 x i8> [[VTBL28_I]]
 poly8x8_t test_vtbl4_p8(poly8x8x4_t a, uint8x8_t b) {
   return vtbl4_p8(a, b);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl4_p8([4 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbl4_p8([4 x <16 x i8>] %a.coerce, <8 x i8> noundef %b) #1 {
 // CHECK:   [[__P0_I:%.*]] = alloca %struct.poly8x16x4_t, align 16
 // CHECK:   [[A:%.*]] = alloca %struct.poly8x16x4_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.poly8x16x4_t, %struct.poly8x16x4_t* [[A]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[A]].coerce, [4 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.poly8x16x4_t, %struct.poly8x16x4_t* [[A]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [4 x <16 x i8>], [4 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.poly8x16x4_t, %struct.poly8x16x4_t* [[__P0_I]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[TMP0]], [4 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

 
 // CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx1_p8(<8 x i8> noundef %a, <16 x i8> noundef %b, <8 x i8> noundef %c) #1 {
 // CHECK:   [[VTBX1_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbx1.v8i8(<8 x i8> %a, <16 x i8> %b, <8 x i8> %c) #3
 // CHECK:   ret <8 x i8> [[VTBX1_I]]
 poly8x8_t test_vqtbx1_p8(poly8x8_t a, uint8x16_t b, uint8x8_t c) {
   return vqtbx1_p8(a, b, c);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx2_p8(<8 x i8> noundef %a, [2 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx2_p8(<8 x i8> noundef %a, [2 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #1 {
 // CHECK:   [[__P1_I:%.*]] = alloca %struct.poly8x16x2_t, align 16
 // CHECK:   [[B:%.*]] = alloca %struct.poly8x16x2_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.poly8x16x2_t, %struct.poly8x16x2_t* [[B]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[B]].coerce, [2 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.poly8x16x2_t, %struct.poly8x16x2_t* [[B]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [2 x <16 x i8>], [2 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.poly8x16x2_t, %struct.poly8x16x2_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   store [2 x <16 x i8>] [[TMP0]], [2 x <16 x i8>]* [[COERCE_DIVE_I]], align 16
 // CHECK:   [[VAL_I:%.*]] = getelementptr inbounds %struct.poly8x16x2_t, %struct.poly8x16x2_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x <16 x i8>], [2 x <16 x i8>]* [[VAL_I]], i64 0, i64 0
 // CHECK:   [[TMP1:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX_I]], align 16
 // CHECK:   [[VAL1_I:%.*]] = getelementptr inbounds %struct.poly8x16x2_t, %struct.poly8x16x2_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2_I:%.*]] = getelementptr inbounds [2 x <16 x i8>], [2 x <16 x i8>]* [[VAL1_I]], i64 0, i64 1
 // CHECK:   [[TMP2:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX2_I]], align 16
 // CHECK:   [[VTBX2_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbx2.v8i8(<8 x i8> %a, <16 x i8> [[TMP1]], <16 x i8> [[TMP2]], <8 x i8> %c) #3
 // CHECK:   ret <8 x i8> [[VTBX2_I]]
 poly8x8_t test_vqtbx2_p8(poly8x8_t a, poly8x16x2_t b, uint8x8_t c) {
   return vqtbx2_p8(a, b, c);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx3_p8(<8 x i8> noundef %a, [3 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx3_p8(<8 x i8> noundef %a, [3 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #1 {
 // CHECK:   [[__P1_I:%.*]] = alloca %struct.poly8x16x3_t, align 16
 // CHECK:   [[B:%.*]] = alloca %struct.poly8x16x3_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.poly8x16x3_t, %struct.poly8x16x3_t* [[B]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[B]].coerce, [3 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.poly8x16x3_t, %struct.poly8x16x3_t* [[B]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [3 x <16 x i8>], [3 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.poly8x16x3_t, %struct.poly8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   store [3 x <16 x i8>] [[TMP0]], [3 x <16 x i8>]* [[COERCE_DIVE_I]], align 16
 // CHECK:   [[VAL_I:%.*]] = getelementptr inbounds %struct.poly8x16x3_t, %struct.poly8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX_I:%.*]] = getelementptr inbounds [3 x <16 x i8>], [3 x <16 x i8>]* [[VAL_I]], i64 0, i64 0
 // CHECK:   [[TMP1:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX_I]], align 16
 // CHECK:   [[VAL1_I:%.*]] = getelementptr inbounds %struct.poly8x16x3_t, %struct.poly8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2_I:%.*]] = getelementptr inbounds [3 x <16 x i8>], [3 x <16 x i8>]* [[VAL1_I]], i64 0, i64 1
 // CHECK:   [[TMP2:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX2_I]], align 16
 // CHECK:   [[VAL3_I:%.*]] = getelementptr inbounds %struct.poly8x16x3_t, %struct.poly8x16x3_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4_I:%.*]] = getelementptr inbounds [3 x <16 x i8>], [3 x <16 x i8>]* [[VAL3_I]], i64 0, i64 2
 // CHECK:   [[TMP3:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4_I]], align 16
 // CHECK:   [[VTBX3_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.tbx3.v8i8(<8 x i8> %a, <16 x i8> [[TMP1]], <16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <8 x i8> %c) #3
 // CHECK:   ret <8 x i8> [[VTBX3_I]]
 poly8x8_t test_vqtbx3_p8(poly8x8_t a, poly8x16x3_t b, uint8x8_t c) {
   return vqtbx3_p8(a, b, c);
 }
 
-// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx4_p8(<8 x i8> noundef %a, [4 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #0 {
+// CHECK-LABEL: define{{.*}} <8 x i8> @test_vqtbx4_p8(<8 x i8> noundef %a, [4 x <16 x i8>] %b.coerce, <8 x i8> noundef %c) #1 {
 // CHECK:   [[__P1_I:%.*]] = alloca %struct.poly8x16x4_t, align 16
 // CHECK:   [[B:%.*]] = alloca %struct.poly8x16x4_t, align 16
 // CHECK:   [[COERCE_DIVE:%.*]] = getelementptr inbounds %struct.poly8x16x4_t, %struct.poly8x16x4_t* [[B]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[B]].coerce, [4 x <16 x i8>]* [[COERCE_DIVE]], align 16
 // CHECK:   [[COERCE_DIVE1:%.*]] = getelementptr inbounds %struct.poly8x16x4_t, %struct.poly8x16x4_t* [[B]], i32 0, i32 0
 // CHECK:   [[TMP0:%.*]] = load [4 x <16 x i8>], [4 x <16 x i8>]* [[COERCE_DIVE1]], align 16
 // CHECK:   [[COERCE_DIVE_I:%.*]] = getelementptr inbounds %struct.poly8x16x4_t, %struct.poly8x16x4_t* [[__P1_I]], i32 0, i32 0
 // CHECK:   store [4 x <16 x i8>] [[TMP0]], [4 x <16 x i8>]* [[COERCE_DIVE_I]], align 16

clang/test/CodeGen/regcall2.c

+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py
+// RUN: %clang_cc1 -emit-llvm %s -o - -ffreestanding -target-feature +avx512vl -triple=x86_64-pc-win32     | FileCheck %s --check-prefix=Win
+// RUN: %clang_cc1 -emit-llvm %s -o - -ffreestanding -target-feature +avx512vl -triple=x86_64-pc-linux-gnu | FileCheck %s --check-prefix=Lin
+
+#include <immintrin.h>
+
+typedef struct {
+  __m512d r1[4];
+  __m512 r2[4];
+} __sVector;
+__sVector A;
+
+__sVector __regcall foo(int a) {
+  return A;
+}
+
+double __regcall bar(__sVector a) {
+  return a.r1[0][4];
+}
+
+// FIXME: Do we need to change for Windows?
+// Win: define dso_local x86_regcallcc void @__regcall3__foo(%struct.__sVector* noalias sret(%struct.__sVector) align 64 %agg.result, i32 noundef %a) #0
+// Win: define dso_local x86_regcallcc double @__regcall3__bar(%struct.__sVector* noundef %a) #0
+// Win: attributes #0 = { noinline nounwind optnone "frame-pointer"="none" "min-legal-vector-width"="0" "no-builtins" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-features"="+avx,+avx2,+avx512f,+avx512vl,+crc32,+cx8,+f16c,+fma,+mmx,+popcnt,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87,+xsave" }
+
+// Lin: define dso_local x86_regcallcc %struct.__sVector @__regcall3__foo(i32 noundef %a) #0
+// Lin: define dso_local x86_regcallcc double @__regcall3__bar([4 x <8 x double>] %a.coerce0, [4 x <16 x float>] %a.coerce1) #0
+// Lin: attributes #0 = { noinline nounwind optnone "frame-pointer"="none" "min-legal-vector-width"="512" "no-builtins" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-features"="+avx,+avx2,+avx512f,+avx512vl,+crc32,+cx8,+f16c,+fma,+mmx,+popcnt,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87,+xsave" }