[CodeGen][ARM] Coerce FP16 vectors to integer vectors when needed

[miyuki]

Summary

On targets that do not support FP16 natively LLVM currently legalizes
vectors of FP16 values by scalarizing them and promoting to FP32. This
causes problems for the following code:

void foo(int, ...);

typedef __attribute__((neon_vector_type(4))) __fp16 float16x4_t;
void bar(float16x4_t x) {
  foo(42, x);
}

According to the AAPCS (appendix A.2) float16x4_t is a containerized
vector fundamental type, so 'foo' expects that the 4 16-bit FP values
are packed into 2 32-bit registers, but instead bar promotes them to
4 single precision values.

Since we already handle scalar FP16 values in the frontend by
bitcasting them to/from integers, this patch adds similar handling for
vector types and homogeneous FP16 vector aggregates.

One existing test required some adjustments because we now generate
more bitcasts (so the patch changes the test to target a machine with
native FP16 support).

[efriedma]

Do we need to check for homogeneous aggregates of half vectors somewhere?

[miyuki]

Fixed handling of homogeneous aggregates of FP16 vectors

[efriedma]

lib/CodeGen/TargetInfo.cpp

5824

Do we need equivalent code in classifyReturnType?

[miyuki]

Handle return of homogeneous aggregates

[miyuki]

ping

[efriedma]

LGTM

lib/CodeGen/TargetInfo.cpp

bool isEABIHF() const {
     }
   }
 
   ABIKind getABIKind() const { return Kind; }
 
 private:
   ABIArgInfo classifyReturnType(QualType RetTy, bool isVariadic) const;
   ABIArgInfo classifyArgumentType(QualType RetTy, bool isVariadic) const;
+  ABIArgInfo classifyHomogeneousAggregate(QualType Ty, const Type *Base,
+                                          uint64_t Members) const;
+  ABIArgInfo coerceIllegalVector(QualType Ty) const;
   bool isIllegalVectorType(QualType Ty) const;
 
   bool isHomogeneousAggregateBaseType(QualType Ty) const override;
   bool isHomogeneousAggregateSmallEnough(const Type *Ty,
                                          uint64_t Members) const override;
 
   void computeInfo(CGFunctionInfo &FI) const override;
 
void ARMABIInfo::setCCs() {
 
   // Don't muddy up the IR with a ton of explicit annotations if
   // they'd just match what LLVM will infer from the triple.
   llvm::CallingConv::ID abiCC = getABIDefaultCC();
   if (abiCC != getLLVMDefaultCC())
     RuntimeCC = abiCC;
 }
 
+ABIArgInfo ARMABIInfo::coerceIllegalVector(QualType Ty) const {
+  uint64_t Size = getContext().getTypeSize(Ty);
+  if (Size <= 32) {
+    llvm::Type *ResType =
+        llvm::Type::getInt32Ty(getVMContext());
+    return ABIArgInfo::getDirect(ResType);
+  }
+  if (Size == 64 || Size == 128) {
+    llvm::Type *ResType = llvm::VectorType::get(
+        llvm::Type::getInt32Ty(getVMContext()), Size / 32);
+    return ABIArgInfo::getDirect(ResType);
+  }
+  return getNaturalAlignIndirect(Ty, /*ByVal=*/false);
+}
+
+ABIArgInfo ARMABIInfo::classifyHomogeneousAggregate(QualType Ty,
+                                                    const Type *Base,
+                                                    uint64_t Members) const {
+  assert(Base && "Base class should be set for homogeneous aggregate");
+  // Base can be a floating-point or a vector.
+  if (const VectorType *VT = Base->getAs<VectorType>()) {
+    // FP16 vectors should be converted to integer vectors
+    if (!getTarget().hasLegalHalfType() &&
+        (VT->getElementType()->isFloat16Type() ||
+          VT->getElementType()->isHalfType())) {
+      uint64_t Size = getContext().getTypeSize(VT);
+      llvm::Type *NewVecTy = llvm::VectorType::get(
+          llvm::Type::getInt32Ty(getVMContext()), Size / 32);
+      llvm::Type *Ty = llvm::ArrayType::get(NewVecTy, Members);
+      return ABIArgInfo::getDirect(Ty, 0, nullptr, false);
+    }
+  }
+  return ABIArgInfo::getDirect(nullptr, 0, nullptr, false);
+}
+
 ABIArgInfo ARMABIInfo::classifyArgumentType(QualType Ty,
                                             bool isVariadic) const {
   // 6.1.2.1 The following argument types are VFP CPRCs:
   //   A single-precision floating-point type (including promoted
   //   half-precision types); A double-precision floating-point type;
   //   A 64-bit or 128-bit containerized vector type; Homogeneous Aggregate
   //   with a Base Type of a single- or double-precision floating-point type,
   //   64-bit containerized vectors or 128-bit containerized vectors with one
   //   to four Elements.
   bool IsEffectivelyAAPCS_VFP = getABIKind() == AAPCS_VFP && !isVariadic;
 
   Ty = useFirstFieldIfTransparentUnion(Ty);
 
   // Handle illegal vector types here.
-  if (isIllegalVectorType(Ty)) {
-    uint64_t Size = getContext().getTypeSize(Ty);
-    if (Size <= 32) {
-      llvm::Type *ResType =
-          llvm::Type::getInt32Ty(getVMContext());
-      return ABIArgInfo::getDirect(ResType);
-    }
-    if (Size == 64) {
-      llvm::Type *ResType = llvm::VectorType::get(
-          llvm::Type::getInt32Ty(getVMContext()), 2);
-      return ABIArgInfo::getDirect(ResType);
-    }
-    if (Size == 128) {
-      llvm::Type *ResType = llvm::VectorType::get(
-          llvm::Type::getInt32Ty(getVMContext()), 4);
-      return ABIArgInfo::getDirect(ResType);
-    }
-    return getNaturalAlignIndirect(Ty, /*ByVal=*/false);
-  }
+  if (isIllegalVectorType(Ty))
+    return coerceIllegalVector(Ty);
 
   // _Float16 and __fp16 get passed as if it were an int or float, but with
   // the top 16 bits unspecified. This is not done for OpenCL as it handles the
   // half type natively, and does not need to interwork with AAPCS code.
   if ((Ty->isFloat16Type() || Ty->isHalfType()) &&
       !getContext().getLangOpts().NativeHalfArgsAndReturns) {
     llvm::Type *ResType = IsEffectivelyAAPCS_VFP ?
       llvm::Type::getFloatTy(getVMContext()) :
ABIArgInfo ARMABIInfo::classifyArgumentType(QualType Ty,
   if (isEmptyRecord(getContext(), Ty, true))
     return ABIArgInfo::getIgnore();
 
   if (IsEffectivelyAAPCS_VFP) {
     // Homogeneous Aggregates need to be expanded when we can fit the aggregate
     // into VFP registers.
     const Type *Base = nullptr;
     uint64_t Members = 0;
-    if (isHomogeneousAggregate(Ty, Base, Members)) {
-      assert(Base && "Base class should be set for homogeneous aggregate");
-      // Base can be a floating-point or a vector.
-      return ABIArgInfo::getDirect(nullptr, 0, nullptr, false);
-    }
+    if (isHomogeneousAggregate(Ty, Base, Members))
+      return classifyHomogeneousAggregate(Ty, Base, Members);
   } else if (getABIKind() == ARMABIInfo::AAPCS16_VFP) {
     // WatchOS does have homogeneous aggregates. Note that we intentionally use
     // this convention even for a variadic function: the backend will use GPRs
     // if needed.
     const Type *Base = nullptr;
     uint64_t Members = 0;
     if (isHomogeneousAggregate(Ty, Base, Members)) {
       assert(Base && Members <= 4 && "unexpected homogeneous aggregate");
       llvm::Type *Ty =
         llvm::ArrayType::get(CGT.ConvertType(QualType(Base, 0)), Members);
       return ABIArgInfo::getDirect(Ty, 0, nullptr, false);
     }
   }
 
   if (getABIKind() == ARMABIInfo::AAPCS16_VFP &&
       getContext().getTypeSizeInChars(Ty) > CharUnits::fromQuantity(16)) {

 ABIArgInfo ARMABIInfo::classifyReturnType(QualType RetTy,
                                           bool isVariadic) const {
   bool IsEffectivelyAAPCS_VFP =
       (getABIKind() == AAPCS_VFP || getABIKind() == AAPCS16_VFP) && !isVariadic;
 
   if (RetTy->isVoidType())
     return ABIArgInfo::getIgnore();
 
+  if (const VectorType *VT = RetTy->getAs<VectorType>()) {
     // Large vector types should be returned via memory.
-  if (RetTy->isVectorType() && getContext().getTypeSize(RetTy) > 128) {
+    if (getContext().getTypeSize(RetTy) > 128)
       return getNaturalAlignIndirect(RetTy);
+    // FP16 vectors should be converted to integer vectors
+    if (!getTarget().hasLegalHalfType() &&
+        (VT->getElementType()->isFloat16Type() ||
+         VT->getElementType()->isHalfType()))
+      return coerceIllegalVector(RetTy);
   }
 
   // _Float16 and __fp16 get returned as if it were an int or float, but with
   // the top 16 bits unspecified. This is not done for OpenCL as it handles the
   // half type natively, and does not need to interwork with AAPCS code.
   if ((RetTy->isFloat16Type() || RetTy->isHalfType()) &&
       !getContext().getLangOpts().NativeHalfArgsAndReturns) {
     llvm::Type *ResType = IsEffectivelyAAPCS_VFP ?
ABIArgInfo ARMABIInfo::classifyReturnType(QualType RetTy,
 
   if (isEmptyRecord(getContext(), RetTy, true))
     return ABIArgInfo::getIgnore();
 
   // Check for homogeneous aggregates with AAPCS-VFP.
   if (IsEffectivelyAAPCS_VFP) {
     const Type *Base = nullptr;
     uint64_t Members = 0;
-    if (isHomogeneousAggregate(RetTy, Base, Members)) {
-      assert(Base && "Base class should be set for homogeneous aggregate");
-      // Homogeneous Aggregates are returned directly.
-      return ABIArgInfo::getDirect(nullptr, 0, nullptr, false);
-    }
+    if (isHomogeneousAggregate(RetTy, Base, Members))
+      return classifyHomogeneousAggregate(RetTy, Base, Members);
   }
 
   // Aggregates <= 4 bytes are returned in r0; other aggregates
   // are returned indirectly.
   uint64_t Size = getContext().getTypeSize(RetTy);
   if (Size <= 32) {
     // On RenderScript, coerce Aggregates <= 4 bytes to an integer array of
     // same size and alignment.
ABIArgInfo ARMABIInfo::classifyReturnType(QualType RetTy,
   }
 
   return getNaturalAlignIndirect(RetTy);
 }
 
 /// isIllegalVector - check whether Ty is an illegal vector type.
 bool ARMABIInfo::isIllegalVectorType(QualType Ty) const {
   if (const VectorType *VT = Ty->getAs<VectorType> ()) {
+    // On targets that don't support FP16, FP16 is expanded into float, and we
+    // don't want the ABI to depend on whether or not FP16 is supported in
+    // hardware. Thus return false to coerce FP16 vectors into integer vectors.
+    if (!getTarget().hasLegalHalfType() &&
+        (VT->getElementType()->isFloat16Type() ||
+         VT->getElementType()->isHalfType()))
+      return true;
     if (isAndroid()) {
       // Android shipped using Clang 3.1, which supported a slightly different
       // vector ABI. The primary differences were that 3-element vector types
       // were legal, and so were sub 32-bit vectors (i.e. <2 x i8>). This path
       // accepts that legacy behavior for Android only.
       // Check whether VT is legal.
       unsigned NumElements = VT->getNumElements();
       // NumElements should be power of 2 or equal to 3.

test/CodeGen/arm-vfp16-arguments.c

+// RUN: %clang_cc1 -triple armv7a--none-eabi -target-abi aapcs \
+// RUN:   -mfloat-abi soft -target-feature +neon -emit-llvm -o - -O1 %s \
+// RUN:   | FileCheck %s --check-prefix=CHECK-SOFT
+// RUN: %clang_cc1 -triple armv7a--none-eabi -target-abi aapcs \
+// RUN:   -mfloat-abi hard -target-feature +neon -emit-llvm -o - -O1 %s \
+// RUN:   | FileCheck %s --check-prefix=CHECK-HARD
+// RUN: %clang_cc1 -triple armv7a--none-eabi -target-abi aapcs \
+// RUN:   -mfloat-abi hard -target-feature +neon -target-feature +fullfp16 \
+// RUN:   -emit-llvm -o - -O1 %s \
+// RUN:   | FileCheck %s --check-prefix=CHECK-FULL
+
+typedef __attribute__((neon_vector_type(4))) __fp16 float16x4_t;
+typedef __attribute__((neon_vector_type(8))) __fp16 float16x8_t;
+
+typedef struct { float16x4_t x[2]; } hfa_t;
+// CHECK-FULL: %struct.hfa_t = type { [2 x <4 x half>] }
+
+float16x4_t g4;
+float16x8_t g8;
+
+void st4(float16x4_t a) { g4 = a; }
+// CHECK-SOFT: define void @st4(<2 x i32> %a.coerce)
+// CHECK-SOFT: store <2 x i32> %a.coerce, <2 x i32>* bitcast (<4 x half>* @g4 to <2 x i32>*)
+//
+// CHECK-HARD: define arm_aapcs_vfpcc void @st4(<2 x i32> %a.coerce)
+// CHECK-HARD: store <2 x i32> %a.coerce, <2 x i32>* bitcast (<4 x half>* @g4 to <2 x i32>*)
+//
+// CHECK-FULL: define arm_aapcs_vfpcc void @st4(<4 x half> %a)
+// CHECK-FULL: store <4 x half> %a, <4 x half>* @g4
+
+float16x4_t ld4(void) { return g4; }
+// CHECK-SOFT: define <2 x i32> @ld4()
+// CHECK-SOFT: %0 = load <2 x i32>, <2 x i32>* bitcast (<4 x half>* @g4 to <2 x i32>*)
+// CHECK-SOFT: ret <2 x i32> %0
+//
+// CHECK-HARD: define arm_aapcs_vfpcc <2 x i32> @ld4()
+// CHECK-HARD: %0 = load <2 x i32>, <2 x i32>* bitcast (<4 x half>* @g4 to <2 x i32>*)
+// CHECK-HARD: ret <2 x i32> %0
+//
+// CHECK-FULL: define arm_aapcs_vfpcc <4 x half> @ld4()
+// CHECK-FULL: %0 = load <4 x half>, <4 x half>* @g4
+// CHECK-FULL: ret <4 x half> %0
+
+void st8(float16x8_t a) { g8 = a; }
+// CHECK-SOFT: define void @st8(<4 x i32> %a.coerce)
+// CHECK-SOFT: store <4 x i32> %a.coerce, <4 x i32>* bitcast (<8 x half>* @g8 to <4 x i32>*)
+//
+// CHECK-HARD: define arm_aapcs_vfpcc void @st8(<4 x i32> %a.coerce)
+// CHECK-HARD: store <4 x i32> %a.coerce, <4 x i32>* bitcast (<8 x half>* @g8 to <4 x i32>*)
+//
+// CHECK-FULL: define arm_aapcs_vfpcc void @st8(<8 x half> %a)
+// CHECK-FULL: store <8 x half> %a, <8 x half>* @g8
+
+float16x8_t ld8(void) { return g8; }
+// CHECK-SOFT: define <4 x i32> @ld8()
+// CHECK-SOFT: %0 = load <4 x i32>, <4 x i32>* bitcast (<8 x half>* @g8 to <4 x i32>*)
+// CHECK-SOFT: ret <4 x i32> %0
+//
+// CHECK-HARD: define arm_aapcs_vfpcc <4 x i32> @ld8()
+// CHECK-HARD: %0 = load <4 x i32>, <4 x i32>* bitcast (<8 x half>* @g8 to <4 x i32>*)
+// CHECK-HARD: ret <4 x i32> %0
+//
+// CHECK-FULL: define arm_aapcs_vfpcc <8 x half> @ld8()
+// CHECK-FULL: %0 = load <8 x half>, <8 x half>* @g8
+// CHECK-FULL: ret <8 x half> %0
+
+void test_hfa(hfa_t a) {}
+// CHECK-SOFT: define void @test_hfa([2 x i64] %a.coerce)
+// CHECK-HARD: define arm_aapcs_vfpcc void @test_hfa([2 x <2 x i32>] %a.coerce)
+// CHECK-FULL: define arm_aapcs_vfpcc void @test_hfa(%struct.hfa_t %a.coerce)
+
+hfa_t ghfa;
+hfa_t test_ret_hfa(void) { return ghfa; }
+// CHECK-SOFT: define void @test_ret_hfa(%struct.hfa_t* noalias nocapture sret %agg.result)
+// CHECK-HARD: define arm_aapcs_vfpcc [2 x <2 x i32>] @test_ret_hfa()
+// CHECK-FULL: define arm_aapcs_vfpcc %struct.hfa_t @test_ret_hfa()

test/CodeGen/arm_neon_intrinsics.c

 // RUN: %clang_cc1 -triple thumbv7s-apple-darwin -target-abi apcs-gnu\
-// RUN:  -target-cpu swift -fallow-half-arguments-and-returns -ffreestanding \
+// RUN:  -target-cpu swift -fallow-half-arguments-and-returns \
+// RUN:  -target-feature +fullfp16 -ffreestanding \
 // RUN:  -disable-O0-optnone -emit-llvm -o - %s \
 // RUN:  | opt -S -mem2reg | FileCheck %s
 
 #include <arm_neon.h>
 
 // CHECK-LABEL: @test_vaba_s8(
 // CHECK:   [[VABD_V_I_I:%.*]] = call <8 x i8> @llvm.arm.neon.vabds.v8i8(<8 x i8> %b, <8 x i8> %c)
 // CHECK:   [[ADD_I:%.*]] = add <8 x i8> %a, [[VABD_V_I_I]]

 // CHECK:   [[VLD1:%.*]] = call <2 x i64> @llvm.arm.neon.vld1.v2i64.p0i8(i8* [[TMP0]], i32 4)
 // CHECK:   ret <2 x i64> [[VLD1]]
 int64x2_t test_vld1q_s64(int64_t const * a) {
   return vld1q_s64(a);
 }
 
 // CHECK-LABEL: @test_vld1q_f16(
 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
-// CHECK:   [[VLD1:%.*]] = call <8 x i16> @llvm.arm.neon.vld1.v8i16.p0i8(i8* [[TMP0]], i32 2)
-// CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> [[VLD1]] to <8 x half>
-// CHECK:   ret <8 x half> [[TMP1]]
+// CHECK:   [[VLD1:%.*]] = call <8 x half> @llvm.arm.neon.vld1.v8f16.p0i8(i8* [[TMP0]], i32 2)
+// CHECK:   ret <8 x half> [[VLD1]]
 float16x8_t test_vld1q_f16(float16_t const * a) {
   return vld1q_f16(a);
 }
 
 // CHECK-LABEL: @test_vld1q_f32(
 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
 // CHECK:   [[VLD1:%.*]] = call <4 x float> @llvm.arm.neon.vld1.v4f32.p0i8(i8* [[TMP0]], i32 4)
 // CHECK:   ret <4 x float> [[VLD1]]

 // CHECK:   [[VLD1:%.*]] = call <1 x i64> @llvm.arm.neon.vld1.v1i64.p0i8(i8* [[TMP0]], i32 4)
 // CHECK:   ret <1 x i64> [[VLD1]]
 int64x1_t test_vld1_s64(int64_t const * a) {
   return vld1_s64(a);
 }
 
 // CHECK-LABEL: @test_vld1_f16(
 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
-// CHECK:   [[VLD1:%.*]] = call <4 x i16> @llvm.arm.neon.vld1.v4i16.p0i8(i8* [[TMP0]], i32 2)
-// CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VLD1]] to <4 x half>
-// CHECK:   ret <4 x half> [[TMP1]]
+// CHECK:   [[VLD1:%.*]] = call <4 x half> @llvm.arm.neon.vld1.v4f16.p0i8(i8* [[TMP0]], i32 2)
+// CHECK:   ret <4 x half> [[VLD1]]
 float16x4_t test_vld1_f16(float16_t const * a) {
   return vld1_f16(a);
 }
 
 // CHECK-LABEL: @test_vld1_f32(
 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
 // CHECK:   [[VLD1:%.*]] = call <2 x float> @llvm.arm.neon.vld1.v2f32.p0i8(i8* [[TMP0]], i32 4)
 // CHECK:   ret <2 x float> [[VLD1]]

 // CHECK:   [[LANE:%.*]] = shufflevector <2 x i64> [[TMP3]], <2 x i64> [[TMP3]], <2 x i32> zeroinitializer
 // CHECK:   ret <2 x i64> [[LANE]]
 int64x2_t test_vld1q_dup_s64(int64_t const * a) {
   return vld1q_dup_s64(a);
 }
 
 // CHECK-LABEL: @test_vld1q_dup_f16(
 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
-// CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i16*
-// CHECK:   [[TMP2:%.*]] = load i16, i16* [[TMP1]], align 2
-// CHECK:   [[TMP3:%.*]] = insertelement <8 x i16> undef, i16 [[TMP2]], i32 0
-// CHECK:   [[LANE:%.*]] = shufflevector <8 x i16> [[TMP3]], <8 x i16> [[TMP3]], <8 x i32> zeroinitializer
-// CHECK:   [[TMP4:%.*]] = bitcast <8 x i16> [[LANE]] to <8 x half>
-// CHECK:   ret <8 x half> [[TMP4]]
+// CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to half*
+// CHECK:   [[TMP2:%.*]] = load half, half* [[TMP1]], align 2
+// CHECK:   [[TMP3:%.*]] = insertelement <8 x half> undef, half [[TMP2]], i32 0
+// CHECK:   [[LANE:%.*]] = shufflevector <8 x half> [[TMP3]], <8 x half> [[TMP3]], <8 x i32> zeroinitializer
+// CHECK:   ret <8 x half> [[LANE]]
 float16x8_t test_vld1q_dup_f16(float16_t const * a) {
   return vld1q_dup_f16(a);
 }
 
 // CHECK-LABEL: @test_vld1q_dup_f32(
 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to float*
 // CHECK:   [[TMP2:%.*]] = load float, float* [[TMP1]], align 4

 // CHECK:   [[LANE:%.*]] = shufflevector <1 x i64> [[TMP3]], <1 x i64> [[TMP3]], <1 x i32> zeroinitializer
 // CHECK:   ret <1 x i64> [[LANE]]
 int64x1_t test_vld1_dup_s64(int64_t const * a) {
   return vld1_dup_s64(a);
 }
 
 // CHECK-LABEL: @test_vld1_dup_f16(
 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
-// CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i16*
-// CHECK:   [[TMP2:%.*]] = load i16, i16* [[TMP1]], align 2
-// CHECK:   [[TMP3:%.*]] = insertelement <4 x i16> undef, i16 [[TMP2]], i32 0
-// CHECK:   [[LANE:%.*]] = shufflevector <4 x i16> [[TMP3]], <4 x i16> [[TMP3]], <4 x i32> zeroinitializer
-// CHECK:   [[TMP4:%.*]] = bitcast <4 x i16> [[LANE]] to <4 x half>
-// CHECK:   ret <4 x half> [[TMP4]]
+// CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to half*
+// CHECK:   [[TMP2:%.*]] = load half, half* [[TMP1]], align 2
+// CHECK:   [[TMP3:%.*]] = insertelement <4 x half> undef, half [[TMP2]], i32 0
+// CHECK:   [[LANE:%.*]] = shufflevector <4 x half> [[TMP3]], <4 x half> [[TMP3]], <4 x i32> zeroinitializer
+// CHECK:   ret <4 x half> [[LANE]]
 float16x4_t test_vld1_dup_f16(float16_t const * a) {
   return vld1_dup_f16(a);
 }
 
 // CHECK-LABEL: @test_vld1_dup_f32(
 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to float*
 // CHECK:   [[TMP2:%.*]] = load float, float* [[TMP1]], align 4

 // CHECK:   ret <2 x i64> [[VLD1Q_LANE]]
 int64x2_t test_vld1q_lane_s64(int64_t const * a, int64x2_t b) {
   return vld1q_lane_s64(a, b, 1);
 }
 
 // CHECK-LABEL: @test_vld1q_lane_f16(
 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <8 x half> %b to <16 x i8>
-// CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
-// CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i16*
-// CHECK:   [[TMP4:%.*]] = load i16, i16* [[TMP3]], align 2
-// CHECK:   [[VLD1_LANE:%.*]] = insertelement <8 x i16> [[TMP2]], i16 [[TMP4]], i32 7
-// CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[VLD1_LANE]] to <8 x half>
-// CHECK:   ret <8 x half> [[TMP5]]
+// CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x half>
+// CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to half*
+// CHECK:   [[TMP4:%.*]] = load half, half* [[TMP3]], align 2
+// CHECK:   [[VLD1_LANE:%.*]] = insertelement <8 x half> [[TMP2]], half [[TMP4]], i32 7
+// CHECK:   ret <8 x half> [[VLD1_LANE]]
 float16x8_t test_vld1q_lane_f16(float16_t const * a, float16x8_t b) {
   return vld1q_lane_f16(a, b, 7);
 }
 
 // CHECK-LABEL: @test_vld1q_lane_f32(
 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>

 // CHECK:   ret <1 x i64> [[VLD1_LANE]]
 int64x1_t test_vld1_lane_s64(int64_t const * a, int64x1_t b) {
   return vld1_lane_s64(a, b, 0);
 }
 
 // CHECK-LABEL: @test_vld1_lane_f16(
 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <4 x half> %b to <8 x i8>
-// CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
-// CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i16*
-// CHECK:   [[TMP4:%.*]] = load i16, i16* [[TMP3]], align 2
-// CHECK:   [[VLD1_LANE:%.*]] = insertelement <4 x i16> [[TMP2]], i16 [[TMP4]], i32 3
-// CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[VLD1_LANE]] to <4 x half>
-// CHECK:   ret <4 x half> [[TMP5]]
+// CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x half>
+// CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to half*
+// CHECK:   [[TMP4:%.*]] = load half, half* [[TMP3]], align 2
+// CHECK:   [[VLD1_LANE:%.*]] = insertelement <4 x half> [[TMP2]], half [[TMP4]], i32 3
+// CHECK:   ret <4 x half> [[VLD1_LANE]]
 float16x4_t test_vld1_lane_f16(float16_t const * a, float16x4_t b) {
   return vld1_lane_f16(a, b, 3);
 }
 
 // CHECK-LABEL: @test_vld1_lane_f32(
 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>

 int32x4x2_t test_vld2q_s32(int32_t const * a) {
   return vld2q_s32(a);
 }
 
 // CHECK-LABEL: @test_vld2q_f16(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float16x8x2_t, align 16
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x8x2_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
-// CHECK:   [[VLD2Q_V:%.*]] = call { <8 x i16>, <8 x i16>
+// CHECK:   [[VLD2Q_V:%.*]] = call { <8 x half>, <8 x half>
 float16x8x2_t test_vld2q_f16(float16_t const * a) {
   return vld2q_f16(a);
 }
 
 // CHECK-LABEL: @test_vld2q_f32(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x4x2_t, align 16
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x4x2_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*

 int64x1x2_t test_vld2_s64(int64_t const * a) {
   return vld2_s64(a);
 }
 
 // CHECK-LABEL: @test_vld2_f16(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float16x4x2_t, align 8
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x4x2_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
-// CHECK:   [[VLD2_V:%.*]] = call { <4 x i16>, <4 x i16>
+// CHECK:   [[VLD2_V:%.*]] = call { <4 x half>, <4 x half>
 float16x4x2_t test_vld2_f16(float16_t const * a) {
   return vld2_f16(a);
 }
 
 // CHECK-LABEL: @test_vld2_f32(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x2x2_t, align 8
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x2_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*

 // CHECK:   [[VAL:%.*]] = getelementptr inbounds %struct.float16x8x2_t, %struct.float16x8x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX:%.*]] = getelementptr inbounds [2 x <8 x half>], [2 x <8 x half>]* [[VAL]], i32 0, i32 0
 // CHECK:   [[TMP5:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX]], align 16
 // CHECK:   [[TMP6:%.*]] = bitcast <8 x half> [[TMP5]] to <16 x i8>
 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x8x2_t, %struct.float16x8x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [2 x <8 x half>], [2 x <8 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP7:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX2]], align 16
 // CHECK:   [[TMP8:%.*]] = bitcast <8 x half> [[TMP7]] to <16 x i8>
-// CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
-// CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
-// CHECK:   [[VLD2Q_LANE_V:%.*]] = call { <8 x i16>, <8 x i16>
+// CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x half>
+// CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x half>
+// CHECK:   [[VLD2Q_LANE_V:%.*]] = call { <8 x half>, <8 x half>
 float16x8x2_t test_vld2q_lane_f16(float16_t const * a, float16x8x2_t b) {
   return vld2q_lane_f16(a, b, 7);
 }
 
 // CHECK-LABEL: @test_vld2q_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x4x2_t, align 16
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x4x2_t, align 16
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x4x2_t, align 16

 // CHECK:   [[VAL:%.*]] = getelementptr inbounds %struct.float16x4x2_t, %struct.float16x4x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX:%.*]] = getelementptr inbounds [2 x <4 x half>], [2 x <4 x half>]* [[VAL]], i32 0, i32 0
 // CHECK:   [[TMP5:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX]], align 8
 // CHECK:   [[TMP6:%.*]] = bitcast <4 x half> [[TMP5]] to <8 x i8>
 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x4x2_t, %struct.float16x4x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [2 x <4 x half>], [2 x <4 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP7:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX2]], align 8
 // CHECK:   [[TMP8:%.*]] = bitcast <4 x half> [[TMP7]] to <8 x i8>
-// CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
-// CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
-// CHECK:   [[VLD2_LANE_V:%.*]] = call { <4 x i16>, <4 x i16>
+// CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x half>
+// CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x half>
+// CHECK:   [[VLD2_LANE_V:%.*]] = call { <4 x half>, <4 x half>
 float16x4x2_t test_vld2_lane_f16(float16_t const * a, float16x4x2_t b) {
   return vld2_lane_f16(a, b, 3);
 }
 
 // CHECK-LABEL: @test_vld2_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x2x2_t, align 8
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x2x2_t, align 8
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x2x2_t, align 8

 int32x4x3_t test_vld3q_s32(int32_t const * a) {
   return vld3q_s32(a);
 }
 
 // CHECK-LABEL: @test_vld3q_f16(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float16x8x3_t, align 16
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x8x3_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
-// CHECK:   [[VLD3Q_V:%.*]] = call { <8 x i16>, <8 x i16>, <8 x i16>
+// CHECK:   [[VLD3Q_V:%.*]] = call { <8 x half>, <8 x half>, <8 x half>
 float16x8x3_t test_vld3q_f16(float16_t const * a) {
   return vld3q_f16(a);
 }
 
 // CHECK-LABEL: @test_vld3q_f32(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x4x3_t, align 16
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x4x3_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*

 int64x1x3_t test_vld3_s64(int64_t const * a) {
   return vld3_s64(a);
 }
 
 // CHECK-LABEL: @test_vld3_f16(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float16x4x3_t, align 8
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x4x3_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
-// CHECK:   [[VLD3_V:%.*]] = call { <4 x i16>, <4 x i16>, <4 x i16>
+// CHECK:   [[VLD3_V:%.*]] = call { <4 x half>, <4 x half>, <4 x half>
 float16x4x3_t test_vld3_f16(float16_t const * a) {
   return vld3_f16(a);
 }
 
 // CHECK-LABEL: @test_vld3_f32(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x2x3_t, align 8
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x3_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*

 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x8x3_t, %struct.float16x8x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [3 x <8 x half>], [3 x <8 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP7:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX2]], align 16
 // CHECK:   [[TMP8:%.*]] = bitcast <8 x half> [[TMP7]] to <16 x i8>
 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x8x3_t, %struct.float16x8x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [3 x <8 x half>], [3 x <8 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP9:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX4]], align 16
 // CHECK:   [[TMP10:%.*]] = bitcast <8 x half> [[TMP9]] to <16 x i8>
-// CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
-// CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
-// CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x i16>
-// CHECK:   [[VLD3Q_LANE_V:%.*]] = call { <8 x i16>, <8 x i16>, <8 x i16>
+// CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x half>
+// CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x half>
+// CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x half>
+// CHECK:   [[VLD3Q_LANE_V:%.*]] = call { <8 x half>, <8 x half>, <8 x half>
 float16x8x3_t test_vld3q_lane_f16(float16_t const * a, float16x8x3_t b) {
   return vld3q_lane_f16(a, b, 7);
 }
 
 // CHECK-LABEL: @test_vld3q_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x4x3_t, align 16
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x4x3_t, align 16
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x4x3_t, align 16

 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x4x3_t, %struct.float16x4x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [3 x <4 x half>], [3 x <4 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP7:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX2]], align 8
 // CHECK:   [[TMP8:%.*]] = bitcast <4 x half> [[TMP7]] to <8 x i8>
 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x4x3_t, %struct.float16x4x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [3 x <4 x half>], [3 x <4 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP9:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX4]], align 8
 // CHECK:   [[TMP10:%.*]] = bitcast <4 x half> [[TMP9]] to <8 x i8>
-// CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
-// CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
-// CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x i16>
-// CHECK:   [[VLD3_LANE_V:%.*]] = call { <4 x i16>, <4 x i16>, <4 x i16>
+// CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x half>
+// CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x half>
+// CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x half>
+// CHECK:   [[VLD3_LANE_V:%.*]] = call { <4 x half>, <4 x half>, <4 x half>
 float16x4x3_t test_vld3_lane_f16(float16_t const * a, float16x4x3_t b) {
   return vld3_lane_f16(a, b, 3);
 }
 
 // CHECK-LABEL: @test_vld3_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x2x3_t, align 8
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x2x3_t, align 8
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x2x3_t, align 8

 int32x4x4_t test_vld4q_s32(int32_t const * a) {
   return vld4q_s32(a);
 }
 
 // CHECK-LABEL: @test_vld4q_f16(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float16x8x4_t, align 16
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x8x4_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
-// CHECK:   [[VLD4Q_V:%.*]] = call { <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16>
+// CHECK:   [[VLD4Q_V:%.*]] = call { <8 x half>, <8 x half>, <8 x half>, <8 x half>
 float16x8x4_t test_vld4q_f16(float16_t const * a) {
   return vld4q_f16(a);
 }
 
 // CHECK-LABEL: @test_vld4q_f32(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x4x4_t, align 16
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x4x4_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*

 int64x1x4_t test_vld4_s64(int64_t const * a) {
   return vld4_s64(a);
 }
 
 // CHECK-LABEL: @test_vld4_f16(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float16x4x4_t, align 8
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x4x4_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
-// CHECK:   [[VLD4_V:%.*]] = call { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16>
+// CHECK:   [[VLD4_V:%.*]] = call { <4 x half>, <4 x half>, <4 x half>, <4 x half>
 float16x4x4_t test_vld4_f16(float16_t const * a) {
   return vld4_f16(a);
 }
 
 // CHECK-LABEL: @test_vld4_f32(
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x2x4_t, align 8
 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x4_t* [[__RET]] to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*

 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x8x4_t, %struct.float16x8x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [4 x <8 x half>], [4 x <8 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP9:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX4]], align 16
 // CHECK:   [[TMP10:%.*]] = bitcast <8 x half> [[TMP9]] to <16 x i8>
 // CHECK:   [[VAL5:%.*]] = getelementptr inbounds %struct.float16x8x4_t, %struct.float16x8x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX6:%.*]] = getelementptr inbounds [4 x <8 x half>], [4 x <8 x half>]* [[VAL5]], i32 0, i32 3
 // CHECK:   [[TMP11:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX6]], align 16
 // CHECK:   [[TMP12:%.*]] = bitcast <8 x half> [[TMP11]] to <16 x i8>
-// CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
-// CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
-// CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x i16>
-// CHECK:   [[TMP16:%.*]] = bitcast <16 x i8> [[TMP12]] to <8 x i16>
-// CHECK:   [[VLD4Q_LANE_V:%.*]] = call { <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16>
+// CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x half>
+// CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x half>
+// CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x half>
+// CHECK:   [[TMP16:%.*]] = bitcast <16 x i8> [[TMP12]] to <8 x half>
+// CHECK:   [[VLD4Q_LANE_V:%.*]] = call { <8 x half>, <8 x half>, <8 x half>, <8 x half>
 float16x8x4_t test_vld4q_lane_f16(float16_t const * a, float16x8x4_t b) {
   return vld4q_lane_f16(a, b, 7);
 }
 
 // CHECK-LABEL: @test_vld4q_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x4x4_t, align 16
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x4x4_t, align 16
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x4x4_t, align 16

 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x4x4_t, %struct.float16x4x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [4 x <4 x half>], [4 x <4 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP9:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX4]], align 8
 // CHECK:   [[TMP10:%.*]] = bitcast <4 x half> [[TMP9]] to <8 x i8>
 // CHECK:   [[VAL5:%.*]] = getelementptr inbounds %struct.float16x4x4_t, %struct.float16x4x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX6:%.*]] = getelementptr inbounds [4 x <4 x half>], [4 x <4 x half>]* [[VAL5]], i32 0, i32 3
 // CHECK:   [[TMP11:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX6]], align 8
 // CHECK:   [[TMP12:%.*]] = bitcast <4 x half> [[TMP11]] to <8 x i8>
-// CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
-// CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
-// CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x i16>
-// CHECK:   [[TMP16:%.*]] = bitcast <8 x i8> [[TMP12]] to <4 x i16>
-// CHECK:   [[VLD4_LANE_V:%.*]] = call { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16>
+// CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x half>
+// CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x half>
+// CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x half>
+// CHECK:   [[TMP16:%.*]] = bitcast <8 x i8> [[TMP12]] to <4 x half>
+// CHECK:   [[VLD4_LANE_V:%.*]] = call { <4 x half>, <4 x half>, <4 x half>, <4 x half>
 float16x4x4_t test_vld4_lane_f16(float16_t const * a, float16x4x4_t b) {
   return vld4_lane_f16(a, b, 3);
 }
 
 // CHECK-LABEL: @test_vld4_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x2x4_t, align 8
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x2x4_t, align 8
 // CHECK:   [[__RET:%.*]] = alloca %struct.float32x2x4_t, align 8

 // CHECK:   ret void
 void test_vst1q_s64(int64_t * a, int64x2_t b) {
   vst1q_s64(a, b);
 }
 
 // CHECK-LABEL: @test_vst1q_f16(
 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <8 x half> %b to <16 x i8>
-// CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
-// CHECK:   call void @llvm.arm.neon.vst1.p0i8.v8i16(i8* [[TMP0]], <8 x i16> [[TMP2]], i32 2)
+// CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x half>
+// CHECK:   call void @llvm.arm.neon.vst1.p0i8.v8f16(i8* [[TMP0]], <8 x half> [[TMP2]], i32 2)
 // CHECK:   ret void
 void test_vst1q_f16(float16_t * a, float16x8_t b) {
   vst1q_f16(a, b);
 }
 
 // CHECK-LABEL: @test_vst1q_f32(
 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>

 // CHECK:   ret void
 void test_vst1_s64(int64_t * a, int64x1_t b) {
   vst1_s64(a, b);
 }
 
 // CHECK-LABEL: @test_vst1_f16(
 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <4 x half> %b to <8 x i8>
-// CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
-// CHECK:   call void @llvm.arm.neon.vst1.p0i8.v4i16(i8* [[TMP0]], <4 x i16> [[TMP2]], i32 2)
+// CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x half>
+// CHECK:   call void @llvm.arm.neon.vst1.p0i8.v4f16(i8* [[TMP0]], <4 x half> [[TMP2]], i32 2)
 // CHECK:   ret void
 void test_vst1_f16(float16_t * a, float16x4_t b) {
   vst1_f16(a, b);
 }
 
 // CHECK-LABEL: @test_vst1_f32(
 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>

 // CHECK:   ret void
 void test_vst1q_lane_s64(int64_t * a, int64x2_t b) {
   vst1q_lane_s64(a, b, 1);
 }
 
 // CHECK-LABEL: @test_vst1q_lane_f16(
 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <8 x half> %b to <16 x i8>
-// CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
-// CHECK:   [[TMP3:%.*]] = extractelement <8 x i16> [[TMP2]], i32 7
-// CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i16*
-// CHECK:   store i16 [[TMP3]], i16* [[TMP4]], align 2
+// CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x half>
+// CHECK:   [[TMP3:%.*]] = extractelement <8 x half> [[TMP2]], i32 7
+// CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to half*
+// CHECK:   store half [[TMP3]], half* [[TMP4]], align 2
 // CHECK:   ret void
 void test_vst1q_lane_f16(float16_t * a, float16x8_t b) {
   vst1q_lane_f16(a, b, 7);
 }
 
 // CHECK-LABEL: @test_vst1q_lane_f32(
 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>

 // CHECK:   ret void
 void test_vst1_lane_s64(int64_t * a, int64x1_t b) {
   vst1_lane_s64(a, b, 0);
 }
 
 // CHECK-LABEL: @test_vst1_lane_f16(
 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <4 x half> %b to <8 x i8>
-// CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
-// CHECK:   [[TMP3:%.*]] = extractelement <4 x i16> [[TMP2]], i32 3
-// CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i16*
-// CHECK:   store i16 [[TMP3]], i16* [[TMP4]], align 2
+// CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x half>
+// CHECK:   [[TMP3:%.*]] = extractelement <4 x half> [[TMP2]], i32 3
+// CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to half*
+// CHECK:   store half [[TMP3]], half* [[TMP4]], align 2
 // CHECK:   ret void
 void test_vst1_lane_f16(float16_t * a, float16x4_t b) {
   vst1_lane_f16(a, b, 3);
 }
 
 // CHECK-LABEL: @test_vst1_lane_f32(
 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>

 // CHECK:   [[VAL:%.*]] = getelementptr inbounds %struct.float16x8x2_t, %struct.float16x8x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX:%.*]] = getelementptr inbounds [2 x <8 x half>], [2 x <8 x half>]* [[VAL]], i32 0, i32 0
 // CHECK:   [[TMP4:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX]], align 16
 // CHECK:   [[TMP5:%.*]] = bitcast <8 x half> [[TMP4]] to <16 x i8>
 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x8x2_t, %struct.float16x8x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [2 x <8 x half>], [2 x <8 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP6:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX2]], align 16
 // CHECK:   [[TMP7:%.*]] = bitcast <8 x half> [[TMP6]] to <16 x i8>
-// CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
-// CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
-// CHECK:   call void @llvm.arm.neon.vst2.p0i8.v8i16(i8* [[TMP3]], <8 x i16> [[TMP8]], <8 x i16> [[TMP9]], i32 2)
+// CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>
+// CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x half>
+// CHECK:   call void @llvm.arm.neon.vst2.p0i8.v8f16(i8* [[TMP3]], <8 x half> [[TMP8]], <8 x half> [[TMP9]], i32 2)
 // CHECK:   ret void
 void test_vst2q_f16(float16_t * a, float16x8x2_t b) {
   vst2q_f16(a, b);
 }
 
 // CHECK-LABEL: @test_vst2q_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x4x2_t, align 16
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x4x2_t, align 16

 // CHECK:   [[VAL:%.*]] = getelementptr inbounds %struct.float16x4x2_t, %struct.float16x4x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX:%.*]] = getelementptr inbounds [2 x <4 x half>], [2 x <4 x half>]* [[VAL]], i32 0, i32 0
 // CHECK:   [[TMP4:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX]], align 8
 // CHECK:   [[TMP5:%.*]] = bitcast <4 x half> [[TMP4]] to <8 x i8>
 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x4x2_t, %struct.float16x4x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [2 x <4 x half>], [2 x <4 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP6:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX2]], align 8
 // CHECK:   [[TMP7:%.*]] = bitcast <4 x half> [[TMP6]] to <8 x i8>
-// CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
-// CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
-// CHECK:   call void @llvm.arm.neon.vst2.p0i8.v4i16(i8* [[TMP3]], <4 x i16> [[TMP8]], <4 x i16> [[TMP9]], i32 2)
+// CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>
+// CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x half>
+// CHECK:   call void @llvm.arm.neon.vst2.p0i8.v4f16(i8* [[TMP3]], <4 x half> [[TMP8]], <4 x half> [[TMP9]], i32 2)
 // CHECK:   ret void
 void test_vst2_f16(float16_t * a, float16x4x2_t b) {
   vst2_f16(a, b);
 }
 
 // CHECK-LABEL: @test_vst2_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x2x2_t, align 8
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x2x2_t, align 8

 // CHECK:   [[VAL:%.*]] = getelementptr inbounds %struct.float16x8x2_t, %struct.float16x8x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX:%.*]] = getelementptr inbounds [2 x <8 x half>], [2 x <8 x half>]* [[VAL]], i32 0, i32 0
 // CHECK:   [[TMP4:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX]], align 16
 // CHECK:   [[TMP5:%.*]] = bitcast <8 x half> [[TMP4]] to <16 x i8>
 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x8x2_t, %struct.float16x8x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [2 x <8 x half>], [2 x <8 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP6:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX2]], align 16
 // CHECK:   [[TMP7:%.*]] = bitcast <8 x half> [[TMP6]] to <16 x i8>
-// CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
-// CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
-// CHECK:   call void @llvm.arm.neon.vst2lane.p0i8.v8i16(i8* [[TMP3]], <8 x i16> [[TMP8]], <8 x i16> [[TMP9]], i32 7, i32 2)
+// CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>
+// CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x half>
+// CHECK:   call void @llvm.arm.neon.vst2lane.p0i8.v8f16(i8* [[TMP3]], <8 x half> [[TMP8]], <8 x half> [[TMP9]], i32 7, i32 2)
 // CHECK:   ret void
 void test_vst2q_lane_f16(float16_t * a, float16x8x2_t b) {
   vst2q_lane_f16(a, b, 7);
 }
 
 // CHECK-LABEL: @test_vst2q_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x4x2_t, align 16
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x4x2_t, align 16

 // CHECK:   [[VAL:%.*]] = getelementptr inbounds %struct.float16x4x2_t, %struct.float16x4x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX:%.*]] = getelementptr inbounds [2 x <4 x half>], [2 x <4 x half>]* [[VAL]], i32 0, i32 0
 // CHECK:   [[TMP4:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX]], align 8
 // CHECK:   [[TMP5:%.*]] = bitcast <4 x half> [[TMP4]] to <8 x i8>
 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x4x2_t, %struct.float16x4x2_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [2 x <4 x half>], [2 x <4 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP6:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX2]], align 8
 // CHECK:   [[TMP7:%.*]] = bitcast <4 x half> [[TMP6]] to <8 x i8>
-// CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
-// CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
-// CHECK:   call void @llvm.arm.neon.vst2lane.p0i8.v4i16(i8* [[TMP3]], <4 x i16> [[TMP8]], <4 x i16> [[TMP9]], i32 3, i32 2)
+// CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>
+// CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x half>
+// CHECK:   call void @llvm.arm.neon.vst2lane.p0i8.v4f16(i8* [[TMP3]], <4 x half> [[TMP8]], <4 x half> [[TMP9]], i32 3, i32 2)
 // CHECK:   ret void
 void test_vst2_lane_f16(float16_t * a, float16x4x2_t b) {
   vst2_lane_f16(a, b, 3);
 }
 
 // CHECK-LABEL: @test_vst2_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x2x2_t, align 8
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x2x2_t, align 8

 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x8x3_t, %struct.float16x8x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [3 x <8 x half>], [3 x <8 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP6:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX2]], align 16
 // CHECK:   [[TMP7:%.*]] = bitcast <8 x half> [[TMP6]] to <16 x i8>
 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x8x3_t, %struct.float16x8x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [3 x <8 x half>], [3 x <8 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP8:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX4]], align 16
 // CHECK:   [[TMP9:%.*]] = bitcast <8 x half> [[TMP8]] to <16 x i8>
-// CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
-// CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
-// CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
-// CHECK:   call void @llvm.arm.neon.vst3.p0i8.v8i16(i8* [[TMP3]], <8 x i16> [[TMP10]], <8 x i16> [[TMP11]], <8 x i16> [[TMP12]], i32 2)
+// CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>
+// CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x half>
+// CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x half>
+// CHECK:   call void @llvm.arm.neon.vst3.p0i8.v8f16(i8* [[TMP3]], <8 x half> [[TMP10]], <8 x half> [[TMP11]], <8 x half> [[TMP12]], i32 2)
 // CHECK:   ret void
 void test_vst3q_f16(float16_t * a, float16x8x3_t b) {
   vst3q_f16(a, b);
 }
 
 // CHECK-LABEL: @test_vst3q_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x4x3_t, align 16
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x4x3_t, align 16

 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x4x3_t, %struct.float16x4x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [3 x <4 x half>], [3 x <4 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP6:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX2]], align 8
 // CHECK:   [[TMP7:%.*]] = bitcast <4 x half> [[TMP6]] to <8 x i8>
 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x4x3_t, %struct.float16x4x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [3 x <4 x half>], [3 x <4 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP8:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX4]], align 8
 // CHECK:   [[TMP9:%.*]] = bitcast <4 x half> [[TMP8]] to <8 x i8>
-// CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
-// CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
-// CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
-// CHECK:   call void @llvm.arm.neon.vst3.p0i8.v4i16(i8* [[TMP3]], <4 x i16> [[TMP10]], <4 x i16> [[TMP11]], <4 x i16> [[TMP12]], i32 2)
+// CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>
+// CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x half>
+// CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x half>
+// CHECK:   call void @llvm.arm.neon.vst3.p0i8.v4f16(i8* [[TMP3]], <4 x half> [[TMP10]], <4 x half> [[TMP11]], <4 x half> [[TMP12]], i32 2)
 // CHECK:   ret void
 void test_vst3_f16(float16_t * a, float16x4x3_t b) {
   vst3_f16(a, b);
 }
 
 // CHECK-LABEL: @test_vst3_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x2x3_t, align 8
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x2x3_t, align 8

 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x8x3_t, %struct.float16x8x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [3 x <8 x half>], [3 x <8 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP6:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX2]], align 16
 // CHECK:   [[TMP7:%.*]] = bitcast <8 x half> [[TMP6]] to <16 x i8>
 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x8x3_t, %struct.float16x8x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [3 x <8 x half>], [3 x <8 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP8:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX4]], align 16
 // CHECK:   [[TMP9:%.*]] = bitcast <8 x half> [[TMP8]] to <16 x i8>
-// CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
-// CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
-// CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
-// CHECK:   call void @llvm.arm.neon.vst3lane.p0i8.v8i16(i8* [[TMP3]], <8 x i16> [[TMP10]], <8 x i16> [[TMP11]], <8 x i16> [[TMP12]], i32 7, i32 2)
+// CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>
+// CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x half>
+// CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x half>
+// CHECK:   call void @llvm.arm.neon.vst3lane.p0i8.v8f16(i8* [[TMP3]], <8 x half> [[TMP10]], <8 x half> [[TMP11]], <8 x half> [[TMP12]], i32 7, i32 2)
 // CHECK:   ret void
 void test_vst3q_lane_f16(float16_t * a, float16x8x3_t b) {
   vst3q_lane_f16(a, b, 7);
 }
 
 // CHECK-LABEL: @test_vst3q_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x4x3_t, align 16
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x4x3_t, align 16

 // CHECK:   [[VAL1:%.*]] = getelementptr inbounds %struct.float16x4x3_t, %struct.float16x4x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX2:%.*]] = getelementptr inbounds [3 x <4 x half>], [3 x <4 x half>]* [[VAL1]], i32 0, i32 1
 // CHECK:   [[TMP6:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX2]], align 8
 // CHECK:   [[TMP7:%.*]] = bitcast <4 x half> [[TMP6]] to <8 x i8>
 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x4x3_t, %struct.float16x4x3_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [3 x <4 x half>], [3 x <4 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP8:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX4]], align 8
 // CHECK:   [[TMP9:%.*]] = bitcast <4 x half> [[TMP8]] to <8 x i8>
-// CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
-// CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
-// CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
-// CHECK:   call void @llvm.arm.neon.vst3lane.p0i8.v4i16(i8* [[TMP3]], <4 x i16> [[TMP10]], <4 x i16> [[TMP11]], <4 x i16> [[TMP12]], i32 3, i32 2)
+// CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>
+// CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x half>
+// CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x half>
+// CHECK:   call void @llvm.arm.neon.vst3lane.p0i8.v4f16(i8* [[TMP3]], <4 x half> [[TMP10]], <4 x half> [[TMP11]], <4 x half> [[TMP12]], i32 3, i32 2)
 // CHECK:   ret void
 void test_vst3_lane_f16(float16_t * a, float16x4x3_t b) {
   vst3_lane_f16(a, b, 3);
 }
 
 // CHECK-LABEL: @test_vst3_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x2x3_t, align 8
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x2x3_t, align 8

 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x8x4_t, %struct.float16x8x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [4 x <8 x half>], [4 x <8 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP8:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX4]], align 16
 // CHECK:   [[TMP9:%.*]] = bitcast <8 x half> [[TMP8]] to <16 x i8>
 // CHECK:   [[VAL5:%.*]] = getelementptr inbounds %struct.float16x8x4_t, %struct.float16x8x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX6:%.*]] = getelementptr inbounds [4 x <8 x half>], [4 x <8 x half>]* [[VAL5]], i32 0, i32 3
 // CHECK:   [[TMP10:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX6]], align 16
 // CHECK:   [[TMP11:%.*]] = bitcast <8 x half> [[TMP10]] to <16 x i8>
-// CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
-// CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
-// CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
-// CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x i16>
-// CHECK:   call void @llvm.arm.neon.vst4.p0i8.v8i16(i8* [[TMP3]], <8 x i16> [[TMP12]], <8 x i16> [[TMP13]], <8 x i16> [[TMP14]], <8 x i16> [[TMP15]], i32 2)
+// CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>
+// CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x half>
+// CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x half>
+// CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x half>
+// CHECK:   call void @llvm.arm.neon.vst4.p0i8.v8f16(i8* [[TMP3]], <8 x half> [[TMP12]], <8 x half> [[TMP13]], <8 x half> [[TMP14]], <8 x half> [[TMP15]], i32 2)
 // CHECK:   ret void
 void test_vst4q_f16(float16_t * a, float16x8x4_t b) {
   vst4q_f16(a, b);
 }
 
 // CHECK-LABEL: @test_vst4q_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x4x4_t, align 16
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x4x4_t, align 16

 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x4x4_t, %struct.float16x4x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [4 x <4 x half>], [4 x <4 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP8:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX4]], align 8
 // CHECK:   [[TMP9:%.*]] = bitcast <4 x half> [[TMP8]] to <8 x i8>
 // CHECK:   [[VAL5:%.*]] = getelementptr inbounds %struct.float16x4x4_t, %struct.float16x4x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX6:%.*]] = getelementptr inbounds [4 x <4 x half>], [4 x <4 x half>]* [[VAL5]], i32 0, i32 3
 // CHECK:   [[TMP10:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX6]], align 8
 // CHECK:   [[TMP11:%.*]] = bitcast <4 x half> [[TMP10]] to <8 x i8>
-// CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
-// CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
-// CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
-// CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x i16>
-// CHECK:   call void @llvm.arm.neon.vst4.p0i8.v4i16(i8* [[TMP3]], <4 x i16> [[TMP12]], <4 x i16> [[TMP13]], <4 x i16> [[TMP14]], <4 x i16> [[TMP15]], i32 2)
+// CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>
+// CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x half>
+// CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x half>
+// CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x half>
+// CHECK:   call void @llvm.arm.neon.vst4.p0i8.v4f16(i8* [[TMP3]], <4 x half> [[TMP12]], <4 x half> [[TMP13]], <4 x half> [[TMP14]], <4 x half> [[TMP15]], i32 2)
 // CHECK:   ret void
 void test_vst4_f16(float16_t * a, float16x4x4_t b) {
   vst4_f16(a, b);
 }
 
 // CHECK-LABEL: @test_vst4_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x2x4_t, align 8
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x2x4_t, align 8

 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x8x4_t, %struct.float16x8x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [4 x <8 x half>], [4 x <8 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP8:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX4]], align 16
 // CHECK:   [[TMP9:%.*]] = bitcast <8 x half> [[TMP8]] to <16 x i8>
 // CHECK:   [[VAL5:%.*]] = getelementptr inbounds %struct.float16x8x4_t, %struct.float16x8x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX6:%.*]] = getelementptr inbounds [4 x <8 x half>], [4 x <8 x half>]* [[VAL5]], i32 0, i32 3
 // CHECK:   [[TMP10:%.*]] = load <8 x half>, <8 x half>* [[ARRAYIDX6]], align 16
 // CHECK:   [[TMP11:%.*]] = bitcast <8 x half> [[TMP10]] to <16 x i8>
-// CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
-// CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
-// CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
-// CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x i16>
-// CHECK:   call void @llvm.arm.neon.vst4lane.p0i8.v8i16(i8* [[TMP3]], <8 x i16> [[TMP12]], <8 x i16> [[TMP13]], <8 x i16> [[TMP14]], <8 x i16> [[TMP15]], i32 7, i32 2)
+// CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x half>
+// CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x half>
+// CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x half>
+// CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x half>
+// CHECK:   call void @llvm.arm.neon.vst4lane.p0i8.v8f16(i8* [[TMP3]], <8 x half> [[TMP12]], <8 x half> [[TMP13]], <8 x half> [[TMP14]], <8 x half> [[TMP15]], i32 7, i32 2)
 // CHECK:   ret void
 void test_vst4q_lane_f16(float16_t * a, float16x8x4_t b) {
   vst4q_lane_f16(a, b, 7);
 }
 
 // CHECK-LABEL: @test_vst4q_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x4x4_t, align 16
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x4x4_t, align 16

 // CHECK:   [[VAL3:%.*]] = getelementptr inbounds %struct.float16x4x4_t, %struct.float16x4x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX4:%.*]] = getelementptr inbounds [4 x <4 x half>], [4 x <4 x half>]* [[VAL3]], i32 0, i32 2
 // CHECK:   [[TMP8:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX4]], align 8
 // CHECK:   [[TMP9:%.*]] = bitcast <4 x half> [[TMP8]] to <8 x i8>
 // CHECK:   [[VAL5:%.*]] = getelementptr inbounds %struct.float16x4x4_t, %struct.float16x4x4_t* [[__S1]], i32 0, i32 0
 // CHECK:   [[ARRAYIDX6:%.*]] = getelementptr inbounds [4 x <4 x half>], [4 x <4 x half>]* [[VAL5]], i32 0, i32 3
 // CHECK:   [[TMP10:%.*]] = load <4 x half>, <4 x half>* [[ARRAYIDX6]], align 8
 // CHECK:   [[TMP11:%.*]] = bitcast <4 x half> [[TMP10]] to <8 x i8>
-// CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
-// CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
-// CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
-// CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x i16>
-// CHECK:   call void @llvm.arm.neon.vst4lane.p0i8.v4i16(i8* [[TMP3]], <4 x i16> [[TMP12]], <4 x i16> [[TMP13]], <4 x i16> [[TMP14]], <4 x i16> [[TMP15]], i32 3, i32 2)
+// CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x half>
+// CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x half>
+// CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x half>
+// CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x half>
+// CHECK:   call void @llvm.arm.neon.vst4lane.p0i8.v4f16(i8* [[TMP3]], <4 x half> [[TMP12]], <4 x half> [[TMP13]], <4 x half> [[TMP14]], <4 x half> [[TMP15]], i32 3, i32 2)
 // CHECK:   ret void
 void test_vst4_lane_f16(float16_t * a, float16x4x4_t b) {
   vst4_lane_f16(a, b, 3);
 }
 
 // CHECK-LABEL: @test_vst4_lane_f32(
 // CHECK:   [[B:%.*]] = alloca %struct.float32x2x4_t, align 8
 // CHECK:   [[__S1:%.*]] = alloca %struct.float32x2x4_t, align 8