[AArch64] ADD ARMv.2-A FP16 vector intrinsics

Differential Revision: https://reviews.llvm.org/D34161

llvm-svn: 305820

4 files changed, 189 insertions, 7 deletions

diff --git a/clang/lib/Basic/Targets.cpp b/clang/lib/Basic/Targets.cpp
index a3b8330707b..e23a93e8ced 100644
--- a/clang/lib/Basic/Targets.cpp
+++ b/clang/lib/Basic/Targets.cpp
@@ -6172,6 +6172,8 @@ class AArch64TargetInfo : public TargetInfo {
   unsigned Crypto;
   unsigned Unaligned;
   unsigned V8_1A;
+  unsigned V8_2A;
+  unsigned HasFullFP16;
 
   static const Builtin::Info BuiltinInfo[];
 
@@ -6303,6 +6305,8 @@ public:
 
     if (V8_1A)
       Builder.defineMacro("__ARM_FEATURE_QRDMX", "1");
+    if (V8_2A && FPU == NeonMode && HasFullFP16)
+      Builder.defineMacro("__ARM_FEATURE_FP16_VECTOR_ARITHMETIC", "1");
 
     // All of the __sync_(bool|val)_compare_and_swap_(1|2|4|8) builtins work.
     Builder.defineMacro("__GCC_HAVE_SYNC_COMPARE_AND_SWAP_1");
@@ -6330,6 +6334,8 @@ public:
     Crypto = 0;
     Unaligned = 1;
     V8_1A = 0;
+    V8_2A = 0;
+    HasFullFP16 = 0;
 
     for (const auto &Feature : Features) {
       if (Feature == "+neon")
@@ -6342,6 +6348,10 @@ public:
         Unaligned = 0;
       if (Feature == "+v8.1a")
         V8_1A = 1;
+      if (Feature == "+v8.2a")
+        V8_2A = 1;
+      if (Feature == "+fullfp16")
+        HasFullFP16 = 1;
     }
 
     setDataLayout();
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 8f0c22d1f7e..a6451b7fc3c 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -2956,8 +2956,9 @@ static llvm::VectorType *GetNeonType(CodeGenFunction *CGF,
     return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad));
   case NeonTypeFlags::Int16:
   case NeonTypeFlags::Poly16:
-  case NeonTypeFlags::Float16:
     return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
+  case NeonTypeFlags::Float16:
+    return llvm::VectorType::get(CGF->HalfTy, V1Ty ? 1 : (4 << IsQuad));
   case NeonTypeFlags::Int32:
     return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad));
   case NeonTypeFlags::Int64:
@@ -2980,6 +2981,8 @@ static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF,
                                           NeonTypeFlags IntTypeFlags) {
   int IsQuad = IntTypeFlags.isQuad();
   switch (IntTypeFlags.getEltType()) {
+  case NeonTypeFlags::Int16:
+    return llvm::VectorType::get(CGF->HalfTy, (4 << IsQuad));
   case NeonTypeFlags::Int32:
     return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad));
   case NeonTypeFlags::Int64:
@@ -3127,55 +3130,80 @@ static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = {
   NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0),
   NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0),
   NEONMAP0(vcvt_f32_v),
+  NEONMAP2(vcvt_n_f16_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
   NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
+  NEONMAP1(vcvt_n_s16_v, arm_neon_vcvtfp2fxs, 0),
   NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0),
   NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0),
+  NEONMAP1(vcvt_n_u16_v, arm_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0),
+  NEONMAP0(vcvt_s16_v),
   NEONMAP0(vcvt_s32_v),
   NEONMAP0(vcvt_s64_v),
+  NEONMAP0(vcvt_u16_v),
   NEONMAP0(vcvt_u32_v),
   NEONMAP0(vcvt_u64_v),
+  NEONMAP1(vcvta_s16_v, arm_neon_vcvtas, 0),
   NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0),
   NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0),
   NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0),
   NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0),
+  NEONMAP1(vcvtaq_s16_v, arm_neon_vcvtas, 0),
   NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0),
   NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0),
+  NEONMAP1(vcvtaq_u16_v, arm_neon_vcvtau, 0),
   NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0),
   NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0),
+  NEONMAP1(vcvtm_s16_v, arm_neon_vcvtms, 0),
   NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0),
   NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0),
+  NEONMAP1(vcvtm_u16_v, arm_neon_vcvtmu, 0),
   NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0),
   NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0),
+  NEONMAP1(vcvtmq_s16_v, arm_neon_vcvtms, 0),
   NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0),
   NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0),
+  NEONMAP1(vcvtmq_u16_v, arm_neon_vcvtmu, 0),
   NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0),
   NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0),
+  NEONMAP1(vcvtn_s16_v, arm_neon_vcvtns, 0),
   NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0),
   NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0),
+  NEONMAP1(vcvtn_u16_v, arm_neon_vcvtnu, 0),
   NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0),
   NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0),
+  NEONMAP1(vcvtnq_s16_v, arm_neon_vcvtns, 0),
   NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0),
   NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0),
+  NEONMAP1(vcvtnq_u16_v, arm_neon_vcvtnu, 0),
   NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0),
   NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0),
+  NEONMAP1(vcvtp_s16_v, arm_neon_vcvtps, 0),
   NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0),
   NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0),
+  NEONMAP1(vcvtp_u16_v, arm_neon_vcvtpu, 0),
   NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0),
   NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0),
+  NEONMAP1(vcvtpq_s16_v, arm_neon_vcvtps, 0),
   NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0),
   NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0),
+  NEONMAP1(vcvtpq_u16_v, arm_neon_vcvtpu, 0),
   NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0),
   NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0),
   NEONMAP0(vcvtq_f32_v),
+  NEONMAP2(vcvtq_n_f16_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
   NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
+  NEONMAP1(vcvtq_n_s16_v, arm_neon_vcvtfp2fxs, 0),
   NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0),
   NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0),
+  NEONMAP1(vcvtq_n_u16_v, arm_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0),
+  NEONMAP0(vcvtq_s16_v),
   NEONMAP0(vcvtq_s32_v),
   NEONMAP0(vcvtq_s64_v),
+  NEONMAP0(vcvtq_u16_v),
   NEONMAP0(vcvtq_u32_v),
   NEONMAP0(vcvtq_u64_v),
   NEONMAP0(vext_v),
@@ -3338,19 +3366,27 @@ static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = {
   NEONMAP1(vcnt_v, ctpop, Add1ArgType),
   NEONMAP1(vcntq_v, ctpop, Add1ArgType),
   NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0),
+  NEONMAP0(vcvt_f16_v),
   NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0),
   NEONMAP0(vcvt_f32_v),
+  NEONMAP2(vcvt_n_f16_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
   NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
   NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
+  NEONMAP1(vcvt_n_s16_v, aarch64_neon_vcvtfp2fxs, 0),
   NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
   NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
+  NEONMAP1(vcvt_n_u16_v, aarch64_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
+  NEONMAP0(vcvtq_f16_v),
   NEONMAP0(vcvtq_f32_v),
+  NEONMAP2(vcvtq_n_f16_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
   NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
   NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
+  NEONMAP1(vcvtq_n_s16_v, aarch64_neon_vcvtfp2fxs, 0),
   NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
   NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
+  NEONMAP1(vcvtq_n_u16_v, aarch64_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType),
@@ -3819,9 +3855,20 @@ Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(
   case NEON::BI__builtin_neon_vcageq_v:
   case NEON::BI__builtin_neon_vcagt_v:
   case NEON::BI__builtin_neon_vcagtq_v: {
-    llvm::Type *VecFlt = llvm::VectorType::get(
-        VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy,
-        VTy->getNumElements());
+    llvm::Type *Ty;
+    switch (VTy->getScalarSizeInBits()) {
+    default: llvm_unreachable("unexpected type");
+    case 32:
+      Ty = FloatTy;
+      break;
+    case 64:
+      Ty = DoubleTy;
+      break;
+    case 16:
+      Ty = HalfTy;
+      break;
+    }
+    llvm::Type *VecFlt = llvm::VectorType::get(Ty, VTy->getNumElements());
     llvm::Type *Tys[] = { VTy, VecFlt };
     Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
     return EmitNeonCall(F, Ops, NameHint);
@@ -3838,8 +3885,16 @@ Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(
     Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad));
     return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
                 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
+  case NEON::BI__builtin_neon_vcvt_f16_v:
+  case NEON::BI__builtin_neon_vcvtq_f16_v:
+    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
+    Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float16, false, Quad));
+    return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
+                : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
+  case NEON::BI__builtin_neon_vcvt_n_f16_v:
   case NEON::BI__builtin_neon_vcvt_n_f32_v:
   case NEON::BI__builtin_neon_vcvt_n_f64_v:
+  case NEON::BI__builtin_neon_vcvtq_n_f16_v:
   case NEON::BI__builtin_neon_vcvtq_n_f32_v:
   case NEON::BI__builtin_neon_vcvtq_n_f64_v: {
     llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty };
@@ -3847,11 +3902,15 @@ Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(
     Function *F = CGM.getIntrinsic(Int, Tys);
     return EmitNeonCall(F, Ops, "vcvt_n");
   }
+  case NEON::BI__builtin_neon_vcvt_n_s16_v:
   case NEON::BI__builtin_neon_vcvt_n_s32_v:
+  case NEON::BI__builtin_neon_vcvt_n_u16_v:
   case NEON::BI__builtin_neon_vcvt_n_u32_v:
   case NEON::BI__builtin_neon_vcvt_n_s64_v:
   case NEON::BI__builtin_neon_vcvt_n_u64_v:
+  case NEON::BI__builtin_neon_vcvtq_n_s16_v:
   case NEON::BI__builtin_neon_vcvtq_n_s32_v:
+  case NEON::BI__builtin_neon_vcvtq_n_u16_v:
   case NEON::BI__builtin_neon_vcvtq_n_u32_v:
   case NEON::BI__builtin_neon_vcvtq_n_s64_v:
   case NEON::BI__builtin_neon_vcvtq_n_u64_v: {
@@ -3863,44 +3922,63 @@ Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(
   case NEON::BI__builtin_neon_vcvt_u32_v:
   case NEON::BI__builtin_neon_vcvt_s64_v:
   case NEON::BI__builtin_neon_vcvt_u64_v:
+  case NEON::BI__builtin_neon_vcvt_s16_v:
+  case NEON::BI__builtin_neon_vcvt_u16_v:
   case NEON::BI__builtin_neon_vcvtq_s32_v:
   case NEON::BI__builtin_neon_vcvtq_u32_v:
   case NEON::BI__builtin_neon_vcvtq_s64_v:
-  case NEON::BI__builtin_neon_vcvtq_u64_v: {
+  case NEON::BI__builtin_neon_vcvtq_u64_v:
+  case NEON::BI__builtin_neon_vcvtq_s16_v:
+  case NEON::BI__builtin_neon_vcvtq_u16_v: {
     Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type));
     return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt")
                 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt");
   }
+  case NEON::BI__builtin_neon_vcvta_s16_v:
   case NEON::BI__builtin_neon_vcvta_s32_v:
   case NEON::BI__builtin_neon_vcvta_s64_v:
   case NEON::BI__builtin_neon_vcvta_u32_v:
   case NEON::BI__builtin_neon_vcvta_u64_v:
+  case NEON::BI__builtin_neon_vcvtaq_s16_v:
   case NEON::BI__builtin_neon_vcvtaq_s32_v:
   case NEON::BI__builtin_neon_vcvtaq_s64_v:
+  case NEON::BI__builtin_neon_vcvtaq_u16_v:
   case NEON::BI__builtin_neon_vcvtaq_u32_v:
   case NEON::BI__builtin_neon_vcvtaq_u64_v:
+  case NEON::BI__builtin_neon_vcvtn_s16_v:
   case NEON::BI__builtin_neon_vcvtn_s32_v:
   case NEON::BI__builtin_neon_vcvtn_s64_v:
+  case NEON::BI__builtin_neon_vcvtn_u16_v:
   case NEON::BI__builtin_neon_vcvtn_u32_v:
   case NEON::BI__builtin_neon_vcvtn_u64_v:
+  case NEON::BI__builtin_neon_vcvtnq_s16_v:
   case NEON::BI__builtin_neon_vcvtnq_s32_v:
   case NEON::BI__builtin_neon_vcvtnq_s64_v:
+  case NEON::BI__builtin_neon_vcvtnq_u16_v:
   case NEON::BI__builtin_neon_vcvtnq_u32_v:
   case NEON::BI__builtin_neon_vcvtnq_u64_v:
+  case NEON::BI__builtin_neon_vcvtp_s16_v:
   case NEON::BI__builtin_neon_vcvtp_s32_v:
   case NEON::BI__builtin_neon_vcvtp_s64_v:
+  case NEON::BI__builtin_neon_vcvtp_u16_v:
   case NEON::BI__builtin_neon_vcvtp_u32_v:
   case NEON::BI__builtin_neon_vcvtp_u64_v:
+  case NEON::BI__builtin_neon_vcvtpq_s16_v:
   case NEON::BI__builtin_neon_vcvtpq_s32_v:
   case NEON::BI__builtin_neon_vcvtpq_s64_v:
+  case NEON::BI__builtin_neon_vcvtpq_u16_v:
   case NEON::BI__builtin_neon_vcvtpq_u32_v:
   case NEON::BI__builtin_neon_vcvtpq_u64_v:
+  case NEON::BI__builtin_neon_vcvtm_s16_v:
   case NEON::BI__builtin_neon_vcvtm_s32_v:
   case NEON::BI__builtin_neon_vcvtm_s64_v:
+  case NEON::BI__builtin_neon_vcvtm_u16_v:
   case NEON::BI__builtin_neon_vcvtm_u32_v:
   case NEON::BI__builtin_neon_vcvtm_u64_v:
+  case NEON::BI__builtin_neon_vcvtmq_s16_v:
   case NEON::BI__builtin_neon_vcvtmq_s32_v:
   case NEON::BI__builtin_neon_vcvtmq_s64_v:
+  case NEON::BI__builtin_neon_vcvtmq_u16_v:
   case NEON::BI__builtin_neon_vcvtmq_u32_v:
   case NEON::BI__builtin_neon_vcvtmq_u64_v: {
     llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
@@ -6110,7 +6188,9 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
     Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3]));
     return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]});
   }
+  case NEON::BI__builtin_neon_vfmah_lane_f16:
   case NEON::BI__builtin_neon_vfmas_lane_f32:
+  case NEON::BI__builtin_neon_vfmah_laneq_f16:
   case NEON::BI__builtin_neon_vfmas_laneq_f32:
   case NEON::BI__builtin_neon_vfmad_lane_f64:
   case NEON::BI__builtin_neon_vfmad_laneq_f64: {
@@ -6285,18 +6365,25 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
   case NEON::BI__builtin_neon_vcvt_u32_v:
   case NEON::BI__builtin_neon_vcvt_s64_v:
   case NEON::BI__builtin_neon_vcvt_u64_v:
+	case NEON::BI__builtin_neon_vcvt_s16_v:
+	case NEON::BI__builtin_neon_vcvt_u16_v:
   case NEON::BI__builtin_neon_vcvtq_s32_v:
   case NEON::BI__builtin_neon_vcvtq_u32_v:
   case NEON::BI__builtin_neon_vcvtq_s64_v:
-  case NEON::BI__builtin_neon_vcvtq_u64_v: {
+  case NEON::BI__builtin_neon_vcvtq_u64_v:
+	case NEON::BI__builtin_neon_vcvtq_s16_v:
+	case NEON::BI__builtin_neon_vcvtq_u16_v: {
     Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type));
     if (usgn)
       return Builder.CreateFPToUI(Ops[0], Ty);
     return Builder.CreateFPToSI(Ops[0], Ty);
   }
+  case NEON::BI__builtin_neon_vcvta_s16_v:
   case NEON::BI__builtin_neon_vcvta_s32_v:
+  case NEON::BI__builtin_neon_vcvtaq_s16_v:
   case NEON::BI__builtin_neon_vcvtaq_s32_v:
   case NEON::BI__builtin_neon_vcvta_u32_v:
+  case NEON::BI__builtin_neon_vcvtaq_u16_v:
   case NEON::BI__builtin_neon_vcvtaq_u32_v:
   case NEON::BI__builtin_neon_vcvta_s64_v:
   case NEON::BI__builtin_neon_vcvtaq_s64_v:
@@ -6306,9 +6393,13 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
     llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta");
   }
+  case NEON::BI__builtin_neon_vcvtm_s16_v:
   case NEON::BI__builtin_neon_vcvtm_s32_v:
+  case NEON::BI__builtin_neon_vcvtmq_s16_v:
   case NEON::BI__builtin_neon_vcvtmq_s32_v:
+  case NEON::BI__builtin_neon_vcvtm_u16_v:
   case NEON::BI__builtin_neon_vcvtm_u32_v:
+  case NEON::BI__builtin_neon_vcvtmq_u16_v:
   case NEON::BI__builtin_neon_vcvtmq_u32_v:
   case NEON::BI__builtin_neon_vcvtm_s64_v:
   case NEON::BI__builtin_neon_vcvtmq_s64_v:
@@ -6318,9 +6409,13 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
     llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm");
   }
+  case NEON::BI__builtin_neon_vcvtn_s16_v:
   case NEON::BI__builtin_neon_vcvtn_s32_v:
+  case NEON::BI__builtin_neon_vcvtnq_s16_v:
   case NEON::BI__builtin_neon_vcvtnq_s32_v:
+  case NEON::BI__builtin_neon_vcvtn_u16_v:
   case NEON::BI__builtin_neon_vcvtn_u32_v:
+  case NEON::BI__builtin_neon_vcvtnq_u16_v:
   case NEON::BI__builtin_neon_vcvtnq_u32_v:
   case NEON::BI__builtin_neon_vcvtn_s64_v:
   case NEON::BI__builtin_neon_vcvtnq_s64_v:
@@ -6330,9 +6425,13 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
     llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn");
   }
+  case NEON::BI__builtin_neon_vcvtp_s16_v:
   case NEON::BI__builtin_neon_vcvtp_s32_v:
+  case NEON::BI__builtin_neon_vcvtpq_s16_v:
   case NEON::BI__builtin_neon_vcvtpq_s32_v:
+  case NEON::BI__builtin_neon_vcvtp_u16_v:
   case NEON::BI__builtin_neon_vcvtp_u32_v:
+  case NEON::BI__builtin_neon_vcvtpq_u16_v:
   case NEON::BI__builtin_neon_vcvtpq_u32_v:
   case NEON::BI__builtin_neon_vcvtp_s64_v:
   case NEON::BI__builtin_neon_vcvtpq_s64_v:
@@ -6505,6 +6604,24 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
     return Builder.CreateTrunc(Ops[0], Int16Ty);
   }
+  case NEON::BI__builtin_neon_vmaxv_f16: {
+    Int = Intrinsic::aarch64_neon_fmaxv;
+    Ty = HalfTy;
+    VTy = llvm::VectorType::get(HalfTy, 4);
+    llvm::Type *Tys[2] = { Ty, VTy };
+    Ops.push_back(EmitScalarExpr(E->getArg(0)));
+    Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
+    return Builder.CreateTrunc(Ops[0], HalfTy);
+  }
+  case NEON::BI__builtin_neon_vmaxvq_f16: {
+    Int = Intrinsic::aarch64_neon_fmaxv;
+    Ty = HalfTy;
+    VTy = llvm::VectorType::get(HalfTy, 8);
+    llvm::Type *Tys[2] = { Ty, VTy };
+    Ops.push_back(EmitScalarExpr(E->getArg(0)));
+    Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
+    return Builder.CreateTrunc(Ops[0], HalfTy);
+  }
   case NEON::BI__builtin_neon_vminv_u8: {
     Int = Intrinsic::aarch64_neon_uminv;
     Ty = Int32Ty;
@@ -6577,6 +6694,60 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
     return Builder.CreateTrunc(Ops[0], Int16Ty);
   }
+  case NEON::BI__builtin_neon_vminv_f16: {
+    Int = Intrinsic::aarch64_neon_fminv;
+    Ty = HalfTy;
+    VTy = llvm::VectorType::get(HalfTy, 4);
+    llvm::Type *Tys[2] = { Ty, VTy };
+    Ops.push_back(EmitScalarExpr(E->getArg(0)));
+    Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
+    return Builder.CreateTrunc(Ops[0], HalfTy);
+  }
+  case NEON::BI__builtin_neon_vminvq_f16: {
+    Int = Intrinsic::aarch64_neon_fminv;
+    Ty = HalfTy;
+    VTy = llvm::VectorType::get(HalfTy, 8);
+    llvm::Type *Tys[2] = { Ty, VTy };
+    Ops.push_back(EmitScalarExpr(E->getArg(0)));
+    Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
+    return Builder.CreateTrunc(Ops[0], HalfTy);
+  }
+  case NEON::BI__builtin_neon_vmaxnmv_f16: {
+    Int = Intrinsic::aarch64_neon_fmaxnmv;
+    Ty = HalfTy;
+    VTy = llvm::VectorType::get(HalfTy, 4);
+    llvm::Type *Tys[2] = { Ty, VTy };
+    Ops.push_back(EmitScalarExpr(E->getArg(0)));
+    Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv");
+    return Builder.CreateTrunc(Ops[0], HalfTy);
+  }
+  case NEON::BI__builtin_neon_vmaxnmvq_f16: {
+    Int = Intrinsic::aarch64_neon_fmaxnmv;
+    Ty = HalfTy;
+    VTy = llvm::VectorType::get(HalfTy, 8);
+    llvm::Type *Tys[2] = { Ty, VTy };
+    Ops.push_back(EmitScalarExpr(E->getArg(0)));
+    Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv");
+    return Builder.CreateTrunc(Ops[0], HalfTy);
+  }
+  case NEON::BI__builtin_neon_vminnmv_f16: {
+    Int = Intrinsic::aarch64_neon_fminnmv;
+    Ty = HalfTy;
+    VTy = llvm::VectorType::get(HalfTy, 4);
+    llvm::Type *Tys[2] = { Ty, VTy };
+    Ops.push_back(EmitScalarExpr(E->getArg(0)));
+    Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv");
+    return Builder.CreateTrunc(Ops[0], HalfTy);
+  }
+  case NEON::BI__builtin_neon_vminnmvq_f16: {
+    Int = Intrinsic::aarch64_neon_fminnmv;
+    Ty = HalfTy;
+    VTy = llvm::VectorType::get(HalfTy, 8);
+    llvm::Type *Tys[2] = { Ty, VTy };
+    Ops.push_back(EmitScalarExpr(E->getArg(0)));
+    Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv");
+    return Builder.CreateTrunc(Ops[0], HalfTy);
+  }
   case NEON::BI__builtin_neon_vmul_n_f64: {
     Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
     Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy);
diff --git a/clang/lib/CodeGen/CodeGenModule.cpp b/clang/lib/CodeGen/CodeGenModule.cpp
index 20d945fe50d..5319ccec163 100644
--- a/clang/lib/CodeGen/CodeGenModule.cpp
+++ b/clang/lib/CodeGen/CodeGenModule.cpp
@@ -98,6 +98,7 @@ CodeGenModule::CodeGenModule(ASTContext &C, const HeaderSearchOptions &HSO,
   Int16Ty = llvm::Type::getInt16Ty(LLVMContext);
   Int32Ty = llvm::Type::getInt32Ty(LLVMContext);
   Int64Ty = llvm::Type::getInt64Ty(LLVMContext);
+  HalfTy = llvm::Type::getHalfTy(LLVMContext);
   FloatTy = llvm::Type::getFloatTy(LLVMContext);
   DoubleTy = llvm::Type::getDoubleTy(LLVMContext);
   PointerWidthInBits = C.getTargetInfo().getPointerWidth(0);
diff --git a/clang/lib/CodeGen/CodeGenTypeCache.h b/clang/lib/CodeGen/CodeGenTypeCache.h
index 450eab48a3b..6910d36733d 100644
--- a/clang/lib/CodeGen/CodeGenTypeCache.h
+++ b/clang/lib/CodeGen/CodeGenTypeCache.h
@@ -36,7 +36,7 @@ struct CodeGenTypeCache {
   /// i8, i16, i32, and i64
   llvm::IntegerType *Int8Ty, *Int16Ty, *Int32Ty, *Int64Ty;
   /// float, double
-  llvm::Type *FloatTy, *DoubleTy;
+  llvm::Type *HalfTy, *FloatTy, *DoubleTy;
 
   /// int
   llvm::IntegerType *IntTy;