Implement aarch64 neon instruction class SIMD Across - Clang

[Jiangning]

Hi,

This is the patch implementing aarch64 neon instruction class SIMD Across. Review, please!

Thanks,
-Jiangning

[t.p.northover]

Hi Jiangning,

Ana's probably the expert on the NEONEmitter, though I didn't see anything obviously wrong. Just one other comment.

Cheers.

Tim.

lib/CodeGen/CGBuiltin.cpp

1958

Couldn't this block be:

if (AcrossVec) {
  ...
} else if (OverloadInt) {
  ...
} else
  ...

for(...) Ops.push_back(...)

[Jiangning]

Tim,

OK, I get it changed accordingly.

Thanks,
-Jiangning

[Eugene.Zelenko]

Committed in rL192029.

include/clang/Basic/arm_neon.td

 // i: constant int
 // l: constant uint64
 // s: scalar of element type
+// r: scalar of double width element type
 // a: scalar of element type (splat to vector type)
 // k: default elt width, double num elts
 // #: array of default vectors

 def VQDMLSL_HIGH : SOpInst<"vqdmlsl_high", "wwkk", "si", OP_QDMLSLHi>;
 
 ////////////////////////////////////////////////////////////////////////////////
+// Across vectors class
+def VADDLV  : SInst<"vaddlv", "rd", "csiUcUsUiQcQsQiQUcQUsQUi">;
+def VMAXV   : SInst<"vmaxv", "sd", "csiUcUsUiQcQsQiQUcQUsQUiQf">;
+def VMINV   : SInst<"vminv", "sd", "csiUcUsUiQcQsQiQUcQUsQUiQf">;
+def VADDV   : SInst<"vaddv", "sd", "csiUcUsUiQcQsQiQUcQUsQUi">;
+def FMAXNMV : SInst<"vmaxnmv", "sd", "Qf">;
+def FMINNMV : SInst<"vminnmv", "sd", "Qf">;
+
+////////////////////////////////////////////////////////////////////////////////
 // Scalar Arithmetic
 
 // Scalar Addition

lib/CodeGen/CGBuiltin.cpp

 static Value *EmitAArch64ScalarBuiltinExpr(CodeGenFunction &CGF,
                                            unsigned BuiltinID,
                                            const CallExpr *E) {
-  NeonTypeFlags::EltType ET;
-  bool usgn;
   unsigned int Int = 0;
-  bool OverloadInt = true;
+  // Scalar result generated across vectors
+  bool AcrossVec = false;
+  // Extend element of one-element vector
+  bool ExtendEle = false;
+  bool OverloadInt = false;
   const char *s = NULL;
 
-  SmallVector<Value *, 4> Ops;
-  for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
-    Ops.push_back(CGF.EmitScalarExpr(E->getArg(i)));
-  }
-
   // AArch64 scalar builtins are not overloaded, they do not have an extra
   // argument that specifies the vector type, need to handle each case.
   switch (BuiltinID) {
   default: break;
   // Scalar Add
   case AArch64::BI__builtin_neon_vaddd_s64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vaddds;
-    s = "vaddds"; usgn = false; OverloadInt = false; break;
+    Int = Intrinsic::aarch64_neon_vaddds;
+    s = "vaddds"; break;
   case AArch64::BI__builtin_neon_vaddd_u64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vadddu;
-    s = "vadddu"; usgn = true; OverloadInt = false; break;
+    Int = Intrinsic::aarch64_neon_vadddu;
+    s = "vadddu"; break;
   // Scalar Sub
   case AArch64::BI__builtin_neon_vsubd_s64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vsubds;
-    s = "vsubds"; usgn = false; OverloadInt = false; break;
+    Int = Intrinsic::aarch64_neon_vsubds;
+    s = "vsubds"; break;
   case AArch64::BI__builtin_neon_vsubd_u64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vsubdu;
-    s = "vsubdu"; usgn = true; OverloadInt = false; break;
+    Int = Intrinsic::aarch64_neon_vsubdu;
+    s = "vsubdu"; break;
   // Scalar Saturating Add
   case AArch64::BI__builtin_neon_vqaddb_s8:
-    ET = NeonTypeFlags::Int8; Int = Intrinsic::aarch64_neon_vqadds;
-    s = "vqadds"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqaddh_s16:
-    ET = NeonTypeFlags::Int16; Int = Intrinsic::aarch64_neon_vqadds;
-    s = "vqadds"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqadds_s32:
-    ET = NeonTypeFlags::Int32; Int = Intrinsic::aarch64_neon_vqadds;
-    s = "vqadds"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqaddd_s64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vqadds;
-    s = "vqadds"; usgn = false; OverloadInt = true; break;
+    Int = Intrinsic::aarch64_neon_vqadds;
+    s = "vqadds"; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqaddb_u8:
-    ET = NeonTypeFlags::Int8; Int = Intrinsic::aarch64_neon_vqaddu;
-    s = "vqaddu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqaddh_u16:
-    ET = NeonTypeFlags::Int16; Int = Intrinsic::aarch64_neon_vqaddu;
-    s = "vqaddu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqadds_u32:
-    ET = NeonTypeFlags::Int32; Int = Intrinsic::aarch64_neon_vqaddu;
-    s = "vqaddu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqaddd_u64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vqaddu;
-    s = "vqaddu"; usgn = true; OverloadInt = true; break;
+    Int = Intrinsic::aarch64_neon_vqaddu;
+    s = "vqaddu"; OverloadInt = true; break;
   // Scalar Saturating Sub
   case AArch64::BI__builtin_neon_vqsubb_s8:
-    ET = NeonTypeFlags::Int8; Int = Intrinsic::aarch64_neon_vqsubs;
-    s = "vqsubs"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqsubh_s16:
-    ET = NeonTypeFlags::Int16; Int = Intrinsic::aarch64_neon_vqsubs;
-    s = "vqsubs"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqsubs_s32:
-    ET = NeonTypeFlags::Int32; Int = Intrinsic::aarch64_neon_vqsubs;
-    s = "vqsubs"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqsubd_s64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vqsubs;
-    s = "vqsubs"; usgn = false; OverloadInt = true; break;
+    Int = Intrinsic::aarch64_neon_vqsubs;
+    s = "vqsubs"; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqsubb_u8:
-    ET = NeonTypeFlags::Int8; Int = Intrinsic::aarch64_neon_vqsubu;
-    s = "vqsubu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqsubh_u16:
-    ET = NeonTypeFlags::Int16; Int = Intrinsic::aarch64_neon_vqsubu;
-    s = "vqsubu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqsubs_u32:
-    ET = NeonTypeFlags::Int32; Int = Intrinsic::aarch64_neon_vqsubu;
-    s = "vqsubu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqsubd_u64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vqsubu;
-    s = "vqsubu"; usgn = true; OverloadInt = true; break;
+    Int = Intrinsic::aarch64_neon_vqsubu;
+    s = "vqsubu"; OverloadInt = true; break;
   // Scalar Shift Left
   case AArch64::BI__builtin_neon_vshld_s64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vshlds;
-    s = "vshlds"; usgn = false; OverloadInt=false; break;
+    Int = Intrinsic::aarch64_neon_vshlds;
+    s = "vshlds"; break;
   case AArch64::BI__builtin_neon_vshld_u64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vshldu;
-    s = "vshldu"; usgn = true; OverloadInt = false; break;
+    Int = Intrinsic::aarch64_neon_vshldu;
+    s = "vshldu"; break;
   // Scalar Saturating Shift Left
   case AArch64::BI__builtin_neon_vqshlb_s8:
-    ET = NeonTypeFlags::Int8; Int = Intrinsic::aarch64_neon_vqshls;
-    s = "vqshls"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqshlh_s16:
-    ET = NeonTypeFlags::Int16; Int = Intrinsic::aarch64_neon_vqshls;
-    s = "vqshls"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqshls_s32:
-    ET = NeonTypeFlags::Int32; Int = Intrinsic::aarch64_neon_vqshls;
-    s = "vqshls"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqshld_s64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vqshls;
-    s = "vqshls"; usgn = false; OverloadInt = true; break;
+    Int = Intrinsic::aarch64_neon_vqshls;
+    s = "vqshls"; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqshlb_u8:
-    ET = NeonTypeFlags::Int8; Int = Intrinsic::aarch64_neon_vqshlu;
-    s = "vqshlu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqshlh_u16:
-    ET = NeonTypeFlags::Int16; Int = Intrinsic::aarch64_neon_vqshlu;
-    s = "vqshlu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqshls_u32:
-    ET = NeonTypeFlags::Int32; Int = Intrinsic::aarch64_neon_vqshlu;
-    s = "vqshlu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqshld_u64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vqshlu;
-    s = "vqshlu"; usgn = true; OverloadInt = true; break;
+    Int = Intrinsic::aarch64_neon_vqshlu;
+    s = "vqshlu"; OverloadInt = true; break;
   // Scalar Rouding Shift Left
   case AArch64::BI__builtin_neon_vrshld_s64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vrshlds;
-    s = "vrshlds"; usgn = false; OverloadInt=false; break;
+    Int = Intrinsic::aarch64_neon_vrshlds;
+    s = "vrshlds"; break;
   case AArch64::BI__builtin_neon_vrshld_u64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vrshldu;
-    s = "vrshldu"; usgn = true; OverloadInt=false; break;
+    Int = Intrinsic::aarch64_neon_vrshldu;
+    s = "vrshldu"; break;
   // Scalar Saturating Rouding Shift Left
   case AArch64::BI__builtin_neon_vqrshlb_s8:
-    ET = NeonTypeFlags::Int8; Int = Intrinsic::aarch64_neon_vqrshls;
-    s = "vqrshls"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqrshlh_s16:
-    ET = NeonTypeFlags::Int16; Int = Intrinsic::aarch64_neon_vqrshls;
-    s = "vqrshls"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqrshls_s32:
-    ET = NeonTypeFlags::Int32; Int = Intrinsic::aarch64_neon_vqrshls;
-    s = "vqrshls"; usgn = false; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqrshld_s64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vqrshls;
-    s = "vqrshls"; usgn = false; OverloadInt = true; break;
+    Int = Intrinsic::aarch64_neon_vqrshls;
+    s = "vqrshls"; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqrshlb_u8:
-    ET = NeonTypeFlags::Int8; Int = Intrinsic::aarch64_neon_vqrshlu;
-    s = "vqrshlu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqrshlh_u16:
-    ET = NeonTypeFlags::Int16; Int = Intrinsic::aarch64_neon_vqrshlu;
-    s = "vqrshlu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqrshls_u32:
-    ET = NeonTypeFlags::Int32; Int = Intrinsic::aarch64_neon_vqrshlu;
-    s = "vqrshlu"; usgn = true; OverloadInt = true; break;
   case AArch64::BI__builtin_neon_vqrshld_u64:
-    ET = NeonTypeFlags::Int64; Int = Intrinsic::aarch64_neon_vqrshlu;
-    s = "vqrshlu"; usgn = true; OverloadInt = true; break;
+    Int = Intrinsic::aarch64_neon_vqrshlu;
+    s = "vqrshlu"; OverloadInt = true; break;
   // Scalar Reduce Pairwise Add
   case AArch64::BI__builtin_neon_vpaddd_s64:
     Int = Intrinsic::aarch64_neon_vpadd; s = "vpadd";
-    OverloadInt = false; break;
+    break;
   case AArch64::BI__builtin_neon_vpadds_f32:
     Int = Intrinsic::aarch64_neon_vpfadd; s = "vpfadd";
-    OverloadInt = false; break;
+    break;
   case AArch64::BI__builtin_neon_vpaddd_f64:
     Int = Intrinsic::aarch64_neon_vpfaddq; s = "vpfaddq";
-    OverloadInt = false; break;
+    break;
   // Scalar Reduce Pairwise Floating Point Max
   case AArch64::BI__builtin_neon_vpmaxs_f32:
     Int = Intrinsic::aarch64_neon_vpmax; s = "vpmax";
-    OverloadInt = false; break;
+    break;
   case AArch64::BI__builtin_neon_vpmaxqd_f64:
     Int = Intrinsic::aarch64_neon_vpmaxq; s = "vpmaxq";
-    OverloadInt = false; break;
+    break;
   // Scalar Reduce Pairwise Floating Point Min
   case AArch64::BI__builtin_neon_vpmins_f32:
     Int = Intrinsic::aarch64_neon_vpmin; s = "vpmin";
-    OverloadInt = false; break;
+    break;
   case AArch64::BI__builtin_neon_vpminqd_f64:
     Int = Intrinsic::aarch64_neon_vpminq; s = "vpminq";
-    OverloadInt = false; break;
+    break;
   // Scalar Reduce Pairwise Floating Point Maxnm
   case AArch64::BI__builtin_neon_vpmaxnms_f32:
     Int = Intrinsic::aarch64_neon_vpfmaxnm; s = "vpfmaxnm";
-    OverloadInt = false; break;
+    break;
   case AArch64::BI__builtin_neon_vpmaxnmqd_f64:
     Int = Intrinsic::aarch64_neon_vpfmaxnmq; s = "vpfmaxnmq";
-    OverloadInt = false; break;
-   // Scalar Reduce Pairwise Floating Point Minnm
+    break;
   case AArch64::BI__builtin_neon_vpminnms_f32:
     Int = Intrinsic::aarch64_neon_vpfminnm; s = "vpfminnm";
-    OverloadInt = false; break;
+    break;
   case AArch64::BI__builtin_neon_vpminnmqd_f64:
     Int = Intrinsic::aarch64_neon_vpfminnmq; s = "vpfminnmq";
-    OverloadInt = false; break;
+    break;
+  // The followings are intrinsics with scalar results generated AcrossVec vectors
+  case AArch64::BI__builtin_neon_vaddlv_s8:
+  case AArch64::BI__builtin_neon_vaddlv_s16:
+  case AArch64::BI__builtin_neon_vaddlvq_s8:
+  case AArch64::BI__builtin_neon_vaddlvq_s16:
+  case AArch64::BI__builtin_neon_vaddlvq_s32:
+    Int = Intrinsic::aarch64_neon_saddlv;
+    AcrossVec = true; ExtendEle = true; s = "saddlv"; break;
+  case AArch64::BI__builtin_neon_vaddlv_u8:
+  case AArch64::BI__builtin_neon_vaddlv_u16:
+  case AArch64::BI__builtin_neon_vaddlvq_u8:
+  case AArch64::BI__builtin_neon_vaddlvq_u16:
+  case AArch64::BI__builtin_neon_vaddlvq_u32:
+    Int = Intrinsic::aarch64_neon_uaddlv;
+    AcrossVec = true; ExtendEle = true; s = "uaddlv"; break;
+  case AArch64::BI__builtin_neon_vmaxv_s8:
+  case AArch64::BI__builtin_neon_vmaxv_s16:
+  case AArch64::BI__builtin_neon_vmaxvq_s8:
+  case AArch64::BI__builtin_neon_vmaxvq_s16:
+  case AArch64::BI__builtin_neon_vmaxvq_s32:
+    Int = Intrinsic::aarch64_neon_smaxv;
+    AcrossVec = true; ExtendEle = false; s = "smaxv"; break;
+  case AArch64::BI__builtin_neon_vmaxv_u8:
+  case AArch64::BI__builtin_neon_vmaxv_u16:
+  case AArch64::BI__builtin_neon_vmaxvq_u8:
+  case AArch64::BI__builtin_neon_vmaxvq_u16:
+  case AArch64::BI__builtin_neon_vmaxvq_u32:
+    Int = Intrinsic::aarch64_neon_umaxv;
+    AcrossVec = true; ExtendEle = false; s = "umaxv"; break;
+  case AArch64::BI__builtin_neon_vminv_s8:
+  case AArch64::BI__builtin_neon_vminv_s16:
+  case AArch64::BI__builtin_neon_vminvq_s8:
+  case AArch64::BI__builtin_neon_vminvq_s16:
+  case AArch64::BI__builtin_neon_vminvq_s32:
+    Int = Intrinsic::aarch64_neon_sminv;
+    AcrossVec = true; ExtendEle = false; s = "sminv"; break;
+  case AArch64::BI__builtin_neon_vminv_u8:
+  case AArch64::BI__builtin_neon_vminv_u16:
+  case AArch64::BI__builtin_neon_vminvq_u8:
+  case AArch64::BI__builtin_neon_vminvq_u16:
+  case AArch64::BI__builtin_neon_vminvq_u32:
+    Int = Intrinsic::aarch64_neon_uminv;
+    AcrossVec = true; ExtendEle = false; s = "uminv"; break;
+  case AArch64::BI__builtin_neon_vaddv_s8:
+  case AArch64::BI__builtin_neon_vaddv_s16:
+  case AArch64::BI__builtin_neon_vaddvq_s8:
+  case AArch64::BI__builtin_neon_vaddvq_s16:
+  case AArch64::BI__builtin_neon_vaddvq_s32:
+  case AArch64::BI__builtin_neon_vaddv_u8:
+  case AArch64::BI__builtin_neon_vaddv_u16:
+  case AArch64::BI__builtin_neon_vaddvq_u8:
+  case AArch64::BI__builtin_neon_vaddvq_u16:
+  case AArch64::BI__builtin_neon_vaddvq_u32:
+    Int = Intrinsic::aarch64_neon_vaddv;
+    AcrossVec = true; ExtendEle = false; s = "vaddv"; break;      
+  case AArch64::BI__builtin_neon_vmaxvq_f32:
+    Int = Intrinsic::aarch64_neon_vmaxv;
+    AcrossVec = true; ExtendEle = false; s = "vmaxv"; break;
+  case AArch64::BI__builtin_neon_vminvq_f32:
+    Int = Intrinsic::aarch64_neon_vminv;
+    AcrossVec = true; ExtendEle = false; s = "vminv"; break;
+  case AArch64::BI__builtin_neon_vmaxnmvq_f32:
+    Int = Intrinsic::aarch64_neon_vmaxnmv;
+    AcrossVec = true; ExtendEle = false; s = "vmaxnmv"; break;
+  case AArch64::BI__builtin_neon_vminnmvq_f32:
+    Int = Intrinsic::aarch64_neon_vminnmv;
+    AcrossVec = true; ExtendEle = false; s = "vminnmv"; break;
   }
 
   if (!Int)
     return 0;
 
   // AArch64 scalar builtin that returns scalar type
-  // and should be mapped to AArch64 intrinsic that takes
-  // one-element vector type arguments and returns
+  // and should be mapped to AArch64 intrinsic that returns
   // one-element vector type.
-  llvm::Type *Ty = 0;
   Function *F = 0;
-  if (OverloadInt) {
+  SmallVector<Value *, 4> Ops;
+  if (AcrossVec) {
+    // Gen arg type
+    const Expr *Arg = E->getArg(E->getNumArgs()-1);
+    llvm::Type *Ty = CGF.ConvertType(Arg->getType());
+    llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
+    llvm::Type *ETy = VTy->getElementType();
+    llvm::VectorType *RTy = llvm::VectorType::get(ETy, 1);
+  
+    if (ExtendEle) {
+      assert(!ETy->isFloatingPointTy());
+      RTy = llvm::VectorType::getExtendedElementVectorType(RTy);
+    }
+
+    llvm::Type *Tys[2] = {RTy, VTy};
+    F = CGF.CGM.getIntrinsic(Int, Tys);
+    assert(E->getNumArgs() == 1);
+  }
+  else if (OverloadInt) {
     // Determine the type of this overloaded AArch64 intrinsic
-    NeonTypeFlags Type(ET, usgn, false);
-    llvm::VectorType *VTy = GetNeonType(&CGF, Type, true);
-    Ty = VTy;
-    if (!Ty)
-      return 0;
-    F = CGF.CGM.getIntrinsic(Int, Ty);
+    const Expr *Arg = E->getArg(E->getNumArgs()-1);
+    llvm::Type *Ty = CGF.ConvertType(Arg->getType());
+    llvm::VectorType *VTy = llvm::VectorType::get(Ty, 1);
+    assert(VTy);
+
+    F = CGF.CGM.getIntrinsic(Int, VTy);
   } else
     F = CGF.CGM.getIntrinsic(Int);
 
+  for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
+    Ops.push_back(CGF.EmitScalarExpr(E->getArg(i)));
+  }
+
   Value *Result = CGF.EmitNeonCall(F, Ops, s);
   llvm::Type *ResultType = CGF.ConvertType(E->getType());
   // AArch64 intrinsic one-element vector type cast to

 }
 
 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
                                                      const CallExpr *E) {
 
   // Process AArch64 scalar builtins
   if (Value *Result = EmitAArch64ScalarBuiltinExpr(*this, BuiltinID, E))

test/CodeGen/aarch64-neon-across.c

+// REQUIRES: aarch64-registered-target
+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +neon \
+// RUN:   -ffp-contract=fast -S -O3 -o - %s | FileCheck %s
+
+// Test new aarch64 intrinsics and types
+
+#include <arm_neon.h>
+
+int16_t test_vaddlv_s8(int8x8_t a) {
+  // CHECK: test_vaddlv_s8
+  return vaddlv_s8(a);
+  // CHECK: saddlv {{h[0-9]+}}, {{v[0-9]+}}.8b
+}
+
+int32_t test_vaddlv_s16(int16x4_t a) {
+  // CHECK: test_vaddlv_s16
+  return vaddlv_s16(a);
+  // CHECK: saddlv {{s[0-9]+}}, {{v[0-9]+}}.4h
+}
+
+uint16_t test_vaddlv_u8(uint8x8_t a) {
+  // CHECK: test_vaddlv_u8
+  return vaddlv_u8(a);
+  // CHECK: uaddlv {{h[0-9]+}}, {{v[0-9]+}}.8b
+}
+
+uint32_t test_vaddlv_u16(uint16x4_t a) {
+  // CHECK: test_vaddlv_u16
+  return vaddlv_u16(a);
+  // CHECK: uaddlv {{s[0-9]+}}, {{v[0-9]+}}.4h
+}
+
+int16_t test_vaddlvq_s8(int8x16_t a) {
+  // CHECK: test_vaddlvq_s8
+  return vaddlvq_s8(a);
+  // CHECK: saddlv {{h[0-9]+}}, {{v[0-9]+}}.16b
+}
+
+int32_t test_vaddlvq_s16(int16x8_t a) {
+  // CHECK: test_vaddlvq_s16
+  return vaddlvq_s16(a);
+  // CHECK: saddlv {{s[0-9]+}}, {{v[0-9]+}}.8h
+}
+
+int64_t test_vaddlvq_s32(int32x4_t a) {
+  // CHECK: test_vaddlvq_s32
+  return vaddlvq_s32(a);
+  // CHECK: saddlv {{d[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+uint16_t test_vaddlvq_u8(uint8x16_t a) {
+  // CHECK: test_vaddlvq_u8
+  return vaddlvq_u8(a);
+  // CHECK: uaddlv {{h[0-9]+}}, {{v[0-9]+}}.16b
+}
+
+uint32_t test_vaddlvq_u16(uint16x8_t a) {
+  // CHECK: test_vaddlvq_u16
+  return vaddlvq_u16(a);
+  // CHECK: uaddlv {{s[0-9]+}}, {{v[0-9]+}}.8h
+}
+
+uint64_t test_vaddlvq_u32(uint32x4_t a) {
+  // CHECK: test_vaddlvq_u32
+  return vaddlvq_u32(a);
+  // CHECK: uaddlv {{d[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+int8_t test_vmaxv_s8(int8x8_t a) {
+  // CHECK: test_vmaxv_s8
+  return vmaxv_s8(a);
+  // CHECK: smaxv {{b[0-9]+}}, {{v[0-9]+}}.8b
+}
+
+int16_t test_vmaxv_s16(int16x4_t a) {
+  // CHECK: test_vmaxv_s16
+  return vmaxv_s16(a);
+  // CHECK: smaxv {{h[0-9]+}}, {{v[0-9]+}}.4h
+}
+
+uint8_t test_vmaxv_u8(uint8x8_t a) {
+  // CHECK: test_vmaxv_u8
+  return vmaxv_u8(a);
+  // CHECK: umaxv {{b[0-9]+}}, {{v[0-9]+}}.8b
+}
+
+uint16_t test_vmaxv_u16(uint16x4_t a) {
+  // CHECK: test_vmaxv_u16
+  return vmaxv_u16(a);
+  // CHECK: umaxv {{h[0-9]+}}, {{v[0-9]+}}.4h
+}
+
+int8_t test_vmaxvq_s8(int8x16_t a) {
+  // CHECK: test_vmaxvq_s8
+  return vmaxvq_s8(a);
+  // CHECK: smaxv {{b[0-9]+}}, {{v[0-9]+}}.16b
+}
+
+int16_t test_vmaxvq_s16(int16x8_t a) {
+  // CHECK: test_vmaxvq_s16
+  return vmaxvq_s16(a);
+  // CHECK: smaxv {{h[0-9]+}}, {{v[0-9]+}}.8h
+}
+
+int32_t test_vmaxvq_s32(int32x4_t a) {
+  // CHECK: test_vmaxvq_s32
+  return vmaxvq_s32(a);
+  // CHECK: smaxv {{s[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+uint8_t test_vmaxvq_u8(uint8x16_t a) {
+  // CHECK: test_vmaxvq_u8
+  return vmaxvq_u8(a);
+  // CHECK: umaxv {{b[0-9]+}}, {{v[0-9]+}}.16b
+}
+
+uint16_t test_vmaxvq_u16(uint16x8_t a) {
+  // CHECK: test_vmaxvq_u16
+  return vmaxvq_u16(a);
+  // CHECK: umaxv {{h[0-9]+}}, {{v[0-9]+}}.8h
+}
+
+uint32_t test_vmaxvq_u32(uint32x4_t a) {
+  // CHECK: test_vmaxvq_u32
+  return vmaxvq_u32(a);
+  // CHECK: umaxv {{s[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+int8_t test_vminv_s8(int8x8_t a) {
+  // CHECK: test_vminv_s8
+  return vminv_s8(a);
+  // CHECK: sminv {{b[0-9]+}}, {{v[0-9]+}}.8b
+}
+
+int16_t test_vminv_s16(int16x4_t a) {
+  // CHECK: test_vminv_s16
+  return vminv_s16(a);
+  // CHECK: sminv {{h[0-9]+}}, {{v[0-9]+}}.4h
+}
+
+uint8_t test_vminv_u8(uint8x8_t a) {
+  // CHECK: test_vminv_u8
+  return vminv_u8(a);
+  // CHECK: uminv {{b[0-9]+}}, {{v[0-9]+}}.8b
+}
+
+uint16_t test_vminv_u16(uint16x4_t a) {
+  // CHECK: test_vminv_u16
+  return vminv_u16(a);
+  // CHECK: uminv {{h[0-9]+}}, {{v[0-9]+}}.4h
+}
+
+int8_t test_vminvq_s8(int8x16_t a) {
+  // CHECK: test_vminvq_s8
+  return vminvq_s8(a);
+  // CHECK: sminv {{b[0-9]+}}, {{v[0-9]+}}.16b
+}
+
+int16_t test_vminvq_s16(int16x8_t a) {
+  // CHECK: test_vminvq_s16
+  return vminvq_s16(a);
+  // CHECK: sminv {{h[0-9]+}}, {{v[0-9]+}}.8h
+}
+
+int32_t test_vminvq_s32(int32x4_t a) {
+  // CHECK: test_vminvq_s32
+  return vminvq_s32(a);
+  // CHECK: sminv {{s[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+uint8_t test_vminvq_u8(uint8x16_t a) {
+  // CHECK: test_vminvq_u8
+  return vminvq_u8(a);
+  // CHECK: uminv {{b[0-9]+}}, {{v[0-9]+}}.16b
+}
+
+uint16_t test_vminvq_u16(uint16x8_t a) {
+  // CHECK: test_vminvq_u16
+  return vminvq_u16(a);
+  // CHECK: uminv {{h[0-9]+}}, {{v[0-9]+}}.8h
+}
+
+uint32_t test_vminvq_u32(uint32x4_t a) {
+  // CHECK: test_vminvq_u32
+  return vminvq_u32(a);
+  // CHECK: uminv {{s[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+int8_t test_vaddv_s8(int8x8_t a) {
+  // CHECK: test_vaddv_s8
+  return vaddv_s8(a);
+  // CHECK: addv {{b[0-9]+}}, {{v[0-9]+}}.8b
+}
+
+int16_t test_vaddv_s16(int16x4_t a) {
+  // CHECK: test_vaddv_s16
+  return vaddv_s16(a);
+  // CHECK: addv {{h[0-9]+}}, {{v[0-9]+}}.4h
+}
+
+uint8_t test_vaddv_u8(uint8x8_t a) {
+  // CHECK: test_vaddv_u8
+  return vaddv_u8(a);
+  // CHECK: addv {{b[0-9]+}}, {{v[0-9]+}}.8b
+}
+
+uint16_t test_vaddv_u16(uint16x4_t a) {
+  // CHECK: test_vaddv_u16
+  return vaddv_u16(a);
+  // CHECK: addv {{h[0-9]+}}, {{v[0-9]+}}.4h
+}
+
+int8_t test_vaddvq_s8(int8x16_t a) {
+  // CHECK: test_vaddvq_s8
+  return vaddvq_s8(a);
+  // CHECK: addv {{b[0-9]+}}, {{v[0-9]+}}.16b
+}
+
+int16_t test_vaddvq_s16(int16x8_t a) {
+  // CHECK: test_vaddvq_s16
+  return vaddvq_s16(a);
+  // CHECK: addv {{h[0-9]+}}, {{v[0-9]+}}.8h
+}
+
+int32_t test_vaddvq_s32(int32x4_t a) {
+  // CHECK: test_vaddvq_s32
+  return vaddvq_s32(a);
+  // CHECK: addv {{s[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+uint8_t test_vaddvq_u8(uint8x16_t a) {
+  // CHECK: test_vaddvq_u8
+  return vaddvq_u8(a);
+  // CHECK: addv {{b[0-9]+}}, {{v[0-9]+}}.16b
+}
+
+uint16_t test_vaddvq_u16(uint16x8_t a) {
+  // CHECK: test_vaddvq_u16
+  return vaddvq_u16(a);
+  // CHECK: addv {{h[0-9]+}}, {{v[0-9]+}}.8h
+}
+
+uint32_t test_vaddvq_u32(uint32x4_t a) {
+  // CHECK: test_vaddvq_u32
+  return vaddvq_u32(a);
+  // CHECK: addv {{s[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+float32_t test_vmaxvq_f32(float32x4_t a) {
+  // CHECK: test_vmaxvq_f32
+  return vmaxvq_f32(a);
+  // CHECK: fmaxv {{s[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+float32_t test_vminvq_f32(float32x4_t a) {
+  // CHECK: test_vminvq_f32
+  return vminvq_f32(a);
+  // CHECK: fminv {{s[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+float32_t test_vmaxnmvq_f32(float32x4_t a) {
+  // CHECK: test_vmaxnmvq_f32
+  return vmaxnmvq_f32(a);
+  // CHECK: fmaxnmv {{s[0-9]+}}, {{v[0-9]+}}.4s
+}
+
+float32_t test_vminnmvq_f32(float32x4_t a) {
+  // CHECK: test_vminnmvq_f32
+  return vminnmvq_f32(a);
+  // CHECK: fminnmv {{s[0-9]+}}, {{v[0-9]+}}.4s
+}

utils/TableGen/NeonEmitter.cpp

       scal = true;
       usgn = true;
       break;
+    case 'r':
+      type = Widen(type);
     case 's':
     case 'a':
       scal = true;

   return Flags.getFlags();
 }
 
+static bool ProtoHasScalar(const std::string proto)
+{
+  return (proto.find('s') != std::string::npos
+          || proto.find('r') != std::string::npos);
+}
+
 // Generate the definition for this intrinsic, e.g. __builtin_neon_cls(a)
 static std::string GenBuiltin(const std::string &name, const std::string &proto,
                               StringRef typestr, ClassKind ck) {

   // Check if the prototype has a scalar operand with the type of the vector
   // elements.  If not, bitcasting the args will take care of arg checking.
   // The actual signedness etc. will be taken care of with special enums.
-  if (proto.find('s') == std::string::npos)
+  if (!ProtoHasScalar(proto))
     ck = ClassB;
 
   if (proto[0] != 'v') {

   // If all types are the same size, bitcasting the args will take care
   // of arg checking.  The actual signedness etc. will be taken care of with
   // special enums.
-  if (proto.find('s') == std::string::npos)
+  if (!ProtoHasScalar(proto))
     ck = ClassB;
 
   s += MangleName(name, typestr, ck);

         else
           PrintFatalError(R->getLoc(),
               "Fixed point convert name should contains \"32\" or \"64\"");
-      } else if (Proto.find('s') == std::string::npos) {
+      } else if (!ProtoHasScalar(Proto)) {
         // Builtins which are overloaded by type will need to have their upper
         // bound computed at Sema time based on the type constant.
         ck = ClassB;

 
     // Functions which have a scalar argument cannot be overloaded, no need to
     // check them if we are emitting the type checking code.
-    if (Proto.find('s') != std::string::npos)
+    if (ProtoHasScalar(Proto))
       continue;
 
     SmallVector<StringRef, 16> TypeVec;