ARM: Emit normal IR for vaddhn/vsubhn NEON intrinsics

These operations "vector add high-half narrow" actually correspond to the
sequence:

    %sum = add <4 x i32> %lhs, %rhs
    %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
    %res = trunc <4 x i32> %high to <4 x i16>

Now that LLVM can spot this, Clang should emit the corresponding LLVM IR.

llvm-svn: 189463

1 files changed, 36 insertions, 6 deletions

diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index c9a2fcbf460..59b25d689f1 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -2202,9 +2202,24 @@ Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID,
   case ARM::BI__builtin_neon_vabsq_v:
     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty),
                         Ops, "vabs");
-  case ARM::BI__builtin_neon_vaddhn_v:
-    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vaddhn, Ty),
-                        Ops, "vaddhn");
+  case ARM::BI__builtin_neon_vaddhn_v: {
+    llvm::VectorType *SrcTy =
+        llvm::VectorType::getExtendedElementVectorType(VTy);
+
+    // %sum = add <4 x i32> %lhs, %rhs
+    Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
+    Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
+    Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn");
+
+    // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
+    Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(),
+                                       SrcTy->getScalarSizeInBits() / 2);
+    ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt);
+    Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn");
+
+    // %res = trunc <4 x i32> %high to <4 x i16>
+    return Builder.CreateTrunc(Ops[0], VTy, "vaddhn");
+  }
   case ARM::BI__builtin_neon_vcale_v:
     std::swap(Ops[0], Ops[1]);
   case ARM::BI__builtin_neon_vcage_v: {
@@ -2782,9 +2797,24 @@ Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID,
     Ops.push_back(Align);
     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty),
                         Ops, "");
-  case ARM::BI__builtin_neon_vsubhn_v:
-    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vsubhn, Ty),
-                        Ops, "vsubhn");
+  case ARM::BI__builtin_neon_vsubhn_v: {
+    llvm::VectorType *SrcTy =
+        llvm::VectorType::getExtendedElementVectorType(VTy);
+
+    // %sum = add <4 x i32> %lhs, %rhs
+    Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
+    Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
+    Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn");
+
+    // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
+    Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(),
+                                       SrcTy->getScalarSizeInBits() / 2);
+    ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt);
+    Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn");
+
+    // %res = trunc <4 x i32> %high to <4 x i16>
+    return Builder.CreateTrunc(Ops[0], VTy, "vsubhn");
+  }
   case ARM::BI__builtin_neon_vtbl1_v:
     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1),
                         Ops, "vtbl1");