[CIR][CIRGen][Builtin][Neon] Lower neon_vaddv_s16 and neon_vaddv_u16 (#…

…1210)
llvm · Dec 6, 2024 · 8d507b1 · 8d507b1
1 parent 4ae9a04
commit 8d507b1
Show file tree

Hide file tree

Showing 2 changed files with 35 additions and 2 deletions.
diff --git a/clang/lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp b/clang/lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp
@@ -4183,10 +4183,17 @@ CIRGenFunction::emitAArch64BuiltinExpr(unsigned BuiltinID, const CallExpr *E,
     llvm_unreachable("NEON::BI__builtin_neon_vaddv_s8 NYI");
   }
   case NEON::BI__builtin_neon_vaddv_u16:
-    llvm_unreachable("NEON::BI__builtin_neon_vaddv_u16 NYI");
+    usgn = true;
     [[fallthrough]];
   case NEON::BI__builtin_neon_vaddv_s16: {
-    llvm_unreachable("NEON::BI__builtin_neon_vaddv_s16 NYI");
+    cir::IntType eltTy = usgn ? UInt16Ty : SInt16Ty;
+    cir::VectorType vTy = cir::VectorType::get(builder.getContext(), eltTy, 4);
+    Ops.push_back(emitScalarExpr(E->getArg(0)));
+    // This is to add across the vector elements, so wider result type needed.
+    Ops[0] = emitNeonCall(builder, {vTy}, Ops,
+                          usgn ? "aarch64.neon.uaddv" : "aarch64.neon.saddv",
+                          SInt32Ty, getLoc(E->getExprLoc()));
+    return builder.createIntCast(Ops[0], eltTy);
   }
   case NEON::BI__builtin_neon_vaddvq_u8:
     llvm_unreachable("NEON::BI__builtin_neon_vaddvq_u8 NYI");

diff --git a/clang/test/CIR/CodeGen/AArch64/neon-arith.c b/clang/test/CIR/CodeGen/AArch64/neon-arith.c
@@ -893,3 +893,29 @@ uint32_t test_vaddlvq_u16(uint16x8_t a) {
   // LLVM: [[VADDLV_I:%.*]] = call i32 @llvm.aarch64.neon.uaddlv.i32.v8i16(<8 x i16> [[A]])
   // LLVM: ret i32 [[VADDLV_I]]
 }
+
+uint16_t test_vaddv_u16(uint16x4_t a) {
+  return vaddv_u16(a);
+
+  // CIR-LABEL: vaddv_u16
+  // CIR: [[VADDV_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.uaddv" {{%.*}} : (!cir.vector<!u16i x 4>) -> !s32i
+  // CIR: cir.cast(integral, [[VADDV_I]] : !s32i), !u16i
+
+  // LLVM: {{.*}}test_vaddv_u16(<4 x i16>{{.*}}[[A:%.*]])
+  // LLVM:    [[VADDV_I:%.*]] = call i32 @llvm.aarch64.neon.uaddv.i32.v4i16(<4 x i16> [[A]])
+  // LLVM-NEXT:    [[TMP0:%.*]] = trunc i32 [[VADDV_I]] to i16
+  // LLVM-NEXT:    ret i16 [[TMP0]]
+}
+
+int16_t test_vaddv_s16(int16x4_t a) {
+  return vaddv_s16(a);
+
+  // CIR-LABEL: vaddv_s16
+  // CIR: [[VADDV_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.saddv" {{%.*}} : (!cir.vector<!s16i x 4>) -> !s32i
+  // CIR: cir.cast(integral, [[VADDV_I]] : !s32i), !s16i
+
+  // LLVM: {{.*}}test_vaddv_s16(<4 x i16>{{.*}}[[A:%.*]])
+  // LLVM:    [[VADDV_I:%.*]] = call i32 @llvm.aarch64.neon.saddv.i32.v4i16(<4 x i16> [[A]])
+  // LLVM-NEXT:    [[TMP0:%.*]] = trunc i32 [[VADDV_I]] to i16
+  // LLVM-NEXT:    ret i16 [[TMP0]]
+}