1 ; RUN: llc < %s -march=arm -mattr=+neon | FileCheck %s
3 %struct.__neon_int8x8x2_t = type { <8 x i8>, <8 x i8> }
4 %struct.__neon_int16x4x2_t = type { <4 x i16>, <4 x i16> }
5 %struct.__neon_int32x2x2_t = type { <2 x i32>, <2 x i32> }
6 %struct.__neon_float32x2x2_t = type { <2 x float>, <2 x float> }
7 %struct.__neon_int64x1x2_t = type { <1 x i64>, <1 x i64> }
9 %struct.__neon_int8x16x2_t = type { <16 x i8>, <16 x i8> }
10 %struct.__neon_int16x8x2_t = type { <8 x i16>, <8 x i16> }
11 %struct.__neon_int32x4x2_t = type { <4 x i32>, <4 x i32> }
12 %struct.__neon_float32x4x2_t = type { <4 x float>, <4 x float> }
14 define <8 x i8> @vld2i8(i8* %A) nounwind {
17 %tmp1 = call %struct.__neon_int8x8x2_t @llvm.arm.neon.vld2.v8i8(i8* %A)
18 %tmp2 = extractvalue %struct.__neon_int8x8x2_t %tmp1, 0
19 %tmp3 = extractvalue %struct.__neon_int8x8x2_t %tmp1, 1
20 %tmp4 = add <8 x i8> %tmp2, %tmp3
24 define <4 x i16> @vld2i16(i16* %A) nounwind {
27 %tmp0 = bitcast i16* %A to i8*
28 %tmp1 = call %struct.__neon_int16x4x2_t @llvm.arm.neon.vld2.v4i16(i8* %tmp0)
29 %tmp2 = extractvalue %struct.__neon_int16x4x2_t %tmp1, 0
30 %tmp3 = extractvalue %struct.__neon_int16x4x2_t %tmp1, 1
31 %tmp4 = add <4 x i16> %tmp2, %tmp3
35 define <2 x i32> @vld2i32(i32* %A) nounwind {
38 %tmp0 = bitcast i32* %A to i8*
39 %tmp1 = call %struct.__neon_int32x2x2_t @llvm.arm.neon.vld2.v2i32(i8* %tmp0)
40 %tmp2 = extractvalue %struct.__neon_int32x2x2_t %tmp1, 0
41 %tmp3 = extractvalue %struct.__neon_int32x2x2_t %tmp1, 1
42 %tmp4 = add <2 x i32> %tmp2, %tmp3
46 define <2 x float> @vld2f(float* %A) nounwind {
49 %tmp0 = bitcast float* %A to i8*
50 %tmp1 = call %struct.__neon_float32x2x2_t @llvm.arm.neon.vld2.v2f32(i8* %tmp0)
51 %tmp2 = extractvalue %struct.__neon_float32x2x2_t %tmp1, 0
52 %tmp3 = extractvalue %struct.__neon_float32x2x2_t %tmp1, 1
53 %tmp4 = add <2 x float> %tmp2, %tmp3
57 define <1 x i64> @vld2i64(i64* %A) nounwind {
60 %tmp0 = bitcast i64* %A to i8*
61 %tmp1 = call %struct.__neon_int64x1x2_t @llvm.arm.neon.vld2.v1i64(i8* %tmp0)
62 %tmp2 = extractvalue %struct.__neon_int64x1x2_t %tmp1, 0
63 %tmp3 = extractvalue %struct.__neon_int64x1x2_t %tmp1, 1
64 %tmp4 = add <1 x i64> %tmp2, %tmp3
68 define <16 x i8> @vld2Qi8(i8* %A) nounwind {
71 %tmp1 = call %struct.__neon_int8x16x2_t @llvm.arm.neon.vld2.v16i8(i8* %A)
72 %tmp2 = extractvalue %struct.__neon_int8x16x2_t %tmp1, 0
73 %tmp3 = extractvalue %struct.__neon_int8x16x2_t %tmp1, 1
74 %tmp4 = add <16 x i8> %tmp2, %tmp3
78 define <8 x i16> @vld2Qi16(i16* %A) nounwind {
81 %tmp0 = bitcast i16* %A to i8*
82 %tmp1 = call %struct.__neon_int16x8x2_t @llvm.arm.neon.vld2.v8i16(i8* %tmp0)
83 %tmp2 = extractvalue %struct.__neon_int16x8x2_t %tmp1, 0
84 %tmp3 = extractvalue %struct.__neon_int16x8x2_t %tmp1, 1
85 %tmp4 = add <8 x i16> %tmp2, %tmp3
89 define <4 x i32> @vld2Qi32(i32* %A) nounwind {
92 %tmp0 = bitcast i32* %A to i8*
93 %tmp1 = call %struct.__neon_int32x4x2_t @llvm.arm.neon.vld2.v4i32(i8* %tmp0)
94 %tmp2 = extractvalue %struct.__neon_int32x4x2_t %tmp1, 0
95 %tmp3 = extractvalue %struct.__neon_int32x4x2_t %tmp1, 1
96 %tmp4 = add <4 x i32> %tmp2, %tmp3
100 define <4 x float> @vld2Qf(float* %A) nounwind {
103 %tmp0 = bitcast float* %A to i8*
104 %tmp1 = call %struct.__neon_float32x4x2_t @llvm.arm.neon.vld2.v4f32(i8* %tmp0)
105 %tmp2 = extractvalue %struct.__neon_float32x4x2_t %tmp1, 0
106 %tmp3 = extractvalue %struct.__neon_float32x4x2_t %tmp1, 1
107 %tmp4 = add <4 x float> %tmp2, %tmp3
108 ret <4 x float> %tmp4
111 declare %struct.__neon_int8x8x2_t @llvm.arm.neon.vld2.v8i8(i8*) nounwind readonly
112 declare %struct.__neon_int16x4x2_t @llvm.arm.neon.vld2.v4i16(i8*) nounwind readonly
113 declare %struct.__neon_int32x2x2_t @llvm.arm.neon.vld2.v2i32(i8*) nounwind readonly
114 declare %struct.__neon_float32x2x2_t @llvm.arm.neon.vld2.v2f32(i8*) nounwind readonly
115 declare %struct.__neon_int64x1x2_t @llvm.arm.neon.vld2.v1i64(i8*) nounwind readonly
117 declare %struct.__neon_int8x16x2_t @llvm.arm.neon.vld2.v16i8(i8*) nounwind readonly
118 declare %struct.__neon_int16x8x2_t @llvm.arm.neon.vld2.v8i16(i8*) nounwind readonly
119 declare %struct.__neon_int32x4x2_t @llvm.arm.neon.vld2.v4i32(i8*) nounwind readonly
120 declare %struct.__neon_float32x4x2_t @llvm.arm.neon.vld2.v4f32(i8*) nounwind readonly