1 ; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs< %s | FileCheck -enable-var-scope -check-prefixes=GCN,SI %s
2 ; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs< %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI %s
4 ; Make sure high constant 0 isn't pointlessly materialized
5 ; GCN-LABEL: {{^}}trunc_bitcast_i64_lshr_32_i16:
7 ; GCN-NEXT: v_mov_b32_e32 v0, v1
8 ; GCN-NEXT: s_setpc_b64
9 define i16 @trunc_bitcast_i64_lshr_32_i16(i64 %bar) {
10 %srl = lshr i64 %bar, 32
11 %trunc = trunc i64 %srl to i16
15 ; GCN-LABEL: {{^}}trunc_bitcast_i64_lshr_32_i32:
17 ; GCN-NEXT: v_mov_b32_e32 v0, v1
18 ; GCN-NEXT: s_setpc_b64
19 define i32 @trunc_bitcast_i64_lshr_32_i32(i64 %bar) {
20 %srl = lshr i64 %bar, 32
21 %trunc = trunc i64 %srl to i32
25 ; GCN-LABEL: {{^}}trunc_bitcast_v2i32_to_i16:
27 ; GCN-NOT: _load_dword
29 ; VI: v_add_u16_e32 v0, 4, v0
30 define i16 @trunc_bitcast_v2i32_to_i16(<2 x i32> %bar) {
31 %load0 = load i32, i32 addrspace(1)* undef
32 %load1 = load i32, i32 addrspace(1)* null
33 %insert.0 = insertelement <2 x i32> undef, i32 %load0, i32 0
34 %insert.1 = insertelement <2 x i32> %insert.0, i32 99, i32 1
35 %bc = bitcast <2 x i32> %insert.1 to i64
36 %trunc = trunc i64 %bc to i16
37 %add = add i16 %trunc, 4
41 ; Make sure there's no crash if the source vector type is FP
42 ; GCN-LABEL: {{^}}trunc_bitcast_v2f32_to_i16:
44 ; GCN-NOT: _load_dword
46 ; VI: v_add_u16_e32 v0, 4, v0
47 define i16 @trunc_bitcast_v2f32_to_i16(<2 x float> %bar) {
48 %load0 = load float, float addrspace(1)* undef
49 %load1 = load float, float addrspace(1)* null
50 %insert.0 = insertelement <2 x float> undef, float %load0, i32 0
51 %insert.1 = insertelement <2 x float> %insert.0, float 4.0, i32 1
52 %bc = bitcast <2 x float> %insert.1 to i64
53 %trunc = trunc i64 %bc to i16
54 %add = add i16 %trunc, 4
58 ; GCN-LABEL: {{^}}truncate_high_elt_extract_vector:
64 ; GCN: v_lshrrev_b32_e32
65 define amdgpu_kernel void @truncate_high_elt_extract_vector(<2 x i16> addrspace(1)* nocapture readonly %arg, <2 x i16> addrspace(1)* nocapture readonly %arg1, <2 x i16> addrspace(1)* nocapture %arg2) local_unnamed_addr {
67 %tmp = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %arg, i64 undef
68 %tmp3 = load <2 x i16>, <2 x i16> addrspace(1)* %tmp, align 4
69 %tmp4 = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %arg1, i64 undef
70 %tmp5 = load <2 x i16>, <2 x i16> addrspace(1)* %tmp4, align 4
71 %tmp6 = sext <2 x i16> %tmp3 to <2 x i32>
72 %tmp7 = sext <2 x i16> %tmp5 to <2 x i32>
73 %tmp8 = extractelement <2 x i32> %tmp6, i64 0
74 %tmp9 = extractelement <2 x i32> %tmp7, i64 0
75 %tmp10 = mul nsw i32 %tmp9, %tmp8
76 %tmp11 = insertelement <2 x i32> undef, i32 %tmp10, i32 0
77 %tmp12 = insertelement <2 x i32> %tmp11, i32 undef, i32 1
78 %tmp13 = lshr <2 x i32> %tmp12, <i32 16, i32 16>
79 %tmp14 = trunc <2 x i32> %tmp13 to <2 x i16>
80 %tmp15 = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %arg2, i64 undef
81 store <2 x i16> %tmp14, <2 x i16> addrspace(1)* %tmp15, align 4
85 ; GCN-LABEL: {{^}}trunc_v2i64_arg_to_v2i16:
86 ; GCN: v_lshlrev_b32_e32 v1, 16, v2
88 ; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
89 ; SI-NEXT: v_or_b32_e32 v0, v0, v1
90 ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v0
92 ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
94 ; GCN-NEXT: s_setpc_b64
95 define <2 x i16> @trunc_v2i64_arg_to_v2i16(<2 x i64> %arg0) #0 {
96 %trunc = trunc <2 x i64> %arg0 to <2 x i16>