1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -march=amdgcn -mcpu=tahiti -verify-machineinstrs | FileCheck %s -enable-var-scope -check-prefixes=GCN,SI
3; RUN: llc < %s -march=amdgcn -mcpu=fiji -verify-machineinstrs | FileCheck %s -enable-var-scope -check-prefixes=GCN,VI
4
5; Make sure high constant 0 isn't pointlessly materialized
6define i16 @trunc_bitcast_i64_lshr_32_i16(i64 %bar) {
7; GCN-LABEL: trunc_bitcast_i64_lshr_32_i16:
8; GCN:       ; %bb.0:
9; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
10; GCN-NEXT:    v_mov_b32_e32 v0, v1
11; GCN-NEXT:    s_setpc_b64 s[30:31]
12  %srl = lshr i64 %bar, 32
13  %trunc = trunc i64 %srl to i16
14  ret i16 %trunc
15}
16
17define i32 @trunc_bitcast_i64_lshr_32_i32(i64 %bar) {
18; GCN-LABEL: trunc_bitcast_i64_lshr_32_i32:
19; GCN:       ; %bb.0:
20; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
21; GCN-NEXT:    v_mov_b32_e32 v0, v1
22; GCN-NEXT:    s_setpc_b64 s[30:31]
23  %srl = lshr i64 %bar, 32
24  %trunc = trunc i64 %srl to i32
25  ret i32 %trunc
26}
27
28define i16 @trunc_bitcast_v2i32_to_i16(<2 x i32> %bar) {
29; SI-LABEL: trunc_bitcast_v2i32_to_i16:
30; SI:       ; %bb.0:
31; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
32; SI-NEXT:    s_mov_b32 s7, 0xf000
33; SI-NEXT:    s_mov_b32 s6, -1
34; SI-NEXT:    buffer_load_dword v0, off, s[4:7], 0
35; SI-NEXT:    s_waitcnt vmcnt(0)
36; SI-NEXT:    v_add_i32_e32 v0, vcc, 4, v0
37; SI-NEXT:    s_setpc_b64 s[30:31]
38;
39; VI-LABEL: trunc_bitcast_v2i32_to_i16:
40; VI:       ; %bb.0:
41; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
42; VI-NEXT:    flat_load_dword v0, v[0:1]
43; VI-NEXT:    s_waitcnt vmcnt(0)
44; VI-NEXT:    v_add_u16_e32 v0, 4, v0
45; VI-NEXT:    s_setpc_b64 s[30:31]
46  %load0 = load i32, i32 addrspace(1)* undef
47  %load1 = load i32, i32 addrspace(1)* null
48  %insert.0 = insertelement <2 x i32> undef, i32 %load0, i32 0
49  %insert.1 = insertelement <2 x i32> %insert.0, i32 99, i32 1
50  %bc = bitcast <2 x i32> %insert.1 to i64
51  %trunc = trunc i64 %bc to i16
52  %add = add i16 %trunc, 4
53  ret i16 %add
54}
55
56; Make sure there's no crash if the source vector type is FP
57define i16 @trunc_bitcast_v2f32_to_i16(<2 x float> %bar) {
58; SI-LABEL: trunc_bitcast_v2f32_to_i16:
59; SI:       ; %bb.0:
60; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
61; SI-NEXT:    s_mov_b32 s7, 0xf000
62; SI-NEXT:    s_mov_b32 s6, -1
63; SI-NEXT:    buffer_load_dword v0, off, s[4:7], 0
64; SI-NEXT:    s_waitcnt vmcnt(0)
65; SI-NEXT:    v_add_i32_e32 v0, vcc, 4, v0
66; SI-NEXT:    s_setpc_b64 s[30:31]
67;
68; VI-LABEL: trunc_bitcast_v2f32_to_i16:
69; VI:       ; %bb.0:
70; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
71; VI-NEXT:    flat_load_dword v0, v[0:1]
72; VI-NEXT:    s_waitcnt vmcnt(0)
73; VI-NEXT:    v_add_u16_e32 v0, 4, v0
74; VI-NEXT:    s_setpc_b64 s[30:31]
75  %load0 = load float, float addrspace(1)* undef
76  %load1 = load float, float addrspace(1)* null
77  %insert.0 = insertelement <2 x float> undef, float %load0, i32 0
78  %insert.1 = insertelement <2 x float> %insert.0, float 4.0, i32 1
79  %bc = bitcast <2 x float> %insert.1 to i64
80  %trunc = trunc i64 %bc to i16
81  %add = add i16 %trunc, 4
82  ret i16 %add
83}
84
85define 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 {
86; SI-LABEL: truncate_high_elt_extract_vector:
87; SI:       ; %bb.0: ; %bb
88; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
89; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
90; SI-NEXT:    s_mov_b32 s3, 0xf000
91; SI-NEXT:    s_mov_b32 s2, -1
92; SI-NEXT:    s_waitcnt lgkmcnt(0)
93; SI-NEXT:    s_load_dword s4, s[4:5], 0x0
94; SI-NEXT:    s_load_dword s5, s[6:7], 0x0
95; SI-NEXT:    s_waitcnt lgkmcnt(0)
96; SI-NEXT:    s_sext_i32_i16 s4, s4
97; SI-NEXT:    s_sext_i32_i16 s5, s5
98; SI-NEXT:    v_mov_b32_e32 v0, s4
99; SI-NEXT:    v_mul_i32_i24_e32 v0, s5, v0
100; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
101; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
102; SI-NEXT:    s_endpgm
103;
104; VI-LABEL: truncate_high_elt_extract_vector:
105; VI:       ; %bb.0: ; %bb
106; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
107; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
108; VI-NEXT:    s_waitcnt lgkmcnt(0)
109; VI-NEXT:    s_load_dword s2, s[4:5], 0x0
110; VI-NEXT:    s_load_dword s3, s[6:7], 0x0
111; VI-NEXT:    v_mov_b32_e32 v0, s0
112; VI-NEXT:    v_mov_b32_e32 v1, s1
113; VI-NEXT:    s_waitcnt lgkmcnt(0)
114; VI-NEXT:    s_sext_i32_i16 s0, s2
115; VI-NEXT:    s_sext_i32_i16 s1, s3
116; VI-NEXT:    v_mov_b32_e32 v2, s0
117; VI-NEXT:    v_mul_i32_i24_e32 v2, s1, v2
118; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
119; VI-NEXT:    flat_store_dword v[0:1], v2
120; VI-NEXT:    s_endpgm
121bb:
122  %tmp = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %arg, i64 undef
123  %tmp3 = load <2 x i16>, <2 x i16> addrspace(1)* %tmp, align 4
124  %tmp4 = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %arg1, i64 undef
125  %tmp5 = load <2 x i16>, <2 x i16> addrspace(1)* %tmp4, align 4
126  %tmp6 = sext <2 x i16> %tmp3 to <2 x i32>
127  %tmp7 = sext <2 x i16> %tmp5 to <2 x i32>
128  %tmp8 = extractelement <2 x i32> %tmp6, i64 0
129  %tmp9 = extractelement <2 x i32> %tmp7, i64 0
130  %tmp10 = mul nsw i32 %tmp9, %tmp8
131  %tmp11 = insertelement <2 x i32> undef, i32 %tmp10, i32 0
132  %tmp12 = insertelement <2 x i32> %tmp11, i32 undef, i32 1
133  %tmp13 = lshr <2 x i32> %tmp12, <i32 16, i32 16>
134  %tmp14 = trunc <2 x i32> %tmp13 to <2 x i16>
135  %tmp15 = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %arg2, i64 undef
136  store <2 x i16> %tmp14, <2 x i16> addrspace(1)* %tmp15, align 4
137  ret void
138}
139
140define <2 x i16> @trunc_v2i64_arg_to_v2i16(<2 x i64> %arg0) #0 {
141; SI-LABEL: trunc_v2i64_arg_to_v2i16:
142; SI:       ; %bb.0:
143; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
144; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
145; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
146; SI-NEXT:    v_or_b32_e32 v0, v0, v1
147; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
148; SI-NEXT:    s_setpc_b64 s[30:31]
149;
150; VI-LABEL: trunc_v2i64_arg_to_v2i16:
151; VI:       ; %bb.0:
152; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
153; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
154; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
155; VI-NEXT:    s_setpc_b64 s[30:31]
156  %trunc = trunc <2 x i64> %arg0 to <2 x i16>
157  ret <2 x i16> %trunc
158}
159