1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,CIVI,HAWAII %s 3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,CIVI,FIJI %s 4; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX9 %s 5 6define void @local_store_i56(i56 addrspace(3)* %ptr, i56 %arg) #0 { 7; CIVI-LABEL: local_store_i56: 8; CIVI: ; %bb.0: 9; CIVI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 10; CIVI-NEXT: s_mov_b32 m0, -1 11; CIVI-NEXT: ds_write_b16 v0, v2 offset:4 12; CIVI-NEXT: ds_write_b32 v0, v1 13; CIVI-NEXT: v_lshrrev_b32_e32 v1, 16, v2 14; CIVI-NEXT: ds_write_b8 v0, v1 offset:6 15; CIVI-NEXT: s_waitcnt lgkmcnt(0) 16; CIVI-NEXT: s_setpc_b64 s[30:31] 17; 18; GFX9-LABEL: local_store_i56: 19; GFX9: ; %bb.0: 20; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 21; GFX9-NEXT: ds_write_b8_d16_hi v0, v2 offset:6 22; GFX9-NEXT: ds_write_b16 v0, v2 offset:4 23; GFX9-NEXT: ds_write_b32 v0, v1 24; GFX9-NEXT: s_waitcnt lgkmcnt(0) 25; GFX9-NEXT: s_setpc_b64 s[30:31] 26 store i56 %arg, i56 addrspace(3)* %ptr, align 8 27 ret void 28} 29 30define amdgpu_kernel void @local_store_i55(i55 addrspace(3)* %ptr, i55 %arg) #0 { 31; HAWAII-LABEL: local_store_i55: 32; HAWAII: ; %bb.0: 33; HAWAII-NEXT: s_or_b32 s0, s4, 14 34; HAWAII-NEXT: v_mov_b32_e32 v0, s0 35; HAWAII-NEXT: v_mov_b32_e32 v1, s5 36; HAWAII-NEXT: flat_load_ubyte v0, v[0:1] 37; HAWAII-NEXT: s_load_dword s2, s[4:5], 0x3 38; HAWAII-NEXT: s_load_dword s0, s[4:5], 0x0 39; HAWAII-NEXT: s_load_dword s1, s[4:5], 0x2 40; HAWAII-NEXT: s_mov_b32 m0, -1 41; HAWAII-NEXT: s_waitcnt lgkmcnt(0) 42; HAWAII-NEXT: v_mov_b32_e32 v1, s0 43; HAWAII-NEXT: v_mov_b32_e32 v3, s2 44; HAWAII-NEXT: v_mov_b32_e32 v2, s1 45; HAWAII-NEXT: ds_write_b16 v1, v3 offset:4 46; HAWAII-NEXT: s_waitcnt vmcnt(0) 47; HAWAII-NEXT: v_and_b32_e32 v0, 0x7f, v0 48; HAWAII-NEXT: ds_write_b8 v1, v0 offset:6 49; HAWAII-NEXT: ds_write_b32 v1, v2 50; HAWAII-NEXT: s_endpgm 51; 52; FIJI-LABEL: local_store_i55: 53; FIJI: ; %bb.0: 54; FIJI-NEXT: s_or_b32 s0, s4, 14 55; FIJI-NEXT: v_mov_b32_e32 v0, s0 56; FIJI-NEXT: v_mov_b32_e32 v1, s5 57; FIJI-NEXT: flat_load_ubyte v0, v[0:1] 58; FIJI-NEXT: s_load_dword s0, s[4:5], 0x0 59; FIJI-NEXT: s_load_dword s1, s[4:5], 0x8 60; FIJI-NEXT: s_load_dword s2, s[4:5], 0xc 61; FIJI-NEXT: s_mov_b32 m0, -1 62; FIJI-NEXT: s_waitcnt lgkmcnt(0) 63; FIJI-NEXT: v_mov_b32_e32 v1, s0 64; FIJI-NEXT: v_mov_b32_e32 v3, s1 65; FIJI-NEXT: s_and_b32 s3, s2, 0xffff 66; FIJI-NEXT: v_mov_b32_e32 v2, s2 67; FIJI-NEXT: ds_write_b16 v1, v2 offset:4 68; FIJI-NEXT: s_waitcnt vmcnt(0) 69; FIJI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 70; FIJI-NEXT: v_or_b32_e32 v0, s3, v0 71; FIJI-NEXT: v_bfe_u32 v0, v0, 16, 7 72; FIJI-NEXT: ds_write_b8 v1, v0 offset:6 73; FIJI-NEXT: ds_write_b32 v1, v3 74; FIJI-NEXT: s_endpgm 75; 76; GFX9-LABEL: local_store_i55: 77; GFX9: ; %bb.0: 78; GFX9-NEXT: v_mov_b32_e32 v0, 0 79; GFX9-NEXT: global_load_ubyte_d16_hi v0, v0, s[4:5] offset:14 80; GFX9-NEXT: s_load_dword s2, s[4:5], 0xc 81; GFX9-NEXT: s_load_dword s0, s[4:5], 0x0 82; GFX9-NEXT: s_load_dword s1, s[4:5], 0x8 83; GFX9-NEXT: s_waitcnt lgkmcnt(0) 84; GFX9-NEXT: s_and_b32 s3, s2, 0xffff 85; GFX9-NEXT: v_mov_b32_e32 v1, s0 86; GFX9-NEXT: v_mov_b32_e32 v2, s2 87; GFX9-NEXT: v_mov_b32_e32 v3, s1 88; GFX9-NEXT: ds_write_b16 v1, v2 offset:4 89; GFX9-NEXT: s_waitcnt vmcnt(0) 90; GFX9-NEXT: v_or_b32_e32 v0, s3, v0 91; GFX9-NEXT: v_and_b32_e32 v0, 0x7fffff, v0 92; GFX9-NEXT: ds_write_b8_d16_hi v1, v0 offset:6 93; GFX9-NEXT: ds_write_b32 v1, v3 94; GFX9-NEXT: s_endpgm 95 store i55 %arg, i55 addrspace(3)* %ptr, align 8 96 ret void 97} 98 99define amdgpu_kernel void @local_store_i48(i48 addrspace(3)* %ptr, i48 %arg) #0 { 100; HAWAII-LABEL: local_store_i48: 101; HAWAII: ; %bb.0: 102; HAWAII-NEXT: s_load_dword s0, s[4:5], 0x0 103; HAWAII-NEXT: s_load_dword s1, s[4:5], 0x2 104; HAWAII-NEXT: s_load_dword s2, s[4:5], 0x3 105; HAWAII-NEXT: s_mov_b32 m0, -1 106; HAWAII-NEXT: s_waitcnt lgkmcnt(0) 107; HAWAII-NEXT: v_mov_b32_e32 v0, s0 108; HAWAII-NEXT: v_mov_b32_e32 v1, s2 109; HAWAII-NEXT: ds_write_b16 v0, v1 offset:4 110; HAWAII-NEXT: v_mov_b32_e32 v1, s1 111; HAWAII-NEXT: ds_write_b32 v0, v1 112; HAWAII-NEXT: s_endpgm 113; 114; FIJI-LABEL: local_store_i48: 115; FIJI: ; %bb.0: 116; FIJI-NEXT: s_load_dword s0, s[4:5], 0x0 117; FIJI-NEXT: s_load_dword s1, s[4:5], 0x8 118; FIJI-NEXT: s_load_dword s2, s[4:5], 0xc 119; FIJI-NEXT: s_mov_b32 m0, -1 120; FIJI-NEXT: s_waitcnt lgkmcnt(0) 121; FIJI-NEXT: v_mov_b32_e32 v0, s0 122; FIJI-NEXT: v_mov_b32_e32 v1, s2 123; FIJI-NEXT: ds_write_b16 v0, v1 offset:4 124; FIJI-NEXT: v_mov_b32_e32 v1, s1 125; FIJI-NEXT: ds_write_b32 v0, v1 126; FIJI-NEXT: s_endpgm 127; 128; GFX9-LABEL: local_store_i48: 129; GFX9: ; %bb.0: 130; GFX9-NEXT: s_load_dword s0, s[4:5], 0x0 131; GFX9-NEXT: s_load_dword s1, s[4:5], 0x8 132; GFX9-NEXT: s_load_dword s2, s[4:5], 0xc 133; GFX9-NEXT: s_waitcnt lgkmcnt(0) 134; GFX9-NEXT: v_mov_b32_e32 v0, s0 135; GFX9-NEXT: v_mov_b32_e32 v2, s1 136; GFX9-NEXT: v_mov_b32_e32 v1, s2 137; GFX9-NEXT: ds_write_b16 v0, v1 offset:4 138; GFX9-NEXT: ds_write_b32 v0, v2 139; GFX9-NEXT: s_endpgm 140 store i48 %arg, i48 addrspace(3)* %ptr, align 8 141 ret void 142} 143 144define amdgpu_kernel void @local_store_i65(i65 addrspace(3)* %ptr, i65 %arg) #0 { 145; HAWAII-LABEL: local_store_i65: 146; HAWAII: ; %bb.0: 147; HAWAII-NEXT: s_load_dword s2, s[4:5], 0x0 148; HAWAII-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2 149; HAWAII-NEXT: s_load_dword s3, s[4:5], 0x4 150; HAWAII-NEXT: s_mov_b32 m0, -1 151; HAWAII-NEXT: s_waitcnt lgkmcnt(0) 152; HAWAII-NEXT: v_mov_b32_e32 v2, s2 153; HAWAII-NEXT: s_and_b32 s3, s3, 1 154; HAWAII-NEXT: v_mov_b32_e32 v0, s3 155; HAWAII-NEXT: ds_write_b8 v2, v0 offset:8 156; HAWAII-NEXT: v_mov_b32_e32 v0, s0 157; HAWAII-NEXT: v_mov_b32_e32 v1, s1 158; HAWAII-NEXT: ds_write_b64 v2, v[0:1] 159; HAWAII-NEXT: s_endpgm 160; 161; FIJI-LABEL: local_store_i65: 162; FIJI: ; %bb.0: 163; FIJI-NEXT: s_load_dword s2, s[4:5], 0x0 164; FIJI-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 165; FIJI-NEXT: s_load_dword s3, s[4:5], 0x10 166; FIJI-NEXT: s_mov_b32 m0, -1 167; FIJI-NEXT: s_waitcnt lgkmcnt(0) 168; FIJI-NEXT: v_mov_b32_e32 v2, s2 169; FIJI-NEXT: s_and_b32 s3, s3, 1 170; FIJI-NEXT: v_mov_b32_e32 v0, s3 171; FIJI-NEXT: ds_write_b8 v2, v0 offset:8 172; FIJI-NEXT: v_mov_b32_e32 v0, s0 173; FIJI-NEXT: v_mov_b32_e32 v1, s1 174; FIJI-NEXT: ds_write_b64 v2, v[0:1] 175; FIJI-NEXT: s_endpgm 176; 177; GFX9-LABEL: local_store_i65: 178; GFX9: ; %bb.0: 179; GFX9-NEXT: s_load_dword s2, s[4:5], 0x0 180; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8 181; GFX9-NEXT: s_load_dword s3, s[4:5], 0x10 182; GFX9-NEXT: s_waitcnt lgkmcnt(0) 183; GFX9-NEXT: v_mov_b32_e32 v2, s2 184; GFX9-NEXT: v_mov_b32_e32 v0, s0 185; GFX9-NEXT: s_and_b32 s3, s3, 1 186; GFX9-NEXT: v_mov_b32_e32 v3, s3 187; GFX9-NEXT: v_mov_b32_e32 v1, s1 188; GFX9-NEXT: ds_write_b8 v2, v3 offset:8 189; GFX9-NEXT: ds_write_b64 v2, v[0:1] 190; GFX9-NEXT: s_endpgm 191 store i65 %arg, i65 addrspace(3)* %ptr, align 8 192 ret void 193} 194 195define void @local_store_i13(i13 addrspace(3)* %ptr, i13 %arg) #0 { 196; CIVI-LABEL: local_store_i13: 197; CIVI: ; %bb.0: 198; CIVI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 199; CIVI-NEXT: v_and_b32_e32 v1, 0x1fff, v1 200; CIVI-NEXT: s_mov_b32 m0, -1 201; CIVI-NEXT: ds_write_b16 v0, v1 202; CIVI-NEXT: s_waitcnt lgkmcnt(0) 203; CIVI-NEXT: s_setpc_b64 s[30:31] 204; 205; GFX9-LABEL: local_store_i13: 206; GFX9: ; %bb.0: 207; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 208; GFX9-NEXT: v_and_b32_e32 v1, 0x1fff, v1 209; GFX9-NEXT: ds_write_b16 v0, v1 210; GFX9-NEXT: s_waitcnt lgkmcnt(0) 211; GFX9-NEXT: s_setpc_b64 s[30:31] 212 store i13 %arg, i13 addrspace(3)* %ptr, align 8 213 ret void 214} 215 216define void @local_store_i17(i17 addrspace(3)* %ptr, i17 %arg) #0 { 217; CIVI-LABEL: local_store_i17: 218; CIVI: ; %bb.0: 219; CIVI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 220; CIVI-NEXT: s_mov_b32 m0, -1 221; CIVI-NEXT: ds_write_b16 v0, v1 222; CIVI-NEXT: v_bfe_u32 v1, v1, 16, 1 223; CIVI-NEXT: ds_write_b8 v0, v1 offset:2 224; CIVI-NEXT: s_waitcnt lgkmcnt(0) 225; CIVI-NEXT: s_setpc_b64 s[30:31] 226; 227; GFX9-LABEL: local_store_i17: 228; GFX9: ; %bb.0: 229; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 230; GFX9-NEXT: ds_write_b16 v0, v1 231; GFX9-NEXT: v_and_b32_e32 v1, 0x1ffff, v1 232; GFX9-NEXT: ds_write_b8_d16_hi v0, v1 offset:2 233; GFX9-NEXT: s_waitcnt lgkmcnt(0) 234; GFX9-NEXT: s_setpc_b64 s[30:31] 235 store i17 %arg, i17 addrspace(3)* %ptr, align 8 236 ret void 237} 238 239attributes #0 = { nounwind } 240