1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -global-isel -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s 3; RUN: llc -global-isel -mtriple=amdgcn-amd-amdpal -mcpu=fiji < %s | FileCheck -check-prefixes=GCN,GFX8 %s 4; RUN: llc -global-isel -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefixes=GCN,GFX6 %s 5 6define i64 @zextload_global_i1_to_i64(i1 addrspace(1)* %ptr) { 7; GFX9-LABEL: zextload_global_i1_to_i64: 8; GFX9: ; %bb.0: 9; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 10; GFX9-NEXT: global_load_ubyte v0, v[0:1], off 11; GFX9-NEXT: v_mov_b32_e32 v1, 0 12; GFX9-NEXT: s_waitcnt vmcnt(0) 13; GFX9-NEXT: v_and_b32_e32 v0, 1, v0 14; GFX9-NEXT: s_setpc_b64 s[30:31] 15; 16; GFX8-LABEL: zextload_global_i1_to_i64: 17; GFX8: ; %bb.0: 18; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 19; GFX8-NEXT: flat_load_ubyte v0, v[0:1] 20; GFX8-NEXT: v_mov_b32_e32 v1, 0 21; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 22; GFX8-NEXT: v_and_b32_e32 v0, 1, v0 23; GFX8-NEXT: s_setpc_b64 s[30:31] 24; 25; GFX6-LABEL: zextload_global_i1_to_i64: 26; GFX6: ; %bb.0: 27; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 28; GFX6-NEXT: s_mov_b32 s6, 0 29; GFX6-NEXT: s_mov_b32 s7, 0xf000 30; GFX6-NEXT: s_mov_b64 s[4:5], 0 31; GFX6-NEXT: buffer_load_ubyte v0, v[0:1], s[4:7], 0 addr64 32; GFX6-NEXT: v_mov_b32_e32 v1, 0 33; GFX6-NEXT: s_waitcnt vmcnt(0) 34; GFX6-NEXT: v_and_b32_e32 v0, 1, v0 35; GFX6-NEXT: s_setpc_b64 s[30:31] 36 %load = load i1, i1 addrspace(1)* %ptr 37 %ext = zext i1 %load to i64 38 ret i64 %ext 39} 40 41define i64 @zextload_global_i8_to_i64(i8 addrspace(1)* %ptr) { 42; GFX9-LABEL: zextload_global_i8_to_i64: 43; GFX9: ; %bb.0: 44; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 45; GFX9-NEXT: global_load_ubyte v0, v[0:1], off 46; GFX9-NEXT: v_mov_b32_e32 v1, 0 47; GFX9-NEXT: s_waitcnt vmcnt(0) 48; GFX9-NEXT: s_setpc_b64 s[30:31] 49; 50; GFX8-LABEL: zextload_global_i8_to_i64: 51; GFX8: ; %bb.0: 52; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 53; GFX8-NEXT: flat_load_ubyte v0, v[0:1] 54; GFX8-NEXT: v_mov_b32_e32 v1, 0 55; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 56; GFX8-NEXT: s_setpc_b64 s[30:31] 57; 58; GFX6-LABEL: zextload_global_i8_to_i64: 59; GFX6: ; %bb.0: 60; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 61; GFX6-NEXT: s_mov_b32 s6, 0 62; GFX6-NEXT: s_mov_b32 s7, 0xf000 63; GFX6-NEXT: s_mov_b64 s[4:5], 0 64; GFX6-NEXT: buffer_load_ubyte v0, v[0:1], s[4:7], 0 addr64 65; GFX6-NEXT: v_mov_b32_e32 v1, 0 66; GFX6-NEXT: s_waitcnt vmcnt(0) 67; GFX6-NEXT: s_setpc_b64 s[30:31] 68 %load = load i8, i8 addrspace(1)* %ptr 69 %ext = zext i8 %load to i64 70 ret i64 %ext 71} 72 73define i64 @zextload_global_i16_to_i64(i16 addrspace(1)* %ptr) { 74; GFX9-LABEL: zextload_global_i16_to_i64: 75; GFX9: ; %bb.0: 76; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 77; GFX9-NEXT: global_load_ushort v0, v[0:1], off 78; GFX9-NEXT: v_mov_b32_e32 v1, 0 79; GFX9-NEXT: s_waitcnt vmcnt(0) 80; GFX9-NEXT: s_setpc_b64 s[30:31] 81; 82; GFX8-LABEL: zextload_global_i16_to_i64: 83; GFX8: ; %bb.0: 84; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 85; GFX8-NEXT: flat_load_ushort v0, v[0:1] 86; GFX8-NEXT: v_mov_b32_e32 v1, 0 87; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 88; GFX8-NEXT: s_setpc_b64 s[30:31] 89; 90; GFX6-LABEL: zextload_global_i16_to_i64: 91; GFX6: ; %bb.0: 92; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 93; GFX6-NEXT: s_mov_b32 s6, 0 94; GFX6-NEXT: s_mov_b32 s7, 0xf000 95; GFX6-NEXT: s_mov_b64 s[4:5], 0 96; GFX6-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64 97; GFX6-NEXT: v_mov_b32_e32 v1, 0 98; GFX6-NEXT: s_waitcnt vmcnt(0) 99; GFX6-NEXT: s_setpc_b64 s[30:31] 100 %load = load i16, i16 addrspace(1)* %ptr 101 %ext = zext i16 %load to i64 102 ret i64 %ext 103} 104 105define i64 @zextload_global_i32_to_i64(i32 addrspace(1)* %ptr) { 106; GFX9-LABEL: zextload_global_i32_to_i64: 107; GFX9: ; %bb.0: 108; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 109; GFX9-NEXT: global_load_dword v0, v[0:1], off 110; GFX9-NEXT: v_mov_b32_e32 v1, 0 111; GFX9-NEXT: s_waitcnt vmcnt(0) 112; GFX9-NEXT: s_setpc_b64 s[30:31] 113; 114; GFX8-LABEL: zextload_global_i32_to_i64: 115; GFX8: ; %bb.0: 116; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 117; GFX8-NEXT: flat_load_dword v0, v[0:1] 118; GFX8-NEXT: v_mov_b32_e32 v1, 0 119; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 120; GFX8-NEXT: s_setpc_b64 s[30:31] 121; 122; GFX6-LABEL: zextload_global_i32_to_i64: 123; GFX6: ; %bb.0: 124; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 125; GFX6-NEXT: s_mov_b32 s6, 0 126; GFX6-NEXT: s_mov_b32 s7, 0xf000 127; GFX6-NEXT: s_mov_b64 s[4:5], 0 128; GFX6-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64 129; GFX6-NEXT: v_mov_b32_e32 v1, 0 130; GFX6-NEXT: s_waitcnt vmcnt(0) 131; GFX6-NEXT: s_setpc_b64 s[30:31] 132 %load = load i32, i32 addrspace(1)* %ptr 133 %ext = zext i32 %load to i64 134 ret i64 %ext 135} 136 137define i96 @zextload_global_i32_to_i96(i32 addrspace(1)* %ptr) { 138; GFX9-LABEL: zextload_global_i32_to_i96: 139; GFX9: ; %bb.0: 140; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 141; GFX9-NEXT: global_load_dword v0, v[0:1], off 142; GFX9-NEXT: s_mov_b32 s4, 0 143; GFX9-NEXT: v_mov_b32_e32 v1, 0 144; GFX9-NEXT: v_mov_b32_e32 v2, s4 145; GFX9-NEXT: s_waitcnt vmcnt(0) 146; GFX9-NEXT: s_setpc_b64 s[30:31] 147; 148; GFX8-LABEL: zextload_global_i32_to_i96: 149; GFX8: ; %bb.0: 150; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 151; GFX8-NEXT: flat_load_dword v0, v[0:1] 152; GFX8-NEXT: s_mov_b32 s4, 0 153; GFX8-NEXT: v_mov_b32_e32 v1, 0 154; GFX8-NEXT: v_mov_b32_e32 v2, s4 155; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 156; GFX8-NEXT: s_setpc_b64 s[30:31] 157; 158; GFX6-LABEL: zextload_global_i32_to_i96: 159; GFX6: ; %bb.0: 160; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 161; GFX6-NEXT: s_mov_b32 s6, 0 162; GFX6-NEXT: s_mov_b32 s7, 0xf000 163; GFX6-NEXT: s_mov_b64 s[4:5], 0 164; GFX6-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64 165; GFX6-NEXT: s_mov_b32 s8, 0 166; GFX6-NEXT: v_mov_b32_e32 v1, 0 167; GFX6-NEXT: v_mov_b32_e32 v2, s8 168; GFX6-NEXT: s_waitcnt vmcnt(0) 169; GFX6-NEXT: s_setpc_b64 s[30:31] 170 %load = load i32, i32 addrspace(1)* %ptr 171 %ext = zext i32 %load to i96 172 ret i96 %ext 173} 174 175define i128 @zextload_global_i32_to_i128(i32 addrspace(1)* %ptr) { 176; GFX9-LABEL: zextload_global_i32_to_i128: 177; GFX9: ; %bb.0: 178; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 179; GFX9-NEXT: global_load_dword v0, v[0:1], off 180; GFX9-NEXT: s_mov_b32 s4, 0 181; GFX9-NEXT: s_mov_b32 s5, 0 182; GFX9-NEXT: v_mov_b32_e32 v1, 0 183; GFX9-NEXT: v_mov_b32_e32 v2, s5 184; GFX9-NEXT: v_mov_b32_e32 v3, s4 185; GFX9-NEXT: s_waitcnt vmcnt(0) 186; GFX9-NEXT: s_setpc_b64 s[30:31] 187; 188; GFX8-LABEL: zextload_global_i32_to_i128: 189; GFX8: ; %bb.0: 190; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 191; GFX8-NEXT: flat_load_dword v0, v[0:1] 192; GFX8-NEXT: s_mov_b32 s4, 0 193; GFX8-NEXT: s_mov_b32 s5, 0 194; GFX8-NEXT: v_mov_b32_e32 v1, 0 195; GFX8-NEXT: v_mov_b32_e32 v2, s5 196; GFX8-NEXT: v_mov_b32_e32 v3, s4 197; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) 198; GFX8-NEXT: s_setpc_b64 s[30:31] 199; 200; GFX6-LABEL: zextload_global_i32_to_i128: 201; GFX6: ; %bb.0: 202; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 203; GFX6-NEXT: s_mov_b32 s6, 0 204; GFX6-NEXT: s_mov_b32 s7, 0xf000 205; GFX6-NEXT: s_mov_b64 s[4:5], 0 206; GFX6-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64 207; GFX6-NEXT: s_mov_b32 s8, 0 208; GFX6-NEXT: s_mov_b32 s4, 0 209; GFX6-NEXT: v_mov_b32_e32 v1, 0 210; GFX6-NEXT: v_mov_b32_e32 v2, s4 211; GFX6-NEXT: v_mov_b32_e32 v3, s8 212; GFX6-NEXT: s_waitcnt vmcnt(0) 213; GFX6-NEXT: s_setpc_b64 s[30:31] 214 %load = load i32, i32 addrspace(1)* %ptr 215 %ext = zext i32 %load to i128 216 ret i128 %ext 217} 218