1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefixes=PREGFX10-UNPACKED %s 3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx810 -verify-machineinstrs < %s | FileCheck -check-prefixes=PREGFX10-PACKED %s 4; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=PREGFX10-PACKED %s 5; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-PACKED %s 6; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-PACKED %s 7 8define amdgpu_kernel void @tbuffer_store_d16_x(ptr addrspace(8) %rsrc, half %data, i32 %vindex) { 9; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_x: 10; PREGFX10-UNPACKED: ; %bb.0: ; %main_body 11; PREGFX10-UNPACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 12; PREGFX10-UNPACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 13; PREGFX10-UNPACKED-NEXT: s_waitcnt lgkmcnt(0) 14; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v0, s4 15; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v1, s5 16; PREGFX10-UNPACKED-NEXT: tbuffer_store_format_d16_x v0, v1, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen 17; PREGFX10-UNPACKED-NEXT: s_endpgm 18; 19; PREGFX10-PACKED-LABEL: tbuffer_store_d16_x: 20; PREGFX10-PACKED: ; %bb.0: ; %main_body 21; PREGFX10-PACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 22; PREGFX10-PACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 23; PREGFX10-PACKED-NEXT: s_waitcnt lgkmcnt(0) 24; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v0, s4 25; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v1, s5 26; PREGFX10-PACKED-NEXT: tbuffer_store_format_d16_x v0, v1, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen 27; PREGFX10-PACKED-NEXT: s_endpgm 28; 29; GFX10-PACKED-LABEL: tbuffer_store_d16_x: 30; GFX10-PACKED: ; %bb.0: ; %main_body 31; GFX10-PACKED-NEXT: s_clause 0x1 32; GFX10-PACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 33; GFX10-PACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 34; GFX10-PACKED-NEXT: s_waitcnt lgkmcnt(0) 35; GFX10-PACKED-NEXT: v_mov_b32_e32 v0, s4 36; GFX10-PACKED-NEXT: v_mov_b32_e32 v1, s5 37; GFX10-PACKED-NEXT: tbuffer_store_format_d16_x v0, v1, s[0:3], 0 format:[BUF_FMT_10_11_11_SSCALED] idxen 38; GFX10-PACKED-NEXT: s_endpgm 39; 40; GFX11-PACKED-LABEL: tbuffer_store_d16_x: 41; GFX11-PACKED: ; %bb.0: ; %main_body 42; GFX11-PACKED-NEXT: s_clause 0x1 43; GFX11-PACKED-NEXT: s_load_b64 s[6:7], s[4:5], 0x10 44; GFX11-PACKED-NEXT: s_load_b128 s[0:3], s[4:5], 0x0 45; GFX11-PACKED-NEXT: s_waitcnt lgkmcnt(0) 46; GFX11-PACKED-NEXT: v_mov_b32_e32 v0, s6 47; GFX11-PACKED-NEXT: v_mov_b32_e32 v1, s7 48; GFX11-PACKED-NEXT: tbuffer_store_d16_format_x v0, v1, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM] idxen 49; GFX11-PACKED-NEXT: s_endpgm 50main_body: 51 call void @llvm.amdgcn.struct.ptr.tbuffer.store.f16(half %data, ptr addrspace(8) %rsrc, i32 %vindex, i32 0, i32 0, i32 33, i32 0) 52 ret void 53} 54 55define amdgpu_kernel void @tbuffer_store_d16_xy(ptr addrspace(8) %rsrc, <2 x half> %data, i32 %vindex) { 56; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_xy: 57; PREGFX10-UNPACKED: ; %bb.0: ; %main_body 58; PREGFX10-UNPACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 59; PREGFX10-UNPACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 60; PREGFX10-UNPACKED-NEXT: s_waitcnt lgkmcnt(0) 61; PREGFX10-UNPACKED-NEXT: s_lshr_b32 s6, s4, 16 62; PREGFX10-UNPACKED-NEXT: s_and_b32 s4, s4, 0xffff 63; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v0, s4 64; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v1, s6 65; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v2, s5 66; PREGFX10-UNPACKED-NEXT: tbuffer_store_format_d16_xy v[0:1], v2, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen 67; PREGFX10-UNPACKED-NEXT: s_endpgm 68; 69; PREGFX10-PACKED-LABEL: tbuffer_store_d16_xy: 70; PREGFX10-PACKED: ; %bb.0: ; %main_body 71; PREGFX10-PACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 72; PREGFX10-PACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 73; PREGFX10-PACKED-NEXT: s_waitcnt lgkmcnt(0) 74; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v0, s4 75; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v1, s5 76; PREGFX10-PACKED-NEXT: tbuffer_store_format_d16_xy v0, v1, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen 77; PREGFX10-PACKED-NEXT: s_endpgm 78; 79; GFX10-PACKED-LABEL: tbuffer_store_d16_xy: 80; GFX10-PACKED: ; %bb.0: ; %main_body 81; GFX10-PACKED-NEXT: s_clause 0x1 82; GFX10-PACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 83; GFX10-PACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 84; GFX10-PACKED-NEXT: s_waitcnt lgkmcnt(0) 85; GFX10-PACKED-NEXT: v_mov_b32_e32 v0, s4 86; GFX10-PACKED-NEXT: v_mov_b32_e32 v1, s5 87; GFX10-PACKED-NEXT: tbuffer_store_format_d16_xy v0, v1, s[0:3], 0 format:[BUF_FMT_10_11_11_SSCALED] idxen 88; GFX10-PACKED-NEXT: s_endpgm 89; 90; GFX11-PACKED-LABEL: tbuffer_store_d16_xy: 91; GFX11-PACKED: ; %bb.0: ; %main_body 92; GFX11-PACKED-NEXT: s_clause 0x1 93; GFX11-PACKED-NEXT: s_load_b64 s[6:7], s[4:5], 0x10 94; GFX11-PACKED-NEXT: s_load_b128 s[0:3], s[4:5], 0x0 95; GFX11-PACKED-NEXT: s_waitcnt lgkmcnt(0) 96; GFX11-PACKED-NEXT: v_mov_b32_e32 v0, s6 97; GFX11-PACKED-NEXT: v_mov_b32_e32 v1, s7 98; GFX11-PACKED-NEXT: tbuffer_store_d16_format_xy v0, v1, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM] idxen 99; GFX11-PACKED-NEXT: s_endpgm 100main_body: 101 call void @llvm.amdgcn.struct.ptr.tbuffer.store.v2f16(<2 x half> %data, ptr addrspace(8) %rsrc, i32 %vindex, i32 0, i32 0, i32 33, i32 0) 102 ret void 103} 104 105define amdgpu_kernel void @tbuffer_store_d16_xyz(ptr addrspace(8) %rsrc, <4 x half> %data, i32 %vindex) { 106; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_xyz: 107; PREGFX10-UNPACKED: ; %bb.0: ; %main_body 108; PREGFX10-UNPACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 109; PREGFX10-UNPACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 110; PREGFX10-UNPACKED-NEXT: s_load_dword s6, s[8:9], 0x18 111; PREGFX10-UNPACKED-NEXT: s_waitcnt lgkmcnt(0) 112; PREGFX10-UNPACKED-NEXT: s_and_b32 s5, s5, 0xffff 113; PREGFX10-UNPACKED-NEXT: s_lshr_b32 s7, s4, 16 114; PREGFX10-UNPACKED-NEXT: s_and_b32 s4, s4, 0xffff 115; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v0, s4 116; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v1, s7 117; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v2, s5 118; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v3, s6 119; PREGFX10-UNPACKED-NEXT: tbuffer_store_format_d16_xyz v[0:2], v3, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen 120; PREGFX10-UNPACKED-NEXT: s_endpgm 121; 122; PREGFX10-PACKED-LABEL: tbuffer_store_d16_xyz: 123; PREGFX10-PACKED: ; %bb.0: ; %main_body 124; PREGFX10-PACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 125; PREGFX10-PACKED-NEXT: s_load_dword s6, s[8:9], 0x18 126; PREGFX10-PACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 127; PREGFX10-PACKED-NEXT: s_waitcnt lgkmcnt(0) 128; PREGFX10-PACKED-NEXT: s_and_b32 s5, s5, 0xffff 129; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v0, s4 130; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v1, s5 131; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v2, s6 132; PREGFX10-PACKED-NEXT: tbuffer_store_format_d16_xyz v[0:1], v2, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen 133; PREGFX10-PACKED-NEXT: s_endpgm 134; 135; GFX10-PACKED-LABEL: tbuffer_store_d16_xyz: 136; GFX10-PACKED: ; %bb.0: ; %main_body 137; GFX10-PACKED-NEXT: s_clause 0x2 138; GFX10-PACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 139; GFX10-PACKED-NEXT: s_load_dword s6, s[8:9], 0x18 140; GFX10-PACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 141; GFX10-PACKED-NEXT: s_waitcnt lgkmcnt(0) 142; GFX10-PACKED-NEXT: s_and_b32 s5, s5, 0xffff 143; GFX10-PACKED-NEXT: v_mov_b32_e32 v0, s4 144; GFX10-PACKED-NEXT: v_mov_b32_e32 v1, s5 145; GFX10-PACKED-NEXT: v_mov_b32_e32 v2, s6 146; GFX10-PACKED-NEXT: tbuffer_store_format_d16_xyz v[0:1], v2, s[0:3], 0 format:[BUF_FMT_10_11_11_SSCALED] idxen 147; GFX10-PACKED-NEXT: s_endpgm 148; 149; GFX11-PACKED-LABEL: tbuffer_store_d16_xyz: 150; GFX11-PACKED: ; %bb.0: ; %main_body 151; GFX11-PACKED-NEXT: s_clause 0x2 152; GFX11-PACKED-NEXT: s_load_b64 s[6:7], s[4:5], 0x10 153; GFX11-PACKED-NEXT: s_load_b32 s8, s[4:5], 0x18 154; GFX11-PACKED-NEXT: s_load_b128 s[0:3], s[4:5], 0x0 155; GFX11-PACKED-NEXT: s_waitcnt lgkmcnt(0) 156; GFX11-PACKED-NEXT: s_and_b32 s4, s7, 0xffff 157; GFX11-PACKED-NEXT: v_mov_b32_e32 v0, s6 158; GFX11-PACKED-NEXT: v_mov_b32_e32 v1, s4 159; GFX11-PACKED-NEXT: v_mov_b32_e32 v2, s8 160; GFX11-PACKED-NEXT: tbuffer_store_d16_format_xyz v[0:1], v2, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM] idxen 161; GFX11-PACKED-NEXT: s_endpgm 162main_body: 163 %data_subvec = shufflevector <4 x half> %data, <4 x half> undef, <3 x i32> <i32 0, i32 1, i32 2> 164 call void @llvm.amdgcn.struct.ptr.tbuffer.store.v3f16(<3 x half> %data_subvec, ptr addrspace(8) %rsrc, i32 %vindex, i32 0, i32 0, i32 33, i32 0) 165 ret void 166} 167 168define amdgpu_kernel void @tbuffer_store_d16_xyzw(ptr addrspace(8) %rsrc, <4 x half> %data, i32 %vindex) { 169; PREGFX10-UNPACKED-LABEL: tbuffer_store_d16_xyzw: 170; PREGFX10-UNPACKED: ; %bb.0: ; %main_body 171; PREGFX10-UNPACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 172; PREGFX10-UNPACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 173; PREGFX10-UNPACKED-NEXT: s_load_dword s6, s[8:9], 0x18 174; PREGFX10-UNPACKED-NEXT: s_waitcnt lgkmcnt(0) 175; PREGFX10-UNPACKED-NEXT: s_lshr_b32 s7, s5, 16 176; PREGFX10-UNPACKED-NEXT: s_and_b32 s5, s5, 0xffff 177; PREGFX10-UNPACKED-NEXT: s_lshr_b32 s8, s4, 16 178; PREGFX10-UNPACKED-NEXT: s_and_b32 s4, s4, 0xffff 179; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v0, s4 180; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v1, s8 181; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v2, s5 182; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v3, s7 183; PREGFX10-UNPACKED-NEXT: v_mov_b32_e32 v4, s6 184; PREGFX10-UNPACKED-NEXT: tbuffer_store_format_d16_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen 185; PREGFX10-UNPACKED-NEXT: s_endpgm 186; 187; PREGFX10-PACKED-LABEL: tbuffer_store_d16_xyzw: 188; PREGFX10-PACKED: ; %bb.0: ; %main_body 189; PREGFX10-PACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 190; PREGFX10-PACKED-NEXT: s_load_dword s6, s[8:9], 0x18 191; PREGFX10-PACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 192; PREGFX10-PACKED-NEXT: s_waitcnt lgkmcnt(0) 193; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v0, s4 194; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v1, s5 195; PREGFX10-PACKED-NEXT: v_mov_b32_e32 v2, s6 196; PREGFX10-PACKED-NEXT: tbuffer_store_format_d16_xyzw v[0:1], v2, s[0:3], 0 format:[BUF_NUM_FORMAT_USCALED] idxen 197; PREGFX10-PACKED-NEXT: s_endpgm 198; 199; GFX10-PACKED-LABEL: tbuffer_store_d16_xyzw: 200; GFX10-PACKED: ; %bb.0: ; %main_body 201; GFX10-PACKED-NEXT: s_clause 0x2 202; GFX10-PACKED-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x10 203; GFX10-PACKED-NEXT: s_load_dword s6, s[8:9], 0x18 204; GFX10-PACKED-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 205; GFX10-PACKED-NEXT: s_waitcnt lgkmcnt(0) 206; GFX10-PACKED-NEXT: v_mov_b32_e32 v0, s4 207; GFX10-PACKED-NEXT: v_mov_b32_e32 v1, s5 208; GFX10-PACKED-NEXT: v_mov_b32_e32 v2, s6 209; GFX10-PACKED-NEXT: tbuffer_store_format_d16_xyzw v[0:1], v2, s[0:3], 0 format:[BUF_FMT_10_11_11_SSCALED] idxen 210; GFX10-PACKED-NEXT: s_endpgm 211; 212; GFX11-PACKED-LABEL: tbuffer_store_d16_xyzw: 213; GFX11-PACKED: ; %bb.0: ; %main_body 214; GFX11-PACKED-NEXT: s_clause 0x2 215; GFX11-PACKED-NEXT: s_load_b64 s[6:7], s[4:5], 0x10 216; GFX11-PACKED-NEXT: s_load_b32 s8, s[4:5], 0x18 217; GFX11-PACKED-NEXT: s_load_b128 s[0:3], s[4:5], 0x0 218; GFX11-PACKED-NEXT: s_waitcnt lgkmcnt(0) 219; GFX11-PACKED-NEXT: v_mov_b32_e32 v0, s6 220; GFX11-PACKED-NEXT: v_mov_b32_e32 v1, s7 221; GFX11-PACKED-NEXT: v_mov_b32_e32 v2, s8 222; GFX11-PACKED-NEXT: tbuffer_store_d16_format_xyzw v[0:1], v2, s[0:3], 0 format:[BUF_FMT_10_10_10_2_SNORM] idxen 223; GFX11-PACKED-NEXT: s_endpgm 224main_body: 225 call void @llvm.amdgcn.struct.ptr.tbuffer.store.v4f16(<4 x half> %data, ptr addrspace(8) %rsrc, i32 %vindex, i32 0, i32 0, i32 33, i32 0) 226 ret void 227} 228 229declare void @llvm.amdgcn.struct.ptr.tbuffer.store.f16(half, ptr addrspace(8), i32, i32, i32, i32, i32) 230declare void @llvm.amdgcn.struct.ptr.tbuffer.store.v2f16(<2 x half>, ptr addrspace(8), i32, i32, i32, i32, i32) 231declare void @llvm.amdgcn.struct.ptr.tbuffer.store.v3f16(<3 x half>, ptr addrspace(8), i32, i32, i32, i32, i32) 232declare void @llvm.amdgcn.struct.ptr.tbuffer.store.v4f16(<4 x half>, ptr addrspace(8), i32, i32, i32, i32, i32) 233