187 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a %s -o - | FileCheck %s3 4; FIXME: We need to test AS6 but the AS6 variants of the following tests fail because of illegal VGPR to SGPR copy.5; FIXME: We also want to test memset, memcpy, and memmove, but it needs to fix the SelectionDAG store merging issue (#90714).6 7define amdgpu_kernel void @store_as4_i8(ptr addrspace(4) %p, i8 %v) {8; CHECK-LABEL: store_as4_i8:9; CHECK: ; %bb.0:10; CHECK-NEXT: s_load_dword s2, s[8:9], 0x811; CHECK-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x012; CHECK-NEXT: v_mov_b32_e32 v0, 013; CHECK-NEXT: s_waitcnt lgkmcnt(0)14; CHECK-NEXT: v_mov_b32_e32 v1, s215; CHECK-NEXT: global_store_byte v0, v1, s[0:1]16; CHECK-NEXT: s_endpgm17 store i8 %v, ptr addrspace(4) %p18 ret void19}20 21define amdgpu_kernel void @store_as4_i16(ptr addrspace(4) %p, i16 %v) {22; CHECK-LABEL: store_as4_i16:23; CHECK: ; %bb.0:24; CHECK-NEXT: s_load_dword s2, s[8:9], 0x825; CHECK-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x026; CHECK-NEXT: v_mov_b32_e32 v0, 027; CHECK-NEXT: s_waitcnt lgkmcnt(0)28; CHECK-NEXT: v_mov_b32_e32 v1, s229; CHECK-NEXT: global_store_short v0, v1, s[0:1]30; CHECK-NEXT: s_endpgm31 store i16 %v, ptr addrspace(4) %p32 ret void33}34 35define amdgpu_kernel void @store_as4_i32(ptr addrspace(4) %p, i32 %v) {36; CHECK-LABEL: store_as4_i32:37; CHECK: ; %bb.0:38; CHECK-NEXT: s_load_dword s2, s[8:9], 0x839; CHECK-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x040; CHECK-NEXT: v_mov_b32_e32 v0, 041; CHECK-NEXT: s_waitcnt lgkmcnt(0)42; CHECK-NEXT: v_mov_b32_e32 v1, s243; CHECK-NEXT: global_store_dword v0, v1, s[0:1]44; CHECK-NEXT: s_endpgm45 store i32 %v, ptr addrspace(4) %p46 ret void47}48 49define amdgpu_kernel void @store_as4_i64(ptr addrspace(4) %p, i64 %v) {50; CHECK-LABEL: store_as4_i64:51; CHECK: ; %bb.0:52; CHECK-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x053; CHECK-NEXT: v_mov_b32_e32 v0, 054; CHECK-NEXT: s_waitcnt lgkmcnt(0)55; CHECK-NEXT: v_mov_b32_e32 v2, s256; CHECK-NEXT: v_mov_b32_e32 v3, s357; CHECK-NEXT: global_store_dwordx2 v0, v[2:3], s[0:1]58; CHECK-NEXT: s_endpgm59 store i64 %v, ptr addrspace(4) %p60 ret void61}62 63define amdgpu_kernel void @store_as4_float(ptr addrspace(4) %p, float %v) {64; CHECK-LABEL: store_as4_float:65; CHECK: ; %bb.0:66; CHECK-NEXT: s_load_dword s2, s[8:9], 0x867; CHECK-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x068; CHECK-NEXT: v_mov_b32_e32 v0, 069; CHECK-NEXT: s_waitcnt lgkmcnt(0)70; CHECK-NEXT: v_mov_b32_e32 v1, s271; CHECK-NEXT: global_store_dword v0, v1, s[0:1]72; CHECK-NEXT: s_endpgm73 store float %v, ptr addrspace(4) %p74 ret void75}76 77define amdgpu_kernel void @store_as4_double(ptr addrspace(4) %p, double %v) {78; CHECK-LABEL: store_as4_double:79; CHECK: ; %bb.0:80; CHECK-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x081; CHECK-NEXT: v_mov_b32_e32 v0, 082; CHECK-NEXT: s_waitcnt lgkmcnt(0)83; CHECK-NEXT: v_mov_b32_e32 v2, s284; CHECK-NEXT: v_mov_b32_e32 v3, s385; CHECK-NEXT: global_store_dwordx2 v0, v[2:3], s[0:1]86; CHECK-NEXT: s_endpgm87 store double %v, ptr addrspace(4) %p88 ret void89}90 91define amdgpu_kernel void @store_as4_half(ptr addrspace(4) %p, half %v) {92; CHECK-LABEL: store_as4_half:93; CHECK: ; %bb.0:94; CHECK-NEXT: s_load_dword s2, s[8:9], 0x895; CHECK-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x096; CHECK-NEXT: v_mov_b32_e32 v0, 097; CHECK-NEXT: s_waitcnt lgkmcnt(0)98; CHECK-NEXT: v_mov_b32_e32 v1, s299; CHECK-NEXT: global_store_short v0, v1, s[0:1]100; CHECK-NEXT: s_endpgm101 store half %v, ptr addrspace(4) %p102 ret void103}104 105define amdgpu_kernel void @store_as4_2xi8(ptr addrspace(4) %p, <2 x i8> %v) {106; CHECK-LABEL: store_as4_2xi8:107; CHECK: ; %bb.0:108; CHECK-NEXT: s_load_dword s2, s[8:9], 0x8109; CHECK-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x0110; CHECK-NEXT: v_mov_b32_e32 v0, 0111; CHECK-NEXT: s_waitcnt lgkmcnt(0)112; CHECK-NEXT: v_mov_b32_e32 v1, s2113; CHECK-NEXT: global_store_short v0, v1, s[0:1]114; CHECK-NEXT: s_endpgm115 store <2 x i8> %v, ptr addrspace(4) %p116 ret void117}118 119define amdgpu_kernel void @store_as4_2xi16(ptr addrspace(4) %p, <2 x i16> %v) {120; CHECK-LABEL: store_as4_2xi16:121; CHECK: ; %bb.0:122; CHECK-NEXT: s_load_dword s2, s[8:9], 0x8123; CHECK-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x0124; CHECK-NEXT: v_mov_b32_e32 v0, 0125; CHECK-NEXT: s_waitcnt lgkmcnt(0)126; CHECK-NEXT: v_mov_b32_e32 v1, s2127; CHECK-NEXT: global_store_dword v0, v1, s[0:1]128; CHECK-NEXT: s_endpgm129 store <2 x i16> %v, ptr addrspace(4) %p130 ret void131}132 133define amdgpu_kernel void @store_as4_2xi32(ptr addrspace(4) %p, <2 x i32> %v) {134; CHECK-LABEL: store_as4_2xi32:135; CHECK: ; %bb.0:136; CHECK-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0137; CHECK-NEXT: v_mov_b32_e32 v0, 0138; CHECK-NEXT: s_waitcnt lgkmcnt(0)139; CHECK-NEXT: v_pk_mov_b32 v[2:3], s[2:3], s[2:3] op_sel:[0,1]140; CHECK-NEXT: global_store_dwordx2 v0, v[2:3], s[0:1]141; CHECK-NEXT: s_endpgm142 store <2 x i32> %v, ptr addrspace(4) %p143 ret void144}145 146define amdgpu_kernel void @store_as4_2xhalf(ptr addrspace(4) %p, <2 x half> %v) {147; CHECK-LABEL: store_as4_2xhalf:148; CHECK: ; %bb.0:149; CHECK-NEXT: s_load_dword s2, s[8:9], 0x8150; CHECK-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x0151; CHECK-NEXT: v_mov_b32_e32 v0, 0152; CHECK-NEXT: s_waitcnt lgkmcnt(0)153; CHECK-NEXT: v_mov_b32_e32 v1, s2154; CHECK-NEXT: global_store_dword v0, v1, s[0:1]155; CHECK-NEXT: s_endpgm156 store <2 x half> %v, ptr addrspace(4) %p157 ret void158}159 160define amdgpu_kernel void @store_as4_2xfloat(ptr addrspace(4) %p, <2 x float> %v) {161; CHECK-LABEL: store_as4_2xfloat:162; CHECK: ; %bb.0:163; CHECK-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0164; CHECK-NEXT: v_mov_b32_e32 v0, 0165; CHECK-NEXT: s_waitcnt lgkmcnt(0)166; CHECK-NEXT: v_pk_mov_b32 v[2:3], s[2:3], s[2:3] op_sel:[0,1]167; CHECK-NEXT: global_store_dwordx2 v0, v[2:3], s[0:1]168; CHECK-NEXT: s_endpgm169 store <2 x float> %v, ptr addrspace(4) %p170 ret void171}172 173define amdgpu_kernel void @store_as4_2xdouble(ptr addrspace(4) %p, <2 x double> %v) {174; CHECK-LABEL: store_as4_2xdouble:175; CHECK: ; %bb.0:176; CHECK-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x10177; CHECK-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0178; CHECK-NEXT: v_mov_b32_e32 v0, 0179; CHECK-NEXT: s_waitcnt lgkmcnt(0)180; CHECK-NEXT: v_pk_mov_b32 v[4:5], s[2:3], s[2:3] op_sel:[0,1]181; CHECK-NEXT: v_pk_mov_b32 v[2:3], s[0:1], s[0:1] op_sel:[0,1]182; CHECK-NEXT: global_store_dwordx4 v0, v[2:5], s[4:5]183; CHECK-NEXT: s_endpgm184 store <2 x double> %v, ptr addrspace(4) %p185 ret void186}187