199 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=amdgcn -mcpu=verde < %s | FileCheck --check-prefix=GCN %s3; RUN: llc -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX8 %s4; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX9 %s5 6declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone7declare i32 @llvm.amdgcn.workitem.id.y() nounwind readnone8 9define amdgpu_kernel void @anyext_i1_i32(ptr addrspace(1) %out, i32 %cond) #0 {10; GCN-LABEL: anyext_i1_i32:11; GCN: ; %bb.0: ; %entry12; GCN-NEXT: s_load_dword s6, s[4:5], 0xb13; GCN-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x914; GCN-NEXT: s_mov_b32 s3, 0xf00015; GCN-NEXT: s_mov_b32 s2, -116; GCN-NEXT: s_waitcnt lgkmcnt(0)17; GCN-NEXT: s_cmp_lg_u32 s6, 018; GCN-NEXT: s_cselect_b64 s[4:5], -1, 019; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5]20; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 021; GCN-NEXT: s_endpgm22;23; GFX8-LABEL: anyext_i1_i32:24; GFX8: ; %bb.0: ; %entry25; GFX8-NEXT: s_load_dword s6, s[4:5], 0x2c26; GFX8-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2427; GFX8-NEXT: s_mov_b32 s3, 0xf00028; GFX8-NEXT: s_mov_b32 s2, -129; GFX8-NEXT: s_waitcnt lgkmcnt(0)30; GFX8-NEXT: s_cmp_lg_u32 s6, 031; GFX8-NEXT: s_cselect_b64 s[4:5], -1, 032; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5]33; GFX8-NEXT: buffer_store_dword v0, off, s[0:3], 034; GFX8-NEXT: s_endpgm35;36; GFX9-LABEL: anyext_i1_i32:37; GFX9: ; %bb.0: ; %entry38; GFX9-NEXT: s_load_dword s6, s[4:5], 0x2c39; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2440; GFX9-NEXT: s_mov_b32 s3, 0xf00041; GFX9-NEXT: s_mov_b32 s2, -142; GFX9-NEXT: s_waitcnt lgkmcnt(0)43; GFX9-NEXT: s_cmp_lg_u32 s6, 044; GFX9-NEXT: s_cselect_b64 s[4:5], -1, 045; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5]46; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 047; GFX9-NEXT: s_endpgm48entry:49 %tmp = icmp eq i32 %cond, 050 %tmp1 = zext i1 %tmp to i851 %tmp2 = xor i8 %tmp1, -152 %tmp3 = and i8 %tmp2, 153 %tmp4 = zext i8 %tmp3 to i3254 store i32 %tmp4, ptr addrspace(1) %out55 ret void56}57 58define amdgpu_kernel void @s_anyext_i16_i32(ptr addrspace(1) %out, ptr addrspace(1) %a, ptr addrspace(1) %b) #0 {59; GCN-LABEL: s_anyext_i16_i32:60; GCN: ; %bb.0: ; %entry61; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x962; GCN-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0xd63; GCN-NEXT: s_mov_b32 s11, 0xf00064; GCN-NEXT: s_mov_b32 s14, 065; GCN-NEXT: s_mov_b32 s15, s1166; GCN-NEXT: s_waitcnt lgkmcnt(0)67; GCN-NEXT: s_mov_b64 s[12:13], s[2:3]68; GCN-NEXT: v_lshlrev_b32_e32 v2, 1, v069; GCN-NEXT: v_mov_b32_e32 v3, 070; GCN-NEXT: v_lshlrev_b32_e32 v0, 1, v171; GCN-NEXT: s_mov_b64 s[6:7], s[14:15]72; GCN-NEXT: v_mov_b32_e32 v1, v373; GCN-NEXT: buffer_load_ushort v2, v[2:3], s[12:15], 0 addr6474; GCN-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr6475; GCN-NEXT: s_mov_b32 s10, -176; GCN-NEXT: s_mov_b32 s8, s077; GCN-NEXT: s_mov_b32 s9, s178; GCN-NEXT: s_waitcnt vmcnt(0)79; GCN-NEXT: v_add_i32_e32 v0, vcc, v2, v080; GCN-NEXT: v_bfi_b32 v0, v0, 0, 181; GCN-NEXT: buffer_store_dword v0, off, s[8:11], 082; GCN-NEXT: s_endpgm83;84; GFX8-LABEL: s_anyext_i16_i32:85; GFX8: ; %bb.0: ; %entry86; GFX8-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x2487; GFX8-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x3488; GFX8-NEXT: v_lshlrev_b32_e32 v0, 1, v089; GFX8-NEXT: s_waitcnt lgkmcnt(0)90; GFX8-NEXT: v_mov_b32_e32 v3, s391; GFX8-NEXT: v_add_u32_e32 v2, vcc, s2, v092; GFX8-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc93; GFX8-NEXT: v_lshlrev_b32_e32 v0, 1, v194; GFX8-NEXT: v_mov_b32_e32 v1, s595; GFX8-NEXT: v_add_u32_e32 v0, vcc, s4, v096; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc97; GFX8-NEXT: flat_load_ushort v2, v[2:3]98; GFX8-NEXT: flat_load_ushort v0, v[0:1]99; GFX8-NEXT: s_mov_b32 s3, 0xf000100; GFX8-NEXT: s_mov_b32 s2, -1101; GFX8-NEXT: s_waitcnt vmcnt(0)102; GFX8-NEXT: v_add_u16_e32 v0, v2, v0103; GFX8-NEXT: v_xor_b32_e32 v0, -1, v0104; GFX8-NEXT: v_and_b32_e32 v0, 1, v0105; GFX8-NEXT: v_and_b32_e32 v0, 0xffff, v0106; GFX8-NEXT: buffer_store_dword v0, off, s[0:3], 0107; GFX8-NEXT: s_endpgm108;109; GFX9-LABEL: s_anyext_i16_i32:110; GFX9: ; %bb.0: ; %entry111; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24112; GFX9-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34113; GFX9-NEXT: v_lshlrev_b32_e32 v0, 1, v0114; GFX9-NEXT: v_lshlrev_b32_e32 v1, 1, v1115; GFX9-NEXT: s_waitcnt lgkmcnt(0)116; GFX9-NEXT: global_load_ushort v2, v0, s[2:3]117; GFX9-NEXT: global_load_ushort v3, v1, s[6:7]118; GFX9-NEXT: s_mov_b32 s3, 0xf000119; GFX9-NEXT: s_mov_b32 s2, -1120; GFX9-NEXT: s_waitcnt vmcnt(0)121; GFX9-NEXT: v_add_u16_e32 v0, v2, v3122; GFX9-NEXT: v_xor_b32_e32 v0, -1, v0123; GFX9-NEXT: v_and_b32_e32 v0, 1, v0124; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0125; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 0126; GFX9-NEXT: s_endpgm127entry:128 %tid.x = call i32 @llvm.amdgcn.workitem.id.x()129 %tid.y = call i32 @llvm.amdgcn.workitem.id.y()130 %a.ptr = getelementptr i16, ptr addrspace(1) %a, i32 %tid.x131 %b.ptr = getelementptr i16, ptr addrspace(1) %b, i32 %tid.y132 %a.l = load i16, ptr addrspace(1) %a.ptr133 %b.l = load i16, ptr addrspace(1) %b.ptr134 %tmp = add i16 %a.l, %b.l135 %tmp1 = trunc i16 %tmp to i8136 %tmp2 = xor i8 %tmp1, -1137 %tmp3 = and i8 %tmp2, 1138 %tmp4 = zext i8 %tmp3 to i32139 store i32 %tmp4, ptr addrspace(1) %out140 ret void141}142 143define amdgpu_kernel void @anyext_v2i16_to_v2i32() #0 {144; GCN-LABEL: anyext_v2i16_to_v2i32:145; GCN: ; %bb.0: ; %bb146; GCN-NEXT: s_mov_b32 s3, 0xf000147; GCN-NEXT: s_mov_b32 s2, -1148; GCN-NEXT: buffer_load_ushort v0, off, s[0:3], 0149; GCN-NEXT: s_waitcnt vmcnt(0)150; GCN-NEXT: v_and_b32_e32 v0, 0x8001, v0151; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0152; GCN-NEXT: v_cmp_eq_f32_e32 vcc, 0, v0153; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc154; GCN-NEXT: buffer_store_byte v0, off, s[0:3], 0155; GCN-NEXT: s_endpgm156;157; GFX8-LABEL: anyext_v2i16_to_v2i32:158; GFX8: ; %bb.0: ; %bb159; GFX8-NEXT: s_mov_b32 s3, 0xf000160; GFX8-NEXT: s_mov_b32 s2, -1161; GFX8-NEXT: buffer_load_ushort v0, off, s[0:3], 0162; GFX8-NEXT: v_mov_b32_e32 v1, 0x8001163; GFX8-NEXT: s_waitcnt vmcnt(0)164; GFX8-NEXT: v_and_b32_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD165; GFX8-NEXT: v_cmp_eq_f32_e32 vcc, 0, v0166; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc167; GFX8-NEXT: buffer_store_byte v0, off, s[0:3], 0168; GFX8-NEXT: s_endpgm169;170; GFX9-LABEL: anyext_v2i16_to_v2i32:171; GFX9: ; %bb.0: ; %bb172; GFX9-NEXT: global_load_short_d16_hi v0, v[0:1], off173; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff174; GFX9-NEXT: s_mov_b32 s3, 0xf000175; GFX9-NEXT: s_mov_b32 s2, -1176; GFX9-NEXT: s_waitcnt vmcnt(0)177; GFX9-NEXT: v_and_b32_e32 v0, 0x80018001, v0178; GFX9-NEXT: v_bfi_b32 v0, v1, 0, v0179; GFX9-NEXT: v_cmp_eq_f32_e32 vcc, 0, v0180; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc181; GFX9-NEXT: buffer_store_byte v0, off, s[0:3], 0182; GFX9-NEXT: s_endpgm183bb:184 %tmp = load i16, ptr addrspace(1) poison, align 2185 %tmp2 = insertelement <2 x i16> poison, i16 %tmp, i32 1186 %tmp4 = and <2 x i16> %tmp2, <i16 -32767, i16 -32767>187 %tmp5 = zext <2 x i16> %tmp4 to <2 x i32>188 %tmp6 = shl nuw <2 x i32> %tmp5, <i32 16, i32 16>189 %tmp7 = or <2 x i32> zeroinitializer, %tmp6190 %tmp8 = bitcast <2 x i32> %tmp7 to <2 x float>191 %tmp10 = fcmp oeq <2 x float> %tmp8, zeroinitializer192 %tmp11 = zext <2 x i1> %tmp10 to <2 x i8>193 %tmp12 = extractelement <2 x i8> %tmp11, i32 1194 store i8 %tmp12, ptr addrspace(1) poison, align 1195 ret void196}197 198attributes #0 = { nounwind }199