139 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefix=GCN %s3 4; Test that when extracting the same unknown vector index from an5; insertelement the dynamic indexing is folded away.6 7declare i32 @llvm.amdgcn.workitem.id.x() #08 9; No dynamic indexing required10define amdgpu_kernel void @extract_insert_same_dynelt_v4i32(ptr addrspace(1) %out, ptr addrspace(1) %in, i32 %val, i32 %idx) #1 {11; GCN-LABEL: extract_insert_same_dynelt_v4i32:12; GCN: ; %bb.0:13; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x914; GCN-NEXT: s_load_dword s4, s[4:5], 0xd15; GCN-NEXT: s_waitcnt lgkmcnt(0)16; GCN-NEXT: s_mov_b32 s3, 0xf00017; GCN-NEXT: s_mov_b32 s2, 018; GCN-NEXT: v_lshlrev_b32_e32 v0, 2, v019; GCN-NEXT: v_mov_b32_e32 v1, 020; GCN-NEXT: v_mov_b32_e32 v2, s421; GCN-NEXT: buffer_store_dword v2, v[0:1], s[0:3], 0 addr6422; GCN-NEXT: s_endpgm23 %id = call i32 @llvm.amdgcn.workitem.id.x()24 %id.ext = sext i32 %id to i6425 %gep.in = getelementptr inbounds <4 x i32>, ptr addrspace(1) %in, i64 %id.ext26 %gep.out = getelementptr inbounds i32, ptr addrspace(1) %out, i64 %id.ext27 %vec = load <4 x i32>, ptr addrspace(1) %gep.in28 %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx29 %extract = extractelement <4 x i32> %insert, i32 %idx30 store i32 %extract, ptr addrspace(1) %gep.out31 ret void32}33 34define amdgpu_kernel void @extract_insert_different_dynelt_v4i32(ptr addrspace(1) %out, ptr addrspace(1) %in, i32 %val, i32 %idx0, i32 %idx1) #1 {35; GCN-LABEL: extract_insert_different_dynelt_v4i32:36; GCN: ; %bb.0:37; GCN-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x938; GCN-NEXT: s_mov_b32 s3, 0xf00039; GCN-NEXT: s_mov_b32 s2, 040; GCN-NEXT: v_lshlrev_b32_e32 v4, 4, v041; GCN-NEXT: v_mov_b32_e32 v5, 042; GCN-NEXT: s_waitcnt lgkmcnt(0)43; GCN-NEXT: s_mov_b64 s[0:1], s[10:11]44; GCN-NEXT: buffer_load_dwordx4 v[1:4], v[4:5], s[0:3], 0 addr6445; GCN-NEXT: s_cmp_eq_u32 s13, 346; GCN-NEXT: s_cselect_b64 vcc, -1, 047; GCN-NEXT: s_cmp_eq_u32 s13, 248; GCN-NEXT: s_cselect_b64 s[0:1], -1, 049; GCN-NEXT: s_cmp_eq_u32 s13, 150; GCN-NEXT: s_mov_b64 s[10:11], s[2:3]51; GCN-NEXT: s_cselect_b64 s[2:3], -1, 052; GCN-NEXT: s_cmp_eq_u32 s13, 053; GCN-NEXT: v_lshlrev_b32_e32 v6, 2, v054; GCN-NEXT: v_mov_b32_e32 v0, s1255; GCN-NEXT: s_cselect_b64 s[4:5], -1, 056; GCN-NEXT: s_cmp_eq_u32 s14, 157; GCN-NEXT: v_mov_b32_e32 v7, v558; GCN-NEXT: s_waitcnt vmcnt(0)59; GCN-NEXT: v_cndmask_b32_e32 v4, v4, v0, vcc60; GCN-NEXT: v_cndmask_b32_e64 v3, v3, v0, s[0:1]61; GCN-NEXT: v_cndmask_b32_e64 v2, v2, v0, s[2:3]62; GCN-NEXT: v_cndmask_b32_e64 v0, v1, v0, s[4:5]63; GCN-NEXT: s_cselect_b64 vcc, -1, 064; GCN-NEXT: s_cmp_eq_u32 s14, 265; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc66; GCN-NEXT: s_cselect_b64 vcc, -1, 067; GCN-NEXT: s_cmp_eq_u32 s14, 368; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc69; GCN-NEXT: s_cselect_b64 vcc, -1, 070; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc71; GCN-NEXT: buffer_store_dword v0, v[6:7], s[8:11], 0 addr6472; GCN-NEXT: s_endpgm73 %id = call i32 @llvm.amdgcn.workitem.id.x()74 %id.ext = sext i32 %id to i6475 %gep.in = getelementptr inbounds <4 x i32>, ptr addrspace(1) %in, i64 %id.ext76 %gep.out = getelementptr inbounds i32, ptr addrspace(1) %out, i64 %id.ext77 %vec = load <4 x i32>, ptr addrspace(1) %gep.in78 %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx079 %extract = extractelement <4 x i32> %insert, i32 %idx180 store i32 %extract, ptr addrspace(1) %gep.out81 ret void82}83 84define amdgpu_kernel void @extract_insert_same_elt2_v4i32(ptr addrspace(1) %out, ptr addrspace(1) %in, i32 %val, i32 %idx) #1 {85; GCN-LABEL: extract_insert_same_elt2_v4i32:86; GCN: ; %bb.0:87; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x988; GCN-NEXT: s_load_dword s4, s[4:5], 0xd89; GCN-NEXT: s_waitcnt lgkmcnt(0)90; GCN-NEXT: s_mov_b32 s3, 0xf00091; GCN-NEXT: s_mov_b32 s2, 092; GCN-NEXT: v_lshlrev_b32_e32 v0, 2, v093; GCN-NEXT: v_mov_b32_e32 v1, 094; GCN-NEXT: v_mov_b32_e32 v2, s495; GCN-NEXT: buffer_store_dword v2, v[0:1], s[0:3], 0 addr6496; GCN-NEXT: s_endpgm97 %id = call i32 @llvm.amdgcn.workitem.id.x()98 %id.ext = sext i32 %id to i6499 %gep.in = getelementptr inbounds <4 x i32>, ptr addrspace(1) %in, i64 %id.ext100 %gep.out = getelementptr inbounds i32, ptr addrspace(1) %out, i64 %id.ext101 %vec = load <4 x i32>, ptr addrspace(1) %gep.in102 %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx103 %extract = extractelement <4 x i32> %insert, i32 %idx104 store i32 %extract, ptr addrspace(1) %gep.out105 ret void106}107 108define amdgpu_kernel void @extract_insert_same_dynelt_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %in, float %val, i32 %idx) #1 {109; GCN-LABEL: extract_insert_same_dynelt_v4f32:110; GCN: ; %bb.0:111; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9112; GCN-NEXT: s_load_dword s8, s[4:5], 0xd113; GCN-NEXT: s_mov_b32 s7, 0xf000114; GCN-NEXT: s_mov_b32 s6, 0115; GCN-NEXT: v_lshlrev_b32_e32 v4, 4, v0116; GCN-NEXT: s_waitcnt lgkmcnt(0)117; GCN-NEXT: s_mov_b64 s[4:5], s[2:3]118; GCN-NEXT: v_mov_b32_e32 v5, 0119; GCN-NEXT: buffer_load_dwordx4 v[1:4], v[4:5], s[4:7], 0 addr64 glc120; GCN-NEXT: s_waitcnt vmcnt(0)121; GCN-NEXT: s_mov_b64 s[2:3], s[6:7]122; GCN-NEXT: v_lshlrev_b32_e32 v4, 2, v0123; GCN-NEXT: v_mov_b32_e32 v0, s8124; GCN-NEXT: buffer_store_dword v0, v[4:5], s[0:3], 0 addr64125; GCN-NEXT: s_endpgm126 %id = call i32 @llvm.amdgcn.workitem.id.x()127 %id.ext = sext i32 %id to i64128 %gep.in = getelementptr inbounds <4 x float>, ptr addrspace(1) %in, i64 %id.ext129 %gep.out = getelementptr inbounds float, ptr addrspace(1) %out, i64 %id.ext130 %vec = load volatile <4 x float>, ptr addrspace(1) %gep.in131 %insert = insertelement <4 x float> %vec, float %val, i32 %idx132 %extract = extractelement <4 x float> %insert, i32 %idx133 store float %extract, ptr addrspace(1) %gep.out134 ret void135}136 137attributes #0 = { nounwind readnone }138attributes #1 = { nounwind }139