179 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=amdgcn-- | FileCheck %s -check-prefix=GCN3 4declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone5 6; Make sure the add and load are reduced to 32-bits even with the7; bitcast to vector.8define amdgpu_kernel void @bitcast_int_to_vector_extract_0(ptr addrspace(1) %out, ptr addrspace(1) %in, i64 %b) {9; GCN-LABEL: bitcast_int_to_vector_extract_0:10; GCN: ; %bb.0:11; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x912; GCN-NEXT: s_load_dword s12, s[4:5], 0xd13; GCN-NEXT: s_mov_b32 s7, 0xf00014; GCN-NEXT: s_mov_b32 s10, 015; GCN-NEXT: v_lshlrev_b32_e32 v0, 3, v016; GCN-NEXT: v_mov_b32_e32 v1, 017; GCN-NEXT: s_mov_b32 s11, s718; GCN-NEXT: s_waitcnt lgkmcnt(0)19; GCN-NEXT: s_mov_b64 s[8:9], s[2:3]20; GCN-NEXT: buffer_load_dword v0, v[0:1], s[8:11], 0 addr6421; GCN-NEXT: s_mov_b32 s6, -122; GCN-NEXT: s_mov_b32 s4, s023; GCN-NEXT: s_mov_b32 s5, s124; GCN-NEXT: s_waitcnt vmcnt(0)25; GCN-NEXT: v_add_i32_e32 v0, vcc, s12, v026; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 027; GCN-NEXT: s_endpgm28 %tid = call i32 @llvm.amdgcn.workitem.id.x()29 %gep = getelementptr i64, ptr addrspace(1) %in, i32 %tid30 %a = load i64, ptr addrspace(1) %gep31 %add = add i64 %a, %b32 %val.bc = bitcast i64 %add to <2 x i32>33 %extract = extractelement <2 x i32> %val.bc, i32 034 store i32 %extract, ptr addrspace(1) %out35 ret void36}37 38define amdgpu_kernel void @bitcast_fp_to_vector_extract_0(ptr addrspace(1) %out, ptr addrspace(1) %in, double %b) {39; GCN-LABEL: bitcast_fp_to_vector_extract_0:40; GCN: ; %bb.0:41; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x942; GCN-NEXT: s_load_dwordx2 s[12:13], s[4:5], 0xd43; GCN-NEXT: s_mov_b32 s7, 0xf00044; GCN-NEXT: s_mov_b32 s10, 045; GCN-NEXT: v_lshlrev_b32_e32 v0, 3, v046; GCN-NEXT: v_mov_b32_e32 v1, 047; GCN-NEXT: s_mov_b32 s11, s748; GCN-NEXT: s_waitcnt lgkmcnt(0)49; GCN-NEXT: s_mov_b64 s[8:9], s[2:3]50; GCN-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[8:11], 0 addr6451; GCN-NEXT: s_mov_b32 s6, -152; GCN-NEXT: s_mov_b32 s4, s053; GCN-NEXT: s_mov_b32 s5, s154; GCN-NEXT: s_waitcnt vmcnt(0)55; GCN-NEXT: v_add_f64 v[0:1], v[0:1], s[12:13]56; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 057; GCN-NEXT: s_endpgm58 %tid = call i32 @llvm.amdgcn.workitem.id.x()59 %gep = getelementptr double, ptr addrspace(1) %in, i32 %tid60 %a = load double, ptr addrspace(1) %gep61 %add = fadd double %a, %b62 %val.bc = bitcast double %add to <2 x i32>63 %extract = extractelement <2 x i32> %val.bc, i32 064 store i32 %extract, ptr addrspace(1) %out65 ret void66}67 68define amdgpu_kernel void @bitcast_int_to_fpvector_extract_0(ptr addrspace(1) %out, ptr addrspace(1) %in, i64 %b) {69; GCN-LABEL: bitcast_int_to_fpvector_extract_0:70; GCN: ; %bb.0:71; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x972; GCN-NEXT: s_load_dword s12, s[4:5], 0xd73; GCN-NEXT: s_mov_b32 s7, 0xf00074; GCN-NEXT: s_mov_b32 s10, 075; GCN-NEXT: v_lshlrev_b32_e32 v0, 3, v076; GCN-NEXT: v_mov_b32_e32 v1, 077; GCN-NEXT: s_mov_b32 s11, s778; GCN-NEXT: s_waitcnt lgkmcnt(0)79; GCN-NEXT: s_mov_b64 s[8:9], s[2:3]80; GCN-NEXT: buffer_load_dword v0, v[0:1], s[8:11], 0 addr6481; GCN-NEXT: s_mov_b32 s6, -182; GCN-NEXT: s_mov_b32 s4, s083; GCN-NEXT: s_mov_b32 s5, s184; GCN-NEXT: s_waitcnt vmcnt(0)85; GCN-NEXT: v_add_i32_e32 v0, vcc, s12, v086; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 087; GCN-NEXT: s_endpgm88 %tid = call i32 @llvm.amdgcn.workitem.id.x()89 %gep = getelementptr i64, ptr addrspace(1) %in, i32 %tid90 %a = load i64, ptr addrspace(1) %gep91 %add = add i64 %a, %b92 %val.bc = bitcast i64 %add to <2 x float>93 %extract = extractelement <2 x float> %val.bc, i32 094 store float %extract, ptr addrspace(1) %out95 ret void96}97 98define amdgpu_kernel void @no_extract_volatile_load_extract0(ptr addrspace(1) %out, ptr addrspace(1) %in) {99; GCN-LABEL: no_extract_volatile_load_extract0:100; GCN: ; %bb.0: ; %entry101; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9102; GCN-NEXT: s_mov_b32 s7, 0xf000103; GCN-NEXT: s_mov_b32 s6, -1104; GCN-NEXT: s_mov_b32 s10, s6105; GCN-NEXT: s_mov_b32 s11, s7106; GCN-NEXT: s_waitcnt lgkmcnt(0)107; GCN-NEXT: s_mov_b32 s8, s2108; GCN-NEXT: s_mov_b32 s9, s3109; GCN-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0 glc110; GCN-NEXT: s_waitcnt vmcnt(0)111; GCN-NEXT: s_mov_b32 s4, s0112; GCN-NEXT: s_mov_b32 s5, s1113; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0114; GCN-NEXT: s_endpgm115entry:116 %vec = load volatile <4 x i32>, ptr addrspace(1) %in117 %elt0 = extractelement <4 x i32> %vec, i32 0118 store i32 %elt0, ptr addrspace(1) %out119 ret void120}121 122define amdgpu_kernel void @no_extract_volatile_load_extract2(ptr addrspace(1) %out, ptr addrspace(1) %in) {123; GCN-LABEL: no_extract_volatile_load_extract2:124; GCN: ; %bb.0: ; %entry125; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9126; GCN-NEXT: s_mov_b32 s7, 0xf000127; GCN-NEXT: s_mov_b32 s6, -1128; GCN-NEXT: s_mov_b32 s10, s6129; GCN-NEXT: s_mov_b32 s11, s7130; GCN-NEXT: s_waitcnt lgkmcnt(0)131; GCN-NEXT: s_mov_b32 s8, s2132; GCN-NEXT: s_mov_b32 s9, s3133; GCN-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0 glc134; GCN-NEXT: s_waitcnt vmcnt(0)135; GCN-NEXT: s_mov_b32 s4, s0136; GCN-NEXT: s_mov_b32 s5, s1137; GCN-NEXT: buffer_store_dword v2, off, s[4:7], 0138; GCN-NEXT: s_endpgm139entry:140 %vec = load volatile <4 x i32>, ptr addrspace(1) %in141 %elt2 = extractelement <4 x i32> %vec, i32 2142 store i32 %elt2, ptr addrspace(1) %out143 ret void144}145 146define amdgpu_kernel void @no_extract_volatile_load_dynextract(ptr addrspace(1) %out, ptr addrspace(1) %in, i32 %idx) {147; GCN-LABEL: no_extract_volatile_load_dynextract:148; GCN: ; %bb.0: ; %entry149; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9150; GCN-NEXT: s_load_dword s12, s[4:5], 0xd151; GCN-NEXT: s_mov_b32 s7, 0xf000152; GCN-NEXT: s_mov_b32 s6, -1153; GCN-NEXT: s_mov_b32 s10, s6154; GCN-NEXT: s_mov_b32 s11, s7155; GCN-NEXT: s_waitcnt lgkmcnt(0)156; GCN-NEXT: s_mov_b32 s4, s0157; GCN-NEXT: s_mov_b32 s5, s1158; GCN-NEXT: s_mov_b32 s8, s2159; GCN-NEXT: s_mov_b32 s9, s3160; GCN-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0 glc161; GCN-NEXT: s_waitcnt vmcnt(0)162; GCN-NEXT: s_cmp_eq_u32 s12, 1163; GCN-NEXT: s_cselect_b64 vcc, -1, 0164; GCN-NEXT: s_cmp_eq_u32 s12, 2165; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc166; GCN-NEXT: s_cselect_b64 vcc, -1, 0167; GCN-NEXT: s_cmp_eq_u32 s12, 3168; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc169; GCN-NEXT: s_cselect_b64 vcc, -1, 0170; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc171; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0172; GCN-NEXT: s_endpgm173entry:174 %vec = load volatile <4 x i32>, ptr addrspace(1) %in175 %eltN = extractelement <4 x i32> %vec, i32 %idx176 store i32 %eltN, ptr addrspace(1) %out177 ret void178}179