1; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s 2; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s 3; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s 4 5declare i7 @llvm.ctlz.i7(i7, i1) nounwind readnone 6declare i8 @llvm.ctlz.i8(i8, i1) nounwind readnone 7declare i16 @llvm.ctlz.i16(i16, i1) nounwind readnone 8 9declare i32 @llvm.ctlz.i32(i32, i1) nounwind readnone 10declare <2 x i32> @llvm.ctlz.v2i32(<2 x i32>, i1) nounwind readnone 11declare <4 x i32> @llvm.ctlz.v4i32(<4 x i32>, i1) nounwind readnone 12 13declare i64 @llvm.ctlz.i64(i64, i1) nounwind readnone 14declare <2 x i64> @llvm.ctlz.v2i64(<2 x i64>, i1) nounwind readnone 15declare <4 x i64> @llvm.ctlz.v4i64(<4 x i64>, i1) nounwind readnone 16 17declare i32 @llvm.r600.read.tidig.x() nounwind readnone 18 19; FUNC-LABEL: {{^}}s_ctlz_i32: 20; SI: s_load_dword [[VAL:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}} 21; SI-DAG: s_flbit_i32_b32 [[CTLZ:s[0-9]+]], [[VAL]] 22; SI-DAG: v_cmp_eq_i32_e64 [[CMPZ:s\[[0-9]+:[0-9]+\]]], 0, [[VAL]] 23; SI-DAG: v_mov_b32_e32 [[VCTLZ:v[0-9]+]], [[CTLZ]] 24; SI: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], [[VCTLZ]], 32, [[CMPZ]] 25; SI: buffer_store_dword [[RESULT]] 26; SI: s_endpgm 27 28; EG: FFBH_UINT 29; EG: CNDE_INT 30define void @s_ctlz_i32(i32 addrspace(1)* noalias %out, i32 %val) nounwind { 31 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 32 store i32 %ctlz, i32 addrspace(1)* %out, align 4 33 ret void 34} 35 36; FUNC-LABEL: {{^}}v_ctlz_i32: 37; SI: buffer_load_dword [[VAL:v[0-9]+]], 38; SI-DAG: v_ffbh_u32_e32 [[CTLZ:v[0-9]+]], [[VAL]] 39; SI-DAG: v_cmp_eq_i32_e32 vcc, 0, [[CTLZ]] 40; SI: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], [[CTLZ]], 32, vcc 41; SI: buffer_store_dword [[RESULT]], 42; SI: s_endpgm 43 44; EG: FFBH_UINT 45; EG: CNDE_INT 46define void @v_ctlz_i32(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind { 47 %val = load i32, i32 addrspace(1)* %valptr, align 4 48 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 49 store i32 %ctlz, i32 addrspace(1)* %out, align 4 50 ret void 51} 52 53; FUNC-LABEL: {{^}}v_ctlz_v2i32: 54; SI: buffer_load_dwordx2 55; SI: v_ffbh_u32_e32 56; SI: v_ffbh_u32_e32 57; SI: buffer_store_dwordx2 58; SI: s_endpgm 59 60; EG: FFBH_UINT 61; EG: CNDE_INT 62; EG: FFBH_UINT 63; EG: CNDE_INT 64define void @v_ctlz_v2i32(<2 x i32> addrspace(1)* noalias %out, <2 x i32> addrspace(1)* noalias %valptr) nounwind { 65 %val = load <2 x i32>, <2 x i32> addrspace(1)* %valptr, align 8 66 %ctlz = call <2 x i32> @llvm.ctlz.v2i32(<2 x i32> %val, i1 false) nounwind readnone 67 store <2 x i32> %ctlz, <2 x i32> addrspace(1)* %out, align 8 68 ret void 69} 70 71; FUNC-LABEL: {{^}}v_ctlz_v4i32: 72; SI: buffer_load_dwordx4 73; SI: v_ffbh_u32_e32 74; SI: v_ffbh_u32_e32 75; SI: v_ffbh_u32_e32 76; SI: v_ffbh_u32_e32 77; SI: buffer_store_dwordx4 78; SI: s_endpgm 79 80 81; EG-DAG: FFBH_UINT 82; EG-DAG: CNDE_INT 83 84; EG-DAG: FFBH_UINT 85; EG-DAG: CNDE_INT 86 87; EG-DAG: FFBH_UINT 88; EG-DAG: CNDE_INT 89 90; EG-DAG: FFBH_UINT 91; EG-DAG: CNDE_INT 92define void @v_ctlz_v4i32(<4 x i32> addrspace(1)* noalias %out, <4 x i32> addrspace(1)* noalias %valptr) nounwind { 93 %val = load <4 x i32>, <4 x i32> addrspace(1)* %valptr, align 16 94 %ctlz = call <4 x i32> @llvm.ctlz.v4i32(<4 x i32> %val, i1 false) nounwind readnone 95 store <4 x i32> %ctlz, <4 x i32> addrspace(1)* %out, align 16 96 ret void 97} 98 99; FUNC-LABEL: {{^}}v_ctlz_i8: 100; SI: buffer_load_ubyte [[VAL:v[0-9]+]], 101; SI-DAG: v_ffbh_u32_e32 [[FFBH:v[0-9]+]], [[VAL]] 102; SI-DAG: v_cmp_eq_i32_e32 vcc, 0, [[CTLZ]] 103; SI-DAG: v_cndmask_b32_e64 [[CORRECTED_FFBH:v[0-9]+]], [[FFBH]], 32, vcc 104; SI: v_add_i32_e32 [[RESULT:v[0-9]+]], vcc, 0xffffffe8, [[CORRECTED_FFBH]] 105; SI: buffer_store_byte [[RESULT]], 106define void @v_ctlz_i8(i8 addrspace(1)* noalias %out, i8 addrspace(1)* noalias %valptr) nounwind { 107 %val = load i8, i8 addrspace(1)* %valptr 108 %ctlz = call i8 @llvm.ctlz.i8(i8 %val, i1 false) nounwind readnone 109 store i8 %ctlz, i8 addrspace(1)* %out 110 ret void 111} 112 113; FUNC-LABEL: {{^}}s_ctlz_i64: 114; SI: s_load_dwordx2 s{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}} 115; SI-DAG: v_cmp_eq_i32_e64 vcc, 0, s[[HI]] 116; SI-DAG: s_flbit_i32_b32 [[FFBH_LO:s[0-9]+]], s[[LO]] 117; SI-DAG: s_add_i32 [[ADD:s[0-9]+]], [[FFBH_LO]], 32 118; SI-DAG: s_flbit_i32_b32 [[FFBH_HI:s[0-9]+]], s[[HI]] 119; SI-DAG: v_mov_b32_e32 [[VFFBH_LO:v[0-9]+]], [[ADD]] 120; SI-DAG: v_mov_b32_e32 [[VFFBH_HI:v[0-9]+]], [[FFBH_HI]] 121; SI-DAG: v_cndmask_b32_e32 v[[CTLZ:[0-9]+]], [[VFFBH_HI]], [[VFFBH_LO]] 122; SI-DAG: v_mov_b32_e32 v[[CTLZ_HI:[0-9]+]], 0{{$}} 123; SI: {{buffer|flat}}_store_dwordx2 {{.*}}v{{\[}}[[CTLZ]]:[[CTLZ_HI]]{{\]}} 124define void @s_ctlz_i64(i64 addrspace(1)* noalias %out, i64 %val) nounwind { 125 %ctlz = call i64 @llvm.ctlz.i64(i64 %val, i1 false) 126 store i64 %ctlz, i64 addrspace(1)* %out 127 ret void 128} 129 130; FUNC-LABEL: {{^}}s_ctlz_i64_trunc: 131define void @s_ctlz_i64_trunc(i32 addrspace(1)* noalias %out, i64 %val) nounwind { 132 %ctlz = call i64 @llvm.ctlz.i64(i64 %val, i1 false) 133 %trunc = trunc i64 %ctlz to i32 134 store i32 %trunc, i32 addrspace(1)* %out 135 ret void 136} 137 138; FUNC-LABEL: {{^}}v_ctlz_i64: 139; SI-DAG: v_mov_b32_e32 v[[CTLZ_HI:[0-9]+]], 0{{$}} 140; SI-DAG: {{buffer|flat}}_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}} 141; SI-DAG: v_cmp_eq_i32_e64 [[CMPHI:s\[[0-9]+:[0-9]+\]]], 0, v[[HI]] 142; SI-DAG: v_ffbh_u32_e32 [[FFBH_LO:v[0-9]+]], v[[LO]] 143; SI-DAG: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, 32, [[FFBH_LO]] 144; SI-DAG: v_ffbh_u32_e32 [[FFBH_HI:v[0-9]+]], v[[HI]] 145; SI-DAG: v_cndmask_b32_e64 v[[CTLZ:[0-9]+]], [[FFBH_HI]], [[ADD]], [[CMPHI]] 146; SI-DAG: v_or_b32_e32 [[OR:v[0-9]+]], v[[LO]], v[[HI]] 147; SI-DAG: v_cmp_eq_i32_e32 vcc, 0, [[OR]] 148; SI-DAG: v_cndmask_b32_e64 v[[CLTZ_LO:[0-9]+]], v[[CTLZ:[0-9]+]], 64, vcc 149; SI: {{buffer|flat}}_store_dwordx2 {{.*}}v{{\[}}[[CLTZ_LO]]:[[CTLZ_HI]]{{\]}} 150define void @v_ctlz_i64(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind { 151 %tid = call i32 @llvm.r600.read.tidig.x() 152 %in.gep = getelementptr i64, i64 addrspace(1)* %in, i32 %tid 153 %out.gep = getelementptr i64, i64 addrspace(1)* %out, i32 %tid 154 %val = load i64, i64 addrspace(1)* %in.gep 155 %ctlz = call i64 @llvm.ctlz.i64(i64 %val, i1 false) 156 store i64 %ctlz, i64 addrspace(1)* %out.gep 157 ret void 158} 159 160; FUNC-LABEL: {{^}}v_ctlz_i64_trunc: 161define void @v_ctlz_i64_trunc(i32 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind { 162 %tid = call i32 @llvm.r600.read.tidig.x() 163 %in.gep = getelementptr i64, i64 addrspace(1)* %in, i32 %tid 164 %out.gep = getelementptr i32, i32 addrspace(1)* %out, i32 %tid 165 %val = load i64, i64 addrspace(1)* %in.gep 166 %ctlz = call i64 @llvm.ctlz.i64(i64 %val, i1 false) 167 %trunc = trunc i64 %ctlz to i32 168 store i32 %trunc, i32 addrspace(1)* %out.gep 169 ret void 170} 171 172; FUNC-LABEL: {{^}}v_ctlz_i32_sel_eq_neg1: 173; SI: buffer_load_dword [[VAL:v[0-9]+]], 174; SI: v_ffbh_u32_e32 [[RESULT:v[0-9]+]], [[VAL]] 175; SI: buffer_store_dword [[RESULT]], 176; SI: s_endpgm 177 define void @v_ctlz_i32_sel_eq_neg1(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind { 178 %val = load i32, i32 addrspace(1)* %valptr 179 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 180 %cmp = icmp eq i32 %val, 0 181 %sel = select i1 %cmp, i32 -1, i32 %ctlz 182 store i32 %sel, i32 addrspace(1)* %out 183 ret void 184} 185 186; FUNC-LABEL: {{^}}v_ctlz_i32_sel_ne_neg1: 187; SI: buffer_load_dword [[VAL:v[0-9]+]], 188; SI: v_ffbh_u32_e32 [[RESULT:v[0-9]+]], [[VAL]] 189; SI: buffer_store_dword [[RESULT]], 190; SI: s_endpgm 191define void @v_ctlz_i32_sel_ne_neg1(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind { 192 %val = load i32, i32 addrspace(1)* %valptr 193 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 194 %cmp = icmp ne i32 %val, 0 195 %sel = select i1 %cmp, i32 %ctlz, i32 -1 196 store i32 %sel, i32 addrspace(1)* %out 197 ret void 198} 199 200; TODO: Should be able to eliminate select here as well. 201; FUNC-LABEL: {{^}}v_ctlz_i32_sel_eq_bitwidth: 202; SI: buffer_load_dword 203; SI: v_ffbh_u32_e32 204; SI: v_cmp 205; SI: v_cndmask 206; SI: s_endpgm 207define void @v_ctlz_i32_sel_eq_bitwidth(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind { 208 %val = load i32, i32 addrspace(1)* %valptr 209 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 210 %cmp = icmp eq i32 %ctlz, 32 211 %sel = select i1 %cmp, i32 -1, i32 %ctlz 212 store i32 %sel, i32 addrspace(1)* %out 213 ret void 214} 215 216; FUNC-LABEL: {{^}}v_ctlz_i32_sel_ne_bitwidth: 217; SI: buffer_load_dword 218; SI: v_ffbh_u32_e32 219; SI: v_cmp 220; SI: v_cndmask 221; SI: s_endpgm 222define void @v_ctlz_i32_sel_ne_bitwidth(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind { 223 %val = load i32, i32 addrspace(1)* %valptr 224 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 225 %cmp = icmp ne i32 %ctlz, 32 226 %sel = select i1 %cmp, i32 %ctlz, i32 -1 227 store i32 %sel, i32 addrspace(1)* %out 228 ret void 229} 230 231; FUNC-LABEL: {{^}}v_ctlz_i8_sel_eq_neg1: 232; SI: buffer_load_ubyte [[VAL:v[0-9]+]], 233; SI: v_ffbh_u32_e32 [[FFBH:v[0-9]+]], [[VAL]] 234; SI: buffer_store_byte [[FFBH]], 235 define void @v_ctlz_i8_sel_eq_neg1(i8 addrspace(1)* noalias %out, i8 addrspace(1)* noalias %valptr) nounwind { 236 %val = load i8, i8 addrspace(1)* %valptr 237 %ctlz = call i8 @llvm.ctlz.i8(i8 %val, i1 false) nounwind readnone 238 %cmp = icmp eq i8 %val, 0 239 %sel = select i1 %cmp, i8 -1, i8 %ctlz 240 store i8 %sel, i8 addrspace(1)* %out 241 ret void 242} 243 244; FUNC-LABEL: {{^}}v_ctlz_i16_sel_eq_neg1: 245; SI: buffer_load_ushort [[VAL:v[0-9]+]], 246; SI: v_ffbh_u32_e32 [[FFBH:v[0-9]+]], [[VAL]] 247; SI: buffer_store_short [[FFBH]], 248 define void @v_ctlz_i16_sel_eq_neg1(i16 addrspace(1)* noalias %out, i16 addrspace(1)* noalias %valptr) nounwind { 249 %val = load i16, i16 addrspace(1)* %valptr 250 %ctlz = call i16 @llvm.ctlz.i16(i16 %val, i1 false) nounwind readnone 251 %cmp = icmp eq i16 %val, 0 252 %sel = select i1 %cmp, i16 -1, i16 %ctlz 253 store i16 %sel, i16 addrspace(1)* %out 254 ret void 255} 256 257; FUNC-LABEL: {{^}}v_ctlz_i7_sel_eq_neg1: 258; SI: buffer_load_ubyte [[VAL:v[0-9]+]], 259; SI: v_ffbh_u32_e32 [[FFBH:v[0-9]+]], [[VAL]] 260; SI: v_and_b32_e32 [[TRUNC:v[0-9]+]], 0x7f, [[FFBH]] 261; SI: buffer_store_byte [[TRUNC]], 262 define void @v_ctlz_i7_sel_eq_neg1(i7 addrspace(1)* noalias %out, i7 addrspace(1)* noalias %valptr) nounwind { 263 %val = load i7, i7 addrspace(1)* %valptr 264 %ctlz = call i7 @llvm.ctlz.i7(i7 %val, i1 false) nounwind readnone 265 %cmp = icmp eq i7 %val, 0 266 %sel = select i1 %cmp, i7 -1, i7 %ctlz 267 store i7 %sel, i7 addrspace(1)* %out 268 ret void 269} 270