1; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI -check-prefix=FUNC %s 2; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI -check-prefix=FUNC %s 3; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -enable-var-scope -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; GCN: s_load_dword [[VAL:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}} 21; GCN-DAG: s_flbit_i32_b32 [[CTLZ:s[0-9]+]], [[VAL]] 22; GCN-DAG: v_cmp_ne_u32_e64 vcc, [[VAL]], 0{{$}} 23; GCN-DAG: v_mov_b32_e32 [[VCTLZ:v[0-9]+]], [[CTLZ]] 24; GCN: v_cndmask_b32_e32 [[RESULT:v[0-9]+]], 32, [[VCTLZ]], vcc 25; GCN: buffer_store_dword [[RESULT]] 26; GCN: s_endpgm 27 28; EG: FFBH_UINT 29; EG: CNDE_INT 30define amdgpu_kernel 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; GCN: {{buffer|flat}}_load_dword [[VAL:v[0-9]+]], 38; GCN: v_ffbh_u32_e32 [[CTLZ:v[0-9]+]], [[VAL]] 39; GCN: v_cmp_ne_u32_e32 vcc, 0, [[VAL]] 40; GCN: v_cndmask_b32_e32 [[RESULT:v[0-9]+]], 32, [[CTLZ]], vcc 41; GCN: buffer_store_dword [[RESULT]], 42; GCN: s_endpgm 43 44; EG: FFBH_UINT 45; EG: CNDE_INT 46define amdgpu_kernel void @v_ctlz_i32(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind { 47 %tid = call i32 @llvm.r600.read.tidig.x() 48 %in.gep = getelementptr i32, i32 addrspace(1)* %valptr, i32 %tid 49 %val = load i32, i32 addrspace(1)* %in.gep, align 4 50 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 51 store i32 %ctlz, i32 addrspace(1)* %out, align 4 52 ret void 53} 54 55; FUNC-LABEL: {{^}}v_ctlz_v2i32: 56; GCN: {{buffer|flat}}_load_dwordx2 57; GCN: v_ffbh_u32_e32 58; GCN: v_ffbh_u32_e32 59; GCN: buffer_store_dwordx2 60; GCN: s_endpgm 61 62; EG: FFBH_UINT 63; EG: CNDE_INT 64; EG: FFBH_UINT 65; EG: CNDE_INT 66define amdgpu_kernel void @v_ctlz_v2i32(<2 x i32> addrspace(1)* noalias %out, <2 x i32> addrspace(1)* noalias %valptr) nounwind { 67 %tid = call i32 @llvm.r600.read.tidig.x() 68 %in.gep = getelementptr <2 x i32>, <2 x i32> addrspace(1)* %valptr, i32 %tid 69 %val = load <2 x i32>, <2 x i32> addrspace(1)* %in.gep, align 8 70 %ctlz = call <2 x i32> @llvm.ctlz.v2i32(<2 x i32> %val, i1 false) nounwind readnone 71 store <2 x i32> %ctlz, <2 x i32> addrspace(1)* %out, align 8 72 ret void 73} 74 75; FUNC-LABEL: {{^}}v_ctlz_v4i32: 76; GCN: {{buffer|flat}}_load_dwordx4 77; GCN: v_ffbh_u32_e32 78; GCN: v_ffbh_u32_e32 79; GCN: v_ffbh_u32_e32 80; GCN: v_ffbh_u32_e32 81; GCN: buffer_store_dwordx4 82; GCN: s_endpgm 83 84 85; EG-DAG: FFBH_UINT 86; EG-DAG: CNDE_INT 87 88; EG-DAG: FFBH_UINT 89; EG-DAG: CNDE_INT 90 91; EG-DAG: FFBH_UINT 92; EG-DAG: CNDE_INT 93 94; EG-DAG: FFBH_UINT 95; EG-DAG: CNDE_INT 96define amdgpu_kernel void @v_ctlz_v4i32(<4 x i32> addrspace(1)* noalias %out, <4 x i32> addrspace(1)* noalias %valptr) nounwind { 97 %tid = call i32 @llvm.r600.read.tidig.x() 98 %in.gep = getelementptr <4 x i32>, <4 x i32> addrspace(1)* %valptr, i32 %tid 99 %val = load <4 x i32>, <4 x i32> addrspace(1)* %in.gep, align 16 100 %ctlz = call <4 x i32> @llvm.ctlz.v4i32(<4 x i32> %val, i1 false) nounwind readnone 101 store <4 x i32> %ctlz, <4 x i32> addrspace(1)* %out, align 16 102 ret void 103} 104 105; FUNC-LABEL: {{^}}v_ctlz_i8: 106; GCN: {{buffer|flat}}_load_ubyte [[VAL:v[0-9]+]], 107; SI-DAG: v_ffbh_u32_e32 [[FFBH:v[0-9]+]], [[VAL]] 108; VI-DAG: v_ffbh_u32_sdwa [[FFBH:v[0-9]+]], [[VAL]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 109; SI: v_cmp_ne_u32_e32 vcc, 0, [[VAL]] 110; VI: v_cmp_ne_u16_e32 vcc, 0, [[VAL]] 111 112; GCN: v_cndmask_b32_e32 [[SELECT:v[0-9]+]], 32, [[FFBH]], vcc 113 114; SI: v_subrev_i32_e32 [[RESULT:v[0-9]+]], vcc, 24, [[SELECT]] 115; VI: v_add_u32_e32 [[RESULT:v[0-9]+]], vcc, -16, [[SELECT]] 116; GCN: buffer_store_byte [[RESULT]], 117; GCN: s_endpgm 118define amdgpu_kernel void @v_ctlz_i8(i8 addrspace(1)* noalias %out, i8 addrspace(1)* noalias %valptr) nounwind { 119 %val = load i8, i8 addrspace(1)* %valptr 120 %ctlz = call i8 @llvm.ctlz.i8(i8 %val, i1 false) nounwind readnone 121 store i8 %ctlz, i8 addrspace(1)* %out 122 ret void 123} 124 125; FUNC-LABEL: {{^}}s_ctlz_i64: 126; GCN: s_load_dwordx2 s{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0x13|0x4c}} 127; GCN-DAG: v_cmp_eq_u32_e64 vcc, s[[HI]], 0{{$}} 128; GCN-DAG: s_flbit_i32_b32 [[FFBH_LO:s[0-9]+]], s[[LO]] 129; GCN-DAG: s_add_i32 [[ADD:s[0-9]+]], [[FFBH_LO]], 32 130; GCN-DAG: s_flbit_i32_b32 [[FFBH_HI:s[0-9]+]], s[[HI]] 131; GCN-DAG: v_mov_b32_e32 [[VFFBH_LO:v[0-9]+]], [[ADD]] 132; GCN-DAG: v_mov_b32_e32 [[VFFBH_HI:v[0-9]+]], [[FFBH_HI]] 133; GCN-DAG: v_cndmask_b32_e32 v[[CTLZ:[0-9]+]], [[VFFBH_HI]], [[VFFBH_LO]] 134; GCN-DAG: v_mov_b32_e32 v[[CTLZ_HI:[0-9]+]], 0{{$}} 135; GCN: {{buffer|flat}}_store_dwordx2 {{.*}}v{{\[}}[[CTLZ]]:[[CTLZ_HI]]{{\]}} 136define amdgpu_kernel void @s_ctlz_i64(i64 addrspace(1)* noalias %out, [8 x i32], i64 %val) nounwind { 137 %ctlz = call i64 @llvm.ctlz.i64(i64 %val, i1 false) 138 store i64 %ctlz, i64 addrspace(1)* %out 139 ret void 140} 141 142; FUNC-LABEL: {{^}}s_ctlz_i64_trunc: 143define amdgpu_kernel void @s_ctlz_i64_trunc(i32 addrspace(1)* noalias %out, i64 %val) nounwind { 144 %ctlz = call i64 @llvm.ctlz.i64(i64 %val, i1 false) 145 %trunc = trunc i64 %ctlz to i32 146 store i32 %trunc, i32 addrspace(1)* %out 147 ret void 148} 149 150; FUNC-LABEL: {{^}}v_ctlz_i64: 151; GCN-DAG: {{buffer|flat}}_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}} 152; GCN-DAG: v_cmp_eq_u32_e32 vcc, 0, v[[HI]] 153; GCN-DAG: v_ffbh_u32_e32 [[FFBH_LO:v[0-9]+]], v[[LO]] 154; GCN-DAG: v_add_{{[iu]}}32_e32 [[ADD:v[0-9]+]], vcc, 32, [[FFBH_LO]] 155; GCN-DAG: v_ffbh_u32_e32 [[FFBH_HI:v[0-9]+]], v[[HI]] 156; GCN-DAG: v_cndmask_b32_e32 v[[CTLZ:[0-9]+]], [[FFBH_HI]], [[ADD]], vcc 157; GCN-DAG: v_or_b32_e32 [[OR:v[0-9]+]], v[[LO]], v[[HI]] 158; GCN-DAG: v_cmp_ne_u32_e32 vcc, 0, [[OR]] 159; GCN-DAG: v_cndmask_b32_e32 v[[CLTZ_LO:[0-9]+]], 64, v[[CTLZ:[0-9]+]], vcc 160; GCN: {{buffer|flat}}_store_dwordx2 {{.*}}v{{\[}}[[CLTZ_LO]]:[[CTLZ_HI:[0-9]+]]{{\]}} 161define amdgpu_kernel void @v_ctlz_i64(i64 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 i64, i64 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 store i64 %ctlz, i64 addrspace(1)* %out.gep 168 ret void 169} 170 171; FUNC-LABEL: {{^}}v_ctlz_i64_trunc: 172define amdgpu_kernel void @v_ctlz_i64_trunc(i32 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind { 173 %tid = call i32 @llvm.r600.read.tidig.x() 174 %in.gep = getelementptr i64, i64 addrspace(1)* %in, i32 %tid 175 %out.gep = getelementptr i32, i32 addrspace(1)* %out, i32 %tid 176 %val = load i64, i64 addrspace(1)* %in.gep 177 %ctlz = call i64 @llvm.ctlz.i64(i64 %val, i1 false) 178 %trunc = trunc i64 %ctlz to i32 179 store i32 %trunc, i32 addrspace(1)* %out.gep 180 ret void 181} 182 183; FUNC-LABEL: {{^}}v_ctlz_i32_sel_eq_neg1: 184; GCN: {{buffer|flat}}_load_dword [[VAL:v[0-9]+]], 185; GCN: v_ffbh_u32_e32 [[RESULT:v[0-9]+]], [[VAL]] 186; GCN: buffer_store_dword [[RESULT]], 187; GCN: s_endpgm 188define amdgpu_kernel void @v_ctlz_i32_sel_eq_neg1(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind { 189 %tid = call i32 @llvm.r600.read.tidig.x() 190 %in.gep = getelementptr i32, i32 addrspace(1)* %valptr, i32 %tid 191 %val = load i32, i32 addrspace(1)* %in.gep 192 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 193 %cmp = icmp eq i32 %val, 0 194 %sel = select i1 %cmp, i32 -1, i32 %ctlz 195 store i32 %sel, i32 addrspace(1)* %out 196 ret void 197} 198 199; FUNC-LABEL: {{^}}v_ctlz_i32_sel_ne_neg1: 200; GCN: {{buffer|flat}}_load_dword [[VAL:v[0-9]+]], 201; GCN: v_ffbh_u32_e32 [[RESULT:v[0-9]+]], [[VAL]] 202; GCN: buffer_store_dword [[RESULT]], 203; GCN: s_endpgm 204define amdgpu_kernel void @v_ctlz_i32_sel_ne_neg1(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind { 205 %tid = call i32 @llvm.r600.read.tidig.x() 206 %in.gep = getelementptr i32, i32 addrspace(1)* %valptr, i32 %tid 207 %val = load i32, i32 addrspace(1)* %in.gep 208 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 209 %cmp = icmp ne i32 %val, 0 210 %sel = select i1 %cmp, i32 %ctlz, i32 -1 211 store i32 %sel, i32 addrspace(1)* %out 212 ret void 213} 214 215; TODO: Should be able to eliminate select here as well. 216; FUNC-LABEL: {{^}}v_ctlz_i32_sel_eq_bitwidth: 217; GCN: {{buffer|flat}}_load_dword 218; GCN: v_ffbh_u32_e32 219; GCN: v_cmp 220; GCN: v_cndmask 221; GCN: s_endpgm 222define amdgpu_kernel void @v_ctlz_i32_sel_eq_bitwidth(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind { 223 %tid = call i32 @llvm.r600.read.tidig.x() 224 %in.gep = getelementptr i32, i32 addrspace(1)* %valptr, i32 %tid 225 %val = load i32, i32 addrspace(1)* %in.gep 226 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 227 %cmp = icmp eq i32 %ctlz, 32 228 %sel = select i1 %cmp, i32 -1, i32 %ctlz 229 store i32 %sel, i32 addrspace(1)* %out 230 ret void 231} 232 233; FUNC-LABEL: {{^}}v_ctlz_i32_sel_ne_bitwidth: 234; GCN: {{buffer|flat}}_load_dword 235; GCN: v_ffbh_u32_e32 236; GCN: v_cmp 237; GCN: v_cndmask 238; GCN: s_endpgm 239define amdgpu_kernel void @v_ctlz_i32_sel_ne_bitwidth(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind { 240 %tid = call i32 @llvm.r600.read.tidig.x() 241 %in.gep = getelementptr i32, i32 addrspace(1)* %valptr, i32 %tid 242 %val = load i32, i32 addrspace(1)* %in.gep 243 %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone 244 %cmp = icmp ne i32 %ctlz, 32 245 %sel = select i1 %cmp, i32 %ctlz, i32 -1 246 store i32 %sel, i32 addrspace(1)* %out 247 ret void 248} 249 250; FUNC-LABEL: {{^}}v_ctlz_i8_sel_eq_neg1: 251; GCN: {{buffer|flat}}_load_ubyte [[VAL:v[0-9]+]], 252; GCN: v_ffbh_u32_e32 [[FFBH:v[0-9]+]], [[VAL]] 253; GCN: {{buffer|flat}}_store_byte [[FFBH]], 254 define amdgpu_kernel void @v_ctlz_i8_sel_eq_neg1(i8 addrspace(1)* noalias %out, i8 addrspace(1)* noalias %valptr) nounwind { 255 %tid = call i32 @llvm.r600.read.tidig.x() 256 %valptr.gep = getelementptr i8, i8 addrspace(1)* %valptr, i32 %tid 257 %val = load i8, i8 addrspace(1)* %valptr.gep 258 %ctlz = call i8 @llvm.ctlz.i8(i8 %val, i1 false) nounwind readnone 259 %cmp = icmp eq i8 %val, 0 260 %sel = select i1 %cmp, i8 -1, i8 %ctlz 261 store i8 %sel, i8 addrspace(1)* %out 262 ret void 263} 264 265; FUNC-LABEL: {{^}}v_ctlz_i16_sel_eq_neg1: 266; SI: {{buffer|flat}}_load_ushort [[VAL:v[0-9]+]], 267; SI: v_ffbh_u32_e32 [[FFBH:v[0-9]+]], [[VAL]] 268; SI: buffer_store_short [[FFBH]], 269 define amdgpu_kernel void @v_ctlz_i16_sel_eq_neg1(i16 addrspace(1)* noalias %out, i16 addrspace(1)* noalias %valptr) nounwind { 270 %val = load i16, i16 addrspace(1)* %valptr 271 %ctlz = call i16 @llvm.ctlz.i16(i16 %val, i1 false) nounwind readnone 272 %cmp = icmp eq i16 %val, 0 273 %sel = select i1 %cmp, i16 -1, i16 %ctlz 274 store i16 %sel, i16 addrspace(1)* %out 275 ret void 276} 277 278; FIXME: Need to handle non-uniform case for function below (load without gep). 279; FUNC-LABEL: {{^}}v_ctlz_i7_sel_eq_neg1: 280; GCN: {{buffer|flat}}_load_ubyte [[VAL:v[0-9]+]], 281; GCN: v_ffbh_u32_e32 [[FFBH:v[0-9]+]], [[VAL]] 282; GCN: v_and_b32_e32 [[TRUNC:v[0-9]+]], 0x7f, [[FFBH]] 283; GCN: {{buffer|flat}}_store_byte [[TRUNC]], 284define amdgpu_kernel void @v_ctlz_i7_sel_eq_neg1(i7 addrspace(1)* noalias %out, i7 addrspace(1)* noalias %valptr) nounwind { 285 %tid = call i32 @llvm.r600.read.tidig.x() 286 %valptr.gep = getelementptr i7, i7 addrspace(1)* %valptr, i32 %tid 287 %val = load i7, i7 addrspace(1)* %valptr.gep 288 %ctlz = call i7 @llvm.ctlz.i7(i7 %val, i1 false) nounwind readnone 289 %cmp = icmp eq i7 %val, 0 290 %sel = select i1 %cmp, i7 -1, i7 %ctlz 291 store i7 %sel, i7 addrspace(1)* %out 292 ret void 293} 294