1; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=GCN -check-prefix=FUNC %s 2; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=GCN -check-prefix=FUNC %s 3; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s 4 5; FUNC-LABEL: {{^}}sext_bool_icmp_eq_0: 6; GCN-NOT: v_cmp 7; GCN: v_cmp_ne_i32_e32 vcc, 8; GCN-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 9; GCN-NEXT:buffer_store_byte [[RESULT]] 10; GCN-NEXT: s_endpgm 11 12; EG: SETNE_INT * [[CMP:T[0-9]+]].[[CMPCHAN:[XYZW]]], KC0[2].Z, KC0[2].W 13; EG: AND_INT T{{[0-9]+.[XYZW]}}, PS, 1 14define void @sext_bool_icmp_eq_0(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 15 %icmp0 = icmp eq i32 %a, %b 16 %ext = sext i1 %icmp0 to i32 17 %icmp1 = icmp eq i32 %ext, 0 18 store i1 %icmp1, i1 addrspace(1)* %out 19 ret void 20} 21 22; FUNC-LABEL: {{^}}sext_bool_icmp_ne_0: 23; GCN-NOT: v_cmp 24; GCN: v_cmp_ne_i32_e32 vcc, 25; GCN-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 26; GCN-NEXT: buffer_store_byte [[RESULT]] 27; GCN-NEXT: s_endpgm 28 29; EG: SETNE_INT * [[CMP:T[0-9]+]].[[CMPCHAN:[XYZW]]], KC0[2].Z, KC0[2].W 30; EG: AND_INT T{{[0-9]+.[XYZW]}}, PS, 1 31define void @sext_bool_icmp_ne_0(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 32 %icmp0 = icmp ne i32 %a, %b 33 %ext = sext i1 %icmp0 to i32 34 %icmp1 = icmp ne i32 %ext, 0 35 store i1 %icmp1, i1 addrspace(1)* %out 36 ret void 37} 38 39; FUNC-LABEL: {{^}}sext_bool_icmp_eq_neg1: 40; GCN-NOT: v_cmp 41; GCN: v_cmp_eq_i32_e32 vcc, 42; GCN-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 43; GCN-NEXT: buffer_store_byte [[RESULT]] 44; GCN-NEXT: s_endpgm 45define void @sext_bool_icmp_eq_neg1(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 46 %icmp0 = icmp eq i32 %a, %b 47 %ext = sext i1 %icmp0 to i32 48 %icmp1 = icmp eq i32 %ext, -1 49 store i1 %icmp1, i1 addrspace(1)* %out 50 ret void 51} 52 53; FUNC-LABEL: {{^}}sext_bool_icmp_ne_neg1: 54; GCN-NOT: v_cmp 55; GCN: v_cmp_eq_i32_e32 vcc, 56; GCN-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 57; GCN-NEXT: buffer_store_byte [[RESULT]] 58; GCN-NEXT: s_endpgm 59define void @sext_bool_icmp_ne_neg1(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 60 %icmp0 = icmp ne i32 %a, %b 61 %ext = sext i1 %icmp0 to i32 62 %icmp1 = icmp ne i32 %ext, -1 63 store i1 %icmp1, i1 addrspace(1)* %out 64 ret void 65} 66 67; FUNC-LABEL: {{^}}zext_bool_icmp_eq_0: 68; GCN-NOT: v_cmp 69; GCN: v_cmp_ne_i32_e32 vcc, 70; GCN-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 71; GCN-NEXT: buffer_store_byte [[RESULT]] 72; GCN-NEXT: s_endpgm 73define void @zext_bool_icmp_eq_0(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 74 %icmp0 = icmp eq i32 %a, %b 75 %ext = zext i1 %icmp0 to i32 76 %icmp1 = icmp eq i32 %ext, 0 77 store i1 %icmp1, i1 addrspace(1)* %out 78 ret void 79} 80 81; FUNC-LABEL: {{^}}zext_bool_icmp_ne_0: 82; GCN-NOT: v_cmp 83; GCN: v_cmp_ne_i32_e32 vcc, 84; GCN-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 85; GCN-NEXT: buffer_store_byte [[RESULT]] 86; GCN-NEXT: s_endpgm 87define void @zext_bool_icmp_ne_0(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 88 %icmp0 = icmp ne i32 %a, %b 89 %ext = zext i1 %icmp0 to i32 90 %icmp1 = icmp ne i32 %ext, 0 91 store i1 %icmp1, i1 addrspace(1)* %out 92 ret void 93} 94 95; FUNC-LABEL: {{^}}zext_bool_icmp_eq_1: 96; GCN-NOT: v_cmp 97; GCN: v_cmp_eq_i32_e32 vcc, 98; GCN-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 99; GCN-NEXT: buffer_store_byte [[RESULT]] 100; GCN-NEXT: s_endpgm 101define void @zext_bool_icmp_eq_1(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 102 %icmp0 = icmp eq i32 %a, %b 103 %ext = zext i1 %icmp0 to i32 104 %icmp1 = icmp eq i32 %ext, 1 105 store i1 %icmp1, i1 addrspace(1)* %out 106 ret void 107} 108 109; FUNC-LABEL: {{^}}zext_bool_icmp_ne_1: 110; GCN-NOT: v_cmp 111; GCN: v_cmp_eq_i32_e32 vcc, 112; GCN-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 113; GCN-NEXT: buffer_store_byte [[RESULT]] 114define void @zext_bool_icmp_ne_1(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 115 %icmp0 = icmp ne i32 %a, %b 116 %ext = zext i1 %icmp0 to i32 117 %icmp1 = icmp ne i32 %ext, 1 118 store i1 %icmp1, i1 addrspace(1)* %out 119 ret void 120} 121 122; Reduces to false: 123; FUNC-LABEL: {{^}}zext_bool_icmp_eq_neg1: 124; GCN: v_mov_b32_e32 [[TMP:v[0-9]+]], 0{{$}} 125; GCN: buffer_store_byte [[TMP]] 126; GCN-NEXT: s_endpgm 127define void @zext_bool_icmp_eq_neg1(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 128 %icmp0 = icmp eq i32 %a, %b 129 %ext = zext i1 %icmp0 to i32 130 %icmp1 = icmp eq i32 %ext, -1 131 store i1 %icmp1, i1 addrspace(1)* %out 132 ret void 133} 134 135; Reduces to true: 136; FUNC-LABEL: {{^}}zext_bool_icmp_ne_neg1: 137; GCN: v_mov_b32_e32 [[TMP:v[0-9]+]], 1{{$}} 138; GCN: buffer_store_byte [[TMP]] 139; GCN-NEXT: s_endpgm 140define void @zext_bool_icmp_ne_neg1(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 141 %icmp0 = icmp ne i32 %a, %b 142 %ext = zext i1 %icmp0 to i32 143 %icmp1 = icmp ne i32 %ext, -1 144 store i1 %icmp1, i1 addrspace(1)* %out 145 ret void 146} 147 148; FUNC-LABEL: {{^}}cmp_zext_k_i8max: 149; SI: s_load_dword [[VALUE:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb 150; VI: s_load_dword [[VALUE:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x2c 151; GCN: s_movk_i32 [[K255:s[0-9]+]], 0xff 152; GCN-DAG: s_and_b32 [[B:s[0-9]+]], [[VALUE]], [[K255]] 153; GCN-DAG: v_mov_b32_e32 [[VK255:v[0-9]+]], [[K255]] 154; GCN: v_cmp_ne_i32_e32 vcc, [[B]], [[VK255]] 155; GCN: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 156; GCN: buffer_store_byte [[RESULT]] 157; GCN: s_endpgm 158define void @cmp_zext_k_i8max(i1 addrspace(1)* %out, i8 %b) nounwind { 159 %b.ext = zext i8 %b to i32 160 %icmp0 = icmp ne i32 %b.ext, 255 161 store i1 %icmp0, i1 addrspace(1)* %out 162 ret void 163} 164 165; FUNC-LABEL: {{^}}cmp_sext_k_neg1: 166; GCN: buffer_load_sbyte [[B:v[0-9]+]] 167; GCN: v_cmp_ne_i32_e32 vcc, -1, [[B]]{{$}} 168; GCN-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 169; GCN: buffer_store_byte [[RESULT]] 170; GCN: s_endpgm 171define void @cmp_sext_k_neg1(i1 addrspace(1)* %out, i8 addrspace(1)* %b.ptr) nounwind { 172 %b = load i8, i8 addrspace(1)* %b.ptr 173 %b.ext = sext i8 %b to i32 174 %icmp0 = icmp ne i32 %b.ext, -1 175 store i1 %icmp0, i1 addrspace(1)* %out 176 ret void 177} 178 179; FUNC-LABEL: {{^}}cmp_sext_k_neg1_i8_sext_arg: 180; GCN: s_load_dword [[B:s[0-9]+]] 181; GCN: v_cmp_ne_i32_e64 [[CMP:s\[[0-9]+:[0-9]+\]]], -1, [[B]] 182; GCN-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, [[CMP]] 183; GCN-NEXT: buffer_store_byte [[RESULT]] 184; GCN: s_endpgm 185define void @cmp_sext_k_neg1_i8_sext_arg(i1 addrspace(1)* %out, i8 signext %b) nounwind { 186 %b.ext = sext i8 %b to i32 187 %icmp0 = icmp ne i32 %b.ext, -1 188 store i1 %icmp0, i1 addrspace(1)* %out 189 ret void 190} 191 192; FIXME: This ends up doing a buffer_load_ubyte, and and compare to 193; 255. Seems to be because of ordering problems when not allowing load widths to be reduced. 194; Should do a buffer_load_sbyte and compare with -1 195 196; FUNC-LABEL: {{^}}cmp_sext_k_neg1_i8_arg: 197; SI: s_load_dword [[VAL:s[0-9]+]], s[{{[0-9]+:[0-9]+}}], 0xb 198; VI: s_load_dword [[VAL:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x2c 199; GCN: s_movk_i32 [[K:s[0-9]+]], 0xff 200; GCN-DAG: s_and_b32 [[B:s[0-9]+]], [[VAL]], [[K]] 201; GCN-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], [[K]] 202; GCN: v_cmp_ne_i32_e32 vcc, [[B]], [[VK]]{{$}} 203; GCN: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, vcc 204; GCN: buffer_store_byte [[RESULT]] 205; GCN: s_endpgm 206define void @cmp_sext_k_neg1_i8_arg(i1 addrspace(1)* %out, i8 %b) nounwind { 207 %b.ext = sext i8 %b to i32 208 %icmp0 = icmp ne i32 %b.ext, -1 209 store i1 %icmp0, i1 addrspace(1)* %out 210 ret void 211} 212 213; FUNC-LABEL: {{^}}cmp_zext_k_neg1: 214; GCN: v_mov_b32_e32 [[RESULT:v[0-9]+]], 1{{$}} 215; GCN: buffer_store_byte [[RESULT]] 216; GCN: s_endpgm 217define void @cmp_zext_k_neg1(i1 addrspace(1)* %out, i8 %b) nounwind { 218 %b.ext = zext i8 %b to i32 219 %icmp0 = icmp ne i32 %b.ext, -1 220 store i1 %icmp0, i1 addrspace(1)* %out 221 ret void 222} 223 224; FUNC-LABEL: {{^}}zext_bool_icmp_ne_k: 225; GCN: v_mov_b32_e32 [[RESULT:v[0-9]+]], 1{{$}} 226; GCN: buffer_store_byte [[RESULT]] 227; GCN-NEXT: s_endpgm 228define void @zext_bool_icmp_ne_k(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 229 %icmp0 = icmp ne i32 %a, %b 230 %ext = zext i1 %icmp0 to i32 231 %icmp1 = icmp ne i32 %ext, 2 232 store i1 %icmp1, i1 addrspace(1)* %out 233 ret void 234} 235 236; FUNC-LABEL: {{^}}zext_bool_icmp_eq_k: 237; GCN: v_mov_b32_e32 [[RESULT:v[0-9]+]], 0{{$}} 238; GCN: buffer_store_byte [[RESULT]] 239; GCN-NEXT: s_endpgm 240define void @zext_bool_icmp_eq_k(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 241 %icmp0 = icmp ne i32 %a, %b 242 %ext = zext i1 %icmp0 to i32 243 %icmp1 = icmp eq i32 %ext, 2 244 store i1 %icmp1, i1 addrspace(1)* %out 245 ret void 246} 247 248; FIXME: These cases should really be able fold to true/false in 249; DAGCombiner 250 251; This really folds away to false 252; FUNC-LABEL: {{^}}sext_bool_icmp_eq_1: 253; GCN: v_mov_b32_e32 [[K:v[0-9]+]], 0{{$}} 254; GCN: buffer_store_byte [[K]] 255define void @sext_bool_icmp_eq_1(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 256 %icmp0 = icmp eq i32 %a, %b 257 %ext = sext i1 %icmp0 to i32 258 %icmp1 = icmp eq i32 %ext, 1 259 store i1 %icmp1, i1 addrspace(1)* %out 260 ret void 261} 262 263; FUNC-LABEL: {{^}}sext_bool_icmp_ne_1: 264; GCN: v_mov_b32_e32 [[K:v[0-9]+]], 1{{$}} 265; GCN: buffer_store_byte [[K]] 266define void @sext_bool_icmp_ne_1(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 267 %icmp0 = icmp ne i32 %a, %b 268 %ext = sext i1 %icmp0 to i32 269 %icmp1 = icmp ne i32 %ext, 1 270 store i1 %icmp1, i1 addrspace(1)* %out 271 ret void 272} 273 274; FUNC-LABEL: {{^}}sext_bool_icmp_ne_k: 275; GCN: v_mov_b32_e32 [[K:v[0-9]+]], 1{{$}} 276; GCN: buffer_store_byte [[K]] 277define void @sext_bool_icmp_ne_k(i1 addrspace(1)* %out, i32 %a, i32 %b) nounwind { 278 %icmp0 = icmp ne i32 %a, %b 279 %ext = sext i1 %icmp0 to i32 280 %icmp1 = icmp ne i32 %ext, 2 281 store i1 %icmp1, i1 addrspace(1)* %out 282 ret void 283} 284