1 ; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck %s --check-prefix=R600-CHECK --check-prefix=FUNC 2 ; RUN: llc -verify-machineinstrs -march=r600 -mcpu=SI < %s | FileCheck %s --check-prefix=SI-CHECK --check-prefix=FUNC 3 4 declare i32 @llvm.r600.read.tidig.x() nounwind readnone 5 6 ; FUNC-LABEL: @mova_same_clause 7 8 ; R600-CHECK: LDS_WRITE 9 ; R600-CHECK: LDS_WRITE 10 ; R600-CHECK: LDS_READ 11 ; R600-CHECK: LDS_READ 12 13 ; SI-CHECK: DS_WRITE_B32 14 ; SI-CHECK: DS_WRITE_B32 15 ; SI-CHECK: DS_READ_B32 16 ; SI-CHECK: DS_READ_B32 17 define void @mova_same_clause(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) { 18 entry: 19 %stack = alloca [5 x i32], align 4 20 %0 = load i32 addrspace(1)* %in, align 4 21 %arrayidx1 = getelementptr inbounds [5 x i32]* %stack, i32 0, i32 %0 22 store i32 4, i32* %arrayidx1, align 4 23 %arrayidx2 = getelementptr inbounds i32 addrspace(1)* %in, i32 1 24 %1 = load i32 addrspace(1)* %arrayidx2, align 4 25 %arrayidx3 = getelementptr inbounds [5 x i32]* %stack, i32 0, i32 %1 26 store i32 5, i32* %arrayidx3, align 4 27 %arrayidx10 = getelementptr inbounds [5 x i32]* %stack, i32 0, i32 0 28 %2 = load i32* %arrayidx10, align 4 29 store i32 %2, i32 addrspace(1)* %out, align 4 30 %arrayidx12 = getelementptr inbounds [5 x i32]* %stack, i32 0, i32 1 31 %3 = load i32* %arrayidx12 32 %arrayidx13 = getelementptr inbounds i32 addrspace(1)* %out, i32 1 33 store i32 %3, i32 addrspace(1)* %arrayidx13 34 ret void 35 } 36 37 ; This test checks that the stack offset is calculated correctly for structs. 38 ; All register loads/stores should be optimized away, so there shouldn't be 39 ; any MOVA instructions. 40 ; 41 ; XXX: This generated code has unnecessary MOVs, we should be able to optimize 42 ; this. 43 44 ; FUNC-LABEL: @multiple_structs 45 ; R600-CHECK-NOT: MOVA_INT 46 ; SI-CHECK-NOT: V_MOVREL 47 %struct.point = type { i32, i32 } 48 49 define void @multiple_structs(i32 addrspace(1)* %out) { 50 entry: 51 %a = alloca %struct.point 52 %b = alloca %struct.point 53 %a.x.ptr = getelementptr %struct.point* %a, i32 0, i32 0 54 %a.y.ptr = getelementptr %struct.point* %a, i32 0, i32 1 55 %b.x.ptr = getelementptr %struct.point* %b, i32 0, i32 0 56 %b.y.ptr = getelementptr %struct.point* %b, i32 0, i32 1 57 store i32 0, i32* %a.x.ptr 58 store i32 1, i32* %a.y.ptr 59 store i32 2, i32* %b.x.ptr 60 store i32 3, i32* %b.y.ptr 61 %a.indirect.ptr = getelementptr %struct.point* %a, i32 0, i32 0 62 %b.indirect.ptr = getelementptr %struct.point* %b, i32 0, i32 0 63 %a.indirect = load i32* %a.indirect.ptr 64 %b.indirect = load i32* %b.indirect.ptr 65 %0 = add i32 %a.indirect, %b.indirect 66 store i32 %0, i32 addrspace(1)* %out 67 ret void 68 } 69 70 ; Test direct access of a private array inside a loop. The private array 71 ; loads and stores should be lowered to copies, so there shouldn't be any 72 ; MOVA instructions. 73 74 ; FUNC-LABEL: @direct_loop 75 ; R600-CHECK-NOT: MOVA_INT 76 ; SI-CHECK-NOT: V_MOVREL 77 78 define void @direct_loop(i32 addrspace(1)* %out, i32 addrspace(1)* %in) { 79 entry: 80 %prv_array_const = alloca [2 x i32] 81 %prv_array = alloca [2 x i32] 82 %a = load i32 addrspace(1)* %in 83 %b_src_ptr = getelementptr i32 addrspace(1)* %in, i32 1 84 %b = load i32 addrspace(1)* %b_src_ptr 85 %a_dst_ptr = getelementptr [2 x i32]* %prv_array_const, i32 0, i32 0 86 store i32 %a, i32* %a_dst_ptr 87 %b_dst_ptr = getelementptr [2 x i32]* %prv_array_const, i32 0, i32 1 88 store i32 %b, i32* %b_dst_ptr 89 br label %for.body 90 91 for.body: 92 %inc = phi i32 [0, %entry], [%count, %for.body] 93 %x_ptr = getelementptr [2 x i32]* %prv_array_const, i32 0, i32 0 94 %x = load i32* %x_ptr 95 %y_ptr = getelementptr [2 x i32]* %prv_array, i32 0, i32 0 96 %y = load i32* %y_ptr 97 %xy = add i32 %x, %y 98 store i32 %xy, i32* %y_ptr 99 %count = add i32 %inc, 1 100 %done = icmp eq i32 %count, 4095 101 br i1 %done, label %for.end, label %for.body 102 103 for.end: 104 %value_ptr = getelementptr [2 x i32]* %prv_array, i32 0, i32 0 105 %value = load i32* %value_ptr 106 store i32 %value, i32 addrspace(1)* %out 107 ret void 108 } 109 110 ; FUNC-LABEL: @short_array 111 112 ; R600-CHECK: MOVA_INT 113 114 ; SI-CHECK: V_MOVRELS_B32_e32 115 define void @short_array(i32 addrspace(1)* %out, i32 %index) { 116 entry: 117 %0 = alloca [2 x i16] 118 %1 = getelementptr [2 x i16]* %0, i32 0, i32 0 119 %2 = getelementptr [2 x i16]* %0, i32 0, i32 1 120 store i16 0, i16* %1 121 store i16 1, i16* %2 122 %3 = getelementptr [2 x i16]* %0, i32 0, i32 %index 123 %4 = load i16* %3 124 %5 = sext i16 %4 to i32 125 store i32 %5, i32 addrspace(1)* %out 126 ret void 127 } 128 129 ; FUNC-LABEL: @char_array 130 131 ; R600-CHECK: MOVA_INT 132 133 ; SI-CHECK: V_OR_B32_e32 v{{[0-9]}}, 0x100 134 ; SI-CHECK: V_MOVRELS_B32_e32 135 define void @char_array(i32 addrspace(1)* %out, i32 %index) { 136 entry: 137 %0 = alloca [2 x i8] 138 %1 = getelementptr [2 x i8]* %0, i32 0, i32 0 139 %2 = getelementptr [2 x i8]* %0, i32 0, i32 1 140 store i8 0, i8* %1 141 store i8 1, i8* %2 142 %3 = getelementptr [2 x i8]* %0, i32 0, i32 %index 143 %4 = load i8* %3 144 %5 = sext i8 %4 to i32 145 store i32 %5, i32 addrspace(1)* %out 146 ret void 147 148 } 149 150 ; Make sure we don't overwrite workitem information with private memory 151 152 ; FUNC-LABEL: @work_item_info 153 ; R600-CHECK-NOT: MOV T0.X 154 ; Additional check in case the move ends up in the last slot 155 ; R600-CHECK-NOT: MOV * TO.X 156 157 ; SI-CHECK-NOT: V_MOV_B32_e{{(32|64)}} v0 158 define void @work_item_info(i32 addrspace(1)* %out, i32 %in) { 159 entry: 160 %0 = alloca [2 x i32] 161 %1 = getelementptr [2 x i32]* %0, i32 0, i32 0 162 %2 = getelementptr [2 x i32]* %0, i32 0, i32 1 163 store i32 0, i32* %1 164 store i32 1, i32* %2 165 %3 = getelementptr [2 x i32]* %0, i32 0, i32 %in 166 %4 = load i32* %3 167 %5 = call i32 @llvm.r600.read.tidig.x() 168 %6 = add i32 %4, %5 169 store i32 %6, i32 addrspace(1)* %out 170 ret void 171 } 172 173 ; Test that two stack objects are not stored in the same register 174 ; The second stack object should be in T3.X 175 ; FUNC-LABEL: @no_overlap 176 ; R600_CHECK: MOV 177 ; R600_CHECK: [[CHAN:[XYZW]]]+ 178 ; R600-CHECK-NOT: [[CHAN]]+ 179 ; SI-CHECK: V_MOV_B32_e32 v3 180 define void @no_overlap(i32 addrspace(1)* %out, i32 %in) { 181 entry: 182 %0 = alloca [3 x i8], align 1 183 %1 = alloca [2 x i8], align 1 184 %2 = getelementptr [3 x i8]* %0, i32 0, i32 0 185 %3 = getelementptr [3 x i8]* %0, i32 0, i32 1 186 %4 = getelementptr [3 x i8]* %0, i32 0, i32 2 187 %5 = getelementptr [2 x i8]* %1, i32 0, i32 0 188 %6 = getelementptr [2 x i8]* %1, i32 0, i32 1 189 store i8 0, i8* %2 190 store i8 1, i8* %3 191 store i8 2, i8* %4 192 store i8 1, i8* %5 193 store i8 0, i8* %6 194 %7 = getelementptr [3 x i8]* %0, i32 0, i32 %in 195 %8 = getelementptr [2 x i8]* %1, i32 0, i32 %in 196 %9 = load i8* %7 197 %10 = load i8* %8 198 %11 = add i8 %9, %10 199 %12 = sext i8 %11 to i32 200 store i32 %12, i32 addrspace(1)* %out 201 ret void 202 } 203 204 define void @char_array_array(i32 addrspace(1)* %out, i32 %index) { 205 entry: 206 %alloca = alloca [2 x [2 x i8]] 207 %gep0 = getelementptr [2 x [2 x i8]]* %alloca, i32 0, i32 0, i32 0 208 %gep1 = getelementptr [2 x [2 x i8]]* %alloca, i32 0, i32 0, i32 1 209 store i8 0, i8* %gep0 210 store i8 1, i8* %gep1 211 %gep2 = getelementptr [2 x [2 x i8]]* %alloca, i32 0, i32 0, i32 %index 212 %load = load i8* %gep2 213 %sext = sext i8 %load to i32 214 store i32 %sext, i32 addrspace(1)* %out 215 ret void 216 } 217 218 define void @i32_array_array(i32 addrspace(1)* %out, i32 %index) { 219 entry: 220 %alloca = alloca [2 x [2 x i32]] 221 %gep0 = getelementptr [2 x [2 x i32]]* %alloca, i32 0, i32 0, i32 0 222 %gep1 = getelementptr [2 x [2 x i32]]* %alloca, i32 0, i32 0, i32 1 223 store i32 0, i32* %gep0 224 store i32 1, i32* %gep1 225 %gep2 = getelementptr [2 x [2 x i32]]* %alloca, i32 0, i32 0, i32 %index 226 %load = load i32* %gep2 227 store i32 %load, i32 addrspace(1)* %out 228 ret void 229 } 230 231 define void @i64_array_array(i64 addrspace(1)* %out, i32 %index) { 232 entry: 233 %alloca = alloca [2 x [2 x i64]] 234 %gep0 = getelementptr [2 x [2 x i64]]* %alloca, i32 0, i32 0, i32 0 235 %gep1 = getelementptr [2 x [2 x i64]]* %alloca, i32 0, i32 0, i32 1 236 store i64 0, i64* %gep0 237 store i64 1, i64* %gep1 238 %gep2 = getelementptr [2 x [2 x i64]]* %alloca, i32 0, i32 0, i32 %index 239 %load = load i64* %gep2 240 store i64 %load, i64 addrspace(1)* %out 241 ret void 242 } 243 244 %struct.pair32 = type { i32, i32 } 245 246 define void @struct_array_array(i32 addrspace(1)* %out, i32 %index) { 247 entry: 248 %alloca = alloca [2 x [2 x %struct.pair32]] 249 %gep0 = getelementptr [2 x [2 x %struct.pair32]]* %alloca, i32 0, i32 0, i32 0, i32 1 250 %gep1 = getelementptr [2 x [2 x %struct.pair32]]* %alloca, i32 0, i32 0, i32 1, i32 1 251 store i32 0, i32* %gep0 252 store i32 1, i32* %gep1 253 %gep2 = getelementptr [2 x [2 x %struct.pair32]]* %alloca, i32 0, i32 0, i32 %index, i32 0 254 %load = load i32* %gep2 255 store i32 %load, i32 addrspace(1)* %out 256 ret void 257 } 258 259 define void @struct_pair32_array(i32 addrspace(1)* %out, i32 %index) { 260 entry: 261 %alloca = alloca [2 x %struct.pair32] 262 %gep0 = getelementptr [2 x %struct.pair32]* %alloca, i32 0, i32 0, i32 1 263 %gep1 = getelementptr [2 x %struct.pair32]* %alloca, i32 0, i32 1, i32 0 264 store i32 0, i32* %gep0 265 store i32 1, i32* %gep1 266 %gep2 = getelementptr [2 x %struct.pair32]* %alloca, i32 0, i32 %index, i32 0 267 %load = load i32* %gep2 268 store i32 %load, i32 addrspace(1)* %out 269 ret void 270 } 271 272 define void @select_private(i32 addrspace(1)* %out, i32 %in) nounwind { 273 entry: 274 %tmp = alloca [2 x i32] 275 %tmp1 = getelementptr [2 x i32]* %tmp, i32 0, i32 0 276 %tmp2 = getelementptr [2 x i32]* %tmp, i32 0, i32 1 277 store i32 0, i32* %tmp1 278 store i32 1, i32* %tmp2 279 %cmp = icmp eq i32 %in, 0 280 %sel = select i1 %cmp, i32* %tmp1, i32* %tmp2 281 %load = load i32* %sel 282 store i32 %load, i32 addrspace(1)* %out 283 ret void 284 } 285 286
