1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -passes=loop-vectorize -force-vector-width=2 -S %s | FileCheck %s 3 4; Tests where the indices of some accesses are clamped to a small range. 5 6; FIXME: At the moment, the runtime checks require that the indices do not wrap 7; and runtime checks are emitted to ensure that. The clamped indices do 8; wrap, so the vector loops are dead at the moment. But it is still 9; possible to compute the bounds of the accesses and generate proper 10; runtime checks. 11 12; The relevant bounds for %gep.A are [%A, %A+4). 13define void @load_clamped_index(i32* %A, i32* %B, i32 %N) { 14; CHECK-LABEL: @load_clamped_index( 15; CHECK-NEXT: entry: 16; CHECK-NEXT: [[A2:%.*]] = ptrtoint i32* [[A:%.*]] to i64 17; CHECK-NEXT: [[B1:%.*]] = ptrtoint i32* [[B:%.*]] to i64 18; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N:%.*]], 2 19; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]] 20; CHECK: vector.scevcheck: 21; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1 22; CHECK-NEXT: [[TMP7:%.*]] = icmp ugt i32 [[TMP0]], 3 23; CHECK-NEXT: br i1 [[TMP7]], label [[SCALAR_PH]], label [[VECTOR_MEMCHECK:%.*]] 24; CHECK: vector.memcheck: 25; CHECK-NEXT: [[TMP2:%.*]] = sub i64 [[B1]], [[A2]] 26; CHECK-NEXT: [[DIFF_CHECK:%.*]] = icmp ult i64 [[TMP2]], 8 27; CHECK-NEXT: br i1 [[DIFF_CHECK]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 28; CHECK: vector.ph: 29; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 2 30; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]] 31; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 32; CHECK: vector.body: 33; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 34; CHECK-NEXT: [[TMP13:%.*]] = add i32 [[INDEX]], 0 35; CHECK-NEXT: [[TMP14:%.*]] = urem i32 [[TMP13]], 4 36; CHECK-NEXT: [[TMP15:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[TMP14]] 37; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, i32* [[TMP15]], i32 0 38; CHECK-NEXT: [[TMP17:%.*]] = bitcast i32* [[TMP16]] to <2 x i32>* 39; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, <2 x i32>* [[TMP17]], align 4 40; CHECK-NEXT: [[TMP18:%.*]] = add <2 x i32> [[WIDE_LOAD]], <i32 10, i32 10> 41; CHECK-NEXT: [[TMP19:%.*]] = getelementptr inbounds i32, i32* [[B]], i32 [[TMP13]] 42; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i32, i32* [[TMP19]], i32 0 43; CHECK-NEXT: [[TMP21:%.*]] = bitcast i32* [[TMP20]] to <2 x i32>* 44; CHECK-NEXT: store <2 x i32> [[TMP18]], <2 x i32>* [[TMP21]], align 4 45; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2 46; CHECK-NEXT: [[TMP22:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 47; CHECK-NEXT: br i1 [[TMP22]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]] 48; CHECK: middle.block: 49; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]] 50; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 51; CHECK: scalar.ph: 52; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ], [ 0, [[VECTOR_SCEVCHECK]] ], [ 0, [[VECTOR_MEMCHECK]] ] 53; CHECK-NEXT: br label [[LOOP:%.*]] 54; CHECK: loop: 55; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 56; CHECK-NEXT: [[CLAMPED_INDEX:%.*]] = urem i32 [[IV]], 4 57; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[CLAMPED_INDEX]] 58; CHECK-NEXT: [[LV:%.*]] = load i32, i32* [[GEP_A]], align 4 59; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV]], 10 60; CHECK-NEXT: [[GEP_B:%.*]] = getelementptr inbounds i32, i32* [[B]], i32 [[IV]] 61; CHECK-NEXT: store i32 [[ADD]], i32* [[GEP_B]], align 4 62; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1 63; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]] 64; CHECK-NEXT: br i1 [[COND]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP7:![0-9]+]] 65; CHECK: exit: 66; CHECK-NEXT: ret void 67; 68entry: 69 br label %loop 70 71loop: 72 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 73 %clamped.index = urem i32 %iv, 4 74 %gep.A = getelementptr inbounds i32, i32* %A, i32 %clamped.index 75 %lv = load i32, i32* %gep.A 76 %add = add i32 %lv, 10 77 %gep.B = getelementptr inbounds i32, i32* %B, i32 %iv 78 store i32 %add, i32* %gep.B 79 %iv.next = add nuw nsw i32 %iv, 1 80 %cond = icmp eq i32 %iv.next, %N 81 br i1 %cond, label %exit, label %loop 82 83exit: 84 ret void 85} 86 87; The relevant bounds for %gep.A are [%A, %A+4). 88define void @store_clamped_index(i32* %A, i32* %B, i32 %N) { 89; CHECK-LABEL: @store_clamped_index( 90; CHECK-NEXT: entry: 91; CHECK-NEXT: [[B2:%.*]] = ptrtoint i32* [[B:%.*]] to i64 92; CHECK-NEXT: [[A1:%.*]] = ptrtoint i32* [[A:%.*]] to i64 93; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N:%.*]], 2 94; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]] 95; CHECK: vector.scevcheck: 96; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1 97; CHECK-NEXT: [[TMP7:%.*]] = icmp ugt i32 [[TMP0]], 3 98; CHECK-NEXT: br i1 [[TMP7]], label [[SCALAR_PH]], label [[VECTOR_MEMCHECK:%.*]] 99; CHECK: vector.memcheck: 100; CHECK-NEXT: [[TMP2:%.*]] = sub i64 [[A1]], [[B2]] 101; CHECK-NEXT: [[DIFF_CHECK:%.*]] = icmp ult i64 [[TMP2]], 8 102; CHECK-NEXT: br i1 [[DIFF_CHECK]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 103; CHECK: vector.ph: 104; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 2 105; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]] 106; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 107; CHECK: vector.body: 108; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 109; CHECK-NEXT: [[TMP13:%.*]] = add i32 [[INDEX]], 0 110; CHECK-NEXT: [[TMP14:%.*]] = urem i32 [[TMP13]], 4 111; CHECK-NEXT: [[TMP15:%.*]] = getelementptr inbounds i32, i32* [[B]], i32 [[TMP13]] 112; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, i32* [[TMP15]], i32 0 113; CHECK-NEXT: [[TMP17:%.*]] = bitcast i32* [[TMP16]] to <2 x i32>* 114; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, <2 x i32>* [[TMP17]], align 4 115; CHECK-NEXT: [[TMP18:%.*]] = add <2 x i32> [[WIDE_LOAD]], <i32 10, i32 10> 116; CHECK-NEXT: [[TMP19:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[TMP14]] 117; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i32, i32* [[TMP19]], i32 0 118; CHECK-NEXT: [[TMP21:%.*]] = bitcast i32* [[TMP20]] to <2 x i32>* 119; CHECK-NEXT: store <2 x i32> [[TMP18]], <2 x i32>* [[TMP21]], align 4 120; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2 121; CHECK-NEXT: [[TMP22:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 122; CHECK-NEXT: br i1 [[TMP22]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP13:![0-9]+]] 123; CHECK: middle.block: 124; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]] 125; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 126; CHECK: scalar.ph: 127; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ], [ 0, [[VECTOR_SCEVCHECK]] ], [ 0, [[VECTOR_MEMCHECK]] ] 128; CHECK-NEXT: br label [[LOOP:%.*]] 129; CHECK: loop: 130; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 131; CHECK-NEXT: [[CLAMPED_INDEX:%.*]] = urem i32 [[IV]], 4 132; CHECK-NEXT: [[GEP_B:%.*]] = getelementptr inbounds i32, i32* [[B]], i32 [[IV]] 133; CHECK-NEXT: [[LV:%.*]] = load i32, i32* [[GEP_B]], align 4 134; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV]], 10 135; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[CLAMPED_INDEX]] 136; CHECK-NEXT: store i32 [[ADD]], i32* [[GEP_A]], align 4 137; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1 138; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]] 139; CHECK-NEXT: br i1 [[COND]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP14:![0-9]+]] 140; CHECK: exit: 141; CHECK-NEXT: ret void 142; 143entry: 144 br label %loop 145 146loop: 147 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 148 %clamped.index = urem i32 %iv, 4 149 %gep.B = getelementptr inbounds i32, i32* %B, i32 %iv 150 %lv = load i32, i32* %gep.B 151 %add = add i32 %lv, 10 152 %gep.A = getelementptr inbounds i32, i32* %A, i32 %clamped.index 153 store i32 %add, i32* %gep.A 154 %iv.next = add nuw nsw i32 %iv, 1 155 %cond = icmp eq i32 %iv.next, %N 156 br i1 %cond, label %exit, label %loop 157 158exit: 159 ret void 160} 161 162define void @clamped_index_dependence_non_clamped(i32* %A, i32* %B, i32 %N) { 163; CHECK-LABEL: @clamped_index_dependence_non_clamped( 164; CHECK-NEXT: entry: 165; CHECK-NEXT: br label [[LOOP:%.*]] 166; CHECK: loop: 167; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 168; CHECK-NEXT: [[GEP_B:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i32 [[IV]] 169; CHECK-NEXT: [[LV:%.*]] = load i32, i32* [[GEP_B]], align 4 170; CHECK-NEXT: [[GEP_A_1:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i32 [[IV]] 171; CHECK-NEXT: [[LV_A:%.*]] = load i32, i32* [[GEP_A_1]], align 4 172; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV]], [[LV_A]] 173; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1 174; CHECK-NEXT: [[CLAMPED_INDEX:%.*]] = urem i32 [[IV_NEXT]], 4 175; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[CLAMPED_INDEX]] 176; CHECK-NEXT: store i32 [[ADD]], i32* [[GEP_A]], align 4 177; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV_NEXT]], [[N:%.*]] 178; CHECK-NEXT: br i1 [[COND]], label [[EXIT:%.*]], label [[LOOP]] 179; CHECK: exit: 180; CHECK-NEXT: ret void 181; 182entry: 183 br label %loop 184 185loop: 186 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 187 %gep.B = getelementptr inbounds i32, i32* %B, i32 %iv 188 %lv = load i32, i32* %gep.B 189 %gep.A.1 = getelementptr inbounds i32, i32* %A, i32 %iv 190 %lv.A = load i32, i32* %gep.A.1 191 %add = add i32 %lv, %lv.A 192 193 %iv.next = add nuw nsw i32 %iv, 1 194 %clamped.index = urem i32 %iv.next, 4 195 %gep.A = getelementptr inbounds i32, i32* %A, i32 %clamped.index 196 store i32 %add, i32* %gep.A 197 %cond = icmp eq i32 %iv.next, %N 198 br i1 %cond, label %exit, label %loop 199 200exit: 201 ret void 202} 203 204define void @clamped_index_dependence_clamped_index(i32* %A, i32* %B, i32 %N) { 205; CHECK-LABEL: @clamped_index_dependence_clamped_index( 206; CHECK-NEXT: entry: 207; CHECK-NEXT: br label [[LOOP:%.*]] 208; CHECK: loop: 209; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 210; CHECK-NEXT: [[CLAMPED_INDEX_1:%.*]] = urem i32 [[IV]], 4 211; CHECK-NEXT: [[GEP_A_1:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i32 [[CLAMPED_INDEX_1]] 212; CHECK-NEXT: [[LV_A:%.*]] = load i32, i32* [[GEP_A_1]], align 4 213; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV_A]], 10 214; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1 215; CHECK-NEXT: [[CLAMPED_INDEX:%.*]] = urem i32 [[IV_NEXT]], 4 216; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[CLAMPED_INDEX]] 217; CHECK-NEXT: store i32 [[ADD]], i32* [[GEP_A]], align 4 218; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV_NEXT]], [[N:%.*]] 219; CHECK-NEXT: br i1 [[COND]], label [[EXIT:%.*]], label [[LOOP]] 220; CHECK: exit: 221; CHECK-NEXT: ret void 222; 223entry: 224 br label %loop 225 226loop: 227 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 228 %clamped.index.1 = urem i32 %iv, 4 229 %gep.A.1 = getelementptr inbounds i32, i32* %A, i32 %clamped.index.1 230 %lv.A = load i32, i32* %gep.A.1 231 %add = add i32 %lv.A, 10 232 233 %iv.next = add nuw nsw i32 %iv, 1 234 %clamped.index = urem i32 %iv.next, 4 235 %gep.A = getelementptr inbounds i32, i32* %A, i32 %clamped.index 236 store i32 %add, i32* %gep.A 237 %cond = icmp eq i32 %iv.next, %N 238 br i1 %cond, label %exit, label %loop 239 240exit: 241 ret void 242} 243 244define void @clamped_index_equal_dependence(i32* %A, i32* %B, i32 %N) { 245; CHECK-LABEL: @clamped_index_equal_dependence( 246; CHECK-NEXT: entry: 247; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N:%.*]], 2 248; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]] 249; CHECK: vector.scevcheck: 250; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1 251; CHECK-NEXT: [[TMP7:%.*]] = icmp ugt i32 [[TMP0]], 3 252; CHECK-NEXT: br i1 [[TMP7]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 253; CHECK: vector.ph: 254; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 2 255; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]] 256; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 257; CHECK: vector.body: 258; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 259; CHECK-NEXT: [[TMP10:%.*]] = add i32 [[INDEX]], 0 260; CHECK-NEXT: [[TMP11:%.*]] = urem i32 [[TMP10]], 4 261; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i32 [[TMP11]] 262; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds i32, i32* [[TMP12]], i32 0 263; CHECK-NEXT: [[TMP14:%.*]] = bitcast i32* [[TMP13]] to <2 x i32>* 264; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, <2 x i32>* [[TMP14]], align 4 265; CHECK-NEXT: [[TMP15:%.*]] = add <2 x i32> [[WIDE_LOAD]], <i32 10, i32 10> 266; CHECK-NEXT: [[TMP16:%.*]] = bitcast i32* [[TMP13]] to <2 x i32>* 267; CHECK-NEXT: store <2 x i32> [[TMP15]], <2 x i32>* [[TMP16]], align 4 268; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2 269; CHECK-NEXT: [[TMP17:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 270; CHECK-NEXT: br i1 [[TMP17]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP15:![0-9]+]] 271; CHECK: middle.block: 272; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]] 273; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 274; CHECK: scalar.ph: 275; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ], [ 0, [[VECTOR_SCEVCHECK]] ] 276; CHECK-NEXT: br label [[LOOP:%.*]] 277; CHECK: loop: 278; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 279; CHECK-NEXT: [[CLAMPED_INDEX:%.*]] = urem i32 [[IV]], 4 280; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[CLAMPED_INDEX]] 281; CHECK-NEXT: [[LV_A:%.*]] = load i32, i32* [[GEP_A]], align 4 282; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV_A]], 10 283; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1 284; CHECK-NEXT: store i32 [[ADD]], i32* [[GEP_A]], align 4 285; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]] 286; CHECK-NEXT: br i1 [[COND]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP16:![0-9]+]] 287; CHECK: exit: 288; CHECK-NEXT: ret void 289; 290entry: 291 br label %loop 292 293loop: 294 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 295 %clamped.index = urem i32 %iv, 4 296 %gep.A = getelementptr inbounds i32, i32* %A, i32 %clamped.index 297 %lv.A = load i32, i32* %gep.A 298 %add = add i32 %lv.A, 10 299 300 %iv.next = add nuw nsw i32 %iv, 1 301 store i32 %add, i32* %gep.A 302 %cond = icmp eq i32 %iv.next, %N 303 br i1 %cond, label %exit, label %loop 304 305exit: 306 ret void 307} 308