1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -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: [[B1:%.*]] = bitcast i32* [[B:%.*]] to i8* 17; CHECK-NEXT: [[A3:%.*]] = bitcast i32* [[A:%.*]] to i8* 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: [[TMP1:%.*]] = trunc i32 [[TMP0]] to i2 23; CHECK-NEXT: [[MUL:%.*]] = call { i2, i1 } @llvm.umul.with.overflow.i2(i2 1, i2 [[TMP1]]) 24; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i2, i1 } [[MUL]], 0 25; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i2, i1 } [[MUL]], 1 26; CHECK-NEXT: [[TMP2:%.*]] = add i2 0, [[MUL_RESULT]] 27; CHECK-NEXT: [[TMP3:%.*]] = sub i2 0, [[MUL_RESULT]] 28; CHECK-NEXT: [[TMP4:%.*]] = icmp ugt i2 [[TMP3]], 0 29; CHECK-NEXT: [[TMP5:%.*]] = icmp ult i2 [[TMP2]], 0 30; CHECK-NEXT: [[TMP6:%.*]] = select i1 false, i1 [[TMP4]], i1 [[TMP5]] 31; CHECK-NEXT: [[TMP7:%.*]] = icmp ugt i32 [[TMP0]], 3 32; CHECK-NEXT: [[TMP8:%.*]] = or i1 [[TMP6]], [[TMP7]] 33; CHECK-NEXT: [[TMP9:%.*]] = or i1 [[TMP8]], [[MUL_OVERFLOW]] 34; CHECK-NEXT: [[TMP10:%.*]] = or i1 false, [[TMP9]] 35; CHECK-NEXT: br i1 [[TMP10]], label [[SCALAR_PH]], label [[VECTOR_MEMCHECK:%.*]] 36; CHECK: vector.memcheck: 37; CHECK-NEXT: [[TMP11:%.*]] = add i32 [[N]], -1 38; CHECK-NEXT: [[TMP12:%.*]] = zext i32 [[TMP11]] to i64 39; CHECK-NEXT: [[TMP13:%.*]] = add nuw nsw i64 [[TMP12]], 1 40; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr i32, i32* [[B]], i64 [[TMP13]] 41; CHECK-NEXT: [[SCEVGEP2:%.*]] = bitcast i32* [[SCEVGEP]] to i8* 42; CHECK-NEXT: [[SCEVGEP4:%.*]] = getelementptr i32, i32* [[A]], i64 [[TMP13]] 43; CHECK-NEXT: [[SCEVGEP45:%.*]] = bitcast i32* [[SCEVGEP4]] to i8* 44; CHECK-NEXT: [[BOUND0:%.*]] = icmp ult i8* [[B1]], [[SCEVGEP45]] 45; CHECK-NEXT: [[BOUND1:%.*]] = icmp ult i8* [[A3]], [[SCEVGEP2]] 46; CHECK-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]] 47; CHECK-NEXT: br i1 [[FOUND_CONFLICT]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 48; CHECK: vector.ph: 49; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 2 50; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]] 51; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 52; CHECK: vector.body: 53; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 54; CHECK-NEXT: [[TMP14:%.*]] = add i32 [[INDEX]], 0 55; CHECK-NEXT: [[TMP15:%.*]] = urem i32 [[TMP14]], 4 56; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[TMP15]] 57; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds i32, i32* [[TMP16]], i32 0 58; CHECK-NEXT: [[TMP18:%.*]] = bitcast i32* [[TMP17]] to <2 x i32>* 59; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, <2 x i32>* [[TMP18]], align 4, !alias.scope !0 60; CHECK-NEXT: [[TMP19:%.*]] = add <2 x i32> [[WIDE_LOAD]], <i32 10, i32 10> 61; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i32, i32* [[B]], i32 [[TMP14]] 62; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds i32, i32* [[TMP20]], i32 0 63; CHECK-NEXT: [[TMP22:%.*]] = bitcast i32* [[TMP21]] to <2 x i32>* 64; CHECK-NEXT: store <2 x i32> [[TMP19]], <2 x i32>* [[TMP22]], align 4, !alias.scope !3, !noalias !0 65; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2 66; CHECK-NEXT: [[TMP23:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 67; CHECK-NEXT: br i1 [[TMP23]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]] 68; CHECK: middle.block: 69; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]] 70; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 71; CHECK: scalar.ph: 72; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ], [ 0, [[VECTOR_SCEVCHECK]] ], [ 0, [[VECTOR_MEMCHECK]] ] 73; CHECK-NEXT: br label [[LOOP:%.*]] 74; CHECK: loop: 75; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 76; CHECK-NEXT: [[CLAMPED_INDEX:%.*]] = urem i32 [[IV]], 4 77; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[CLAMPED_INDEX]] 78; CHECK-NEXT: [[LV:%.*]] = load i32, i32* [[GEP_A]], align 4 79; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV]], 10 80; CHECK-NEXT: [[GEP_B:%.*]] = getelementptr inbounds i32, i32* [[B]], i32 [[IV]] 81; CHECK-NEXT: store i32 [[ADD]], i32* [[GEP_B]], align 4 82; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1 83; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]] 84; CHECK-NEXT: br i1 [[COND]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP7:![0-9]+]] 85; CHECK: exit: 86; CHECK-NEXT: ret void 87; 88entry: 89 br label %loop 90 91loop: 92 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 93 %clamped.index = urem i32 %iv, 4 94 %gep.A = getelementptr inbounds i32, i32* %A, i32 %clamped.index 95 %lv = load i32, i32* %gep.A 96 %add = add i32 %lv, 10 97 %gep.B = getelementptr inbounds i32, i32* %B, i32 %iv 98 store i32 %add, i32* %gep.B 99 %iv.next = add nuw nsw i32 %iv, 1 100 %cond = icmp eq i32 %iv.next, %N 101 br i1 %cond, label %exit, label %loop 102 103exit: 104 ret void 105} 106 107; The relevant bounds for %gep.A are [%A, %A+4). 108define void @store_clamped_index(i32* %A, i32* %B, i32 %N) { 109; CHECK-LABEL: @store_clamped_index( 110; CHECK-NEXT: entry: 111; CHECK-NEXT: [[B1:%.*]] = bitcast i32* [[B:%.*]] to i8* 112; CHECK-NEXT: [[A3:%.*]] = bitcast i32* [[A:%.*]] to i8* 113; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N:%.*]], 2 114; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]] 115; CHECK: vector.scevcheck: 116; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1 117; CHECK-NEXT: [[TMP1:%.*]] = trunc i32 [[TMP0]] to i2 118; CHECK-NEXT: [[MUL:%.*]] = call { i2, i1 } @llvm.umul.with.overflow.i2(i2 1, i2 [[TMP1]]) 119; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i2, i1 } [[MUL]], 0 120; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i2, i1 } [[MUL]], 1 121; CHECK-NEXT: [[TMP2:%.*]] = add i2 0, [[MUL_RESULT]] 122; CHECK-NEXT: [[TMP3:%.*]] = sub i2 0, [[MUL_RESULT]] 123; CHECK-NEXT: [[TMP4:%.*]] = icmp ugt i2 [[TMP3]], 0 124; CHECK-NEXT: [[TMP5:%.*]] = icmp ult i2 [[TMP2]], 0 125; CHECK-NEXT: [[TMP6:%.*]] = select i1 false, i1 [[TMP4]], i1 [[TMP5]] 126; CHECK-NEXT: [[TMP7:%.*]] = icmp ugt i32 [[TMP0]], 3 127; CHECK-NEXT: [[TMP8:%.*]] = or i1 [[TMP6]], [[TMP7]] 128; CHECK-NEXT: [[TMP9:%.*]] = or i1 [[TMP8]], [[MUL_OVERFLOW]] 129; CHECK-NEXT: [[TMP10:%.*]] = or i1 false, [[TMP9]] 130; CHECK-NEXT: br i1 [[TMP10]], label [[SCALAR_PH]], label [[VECTOR_MEMCHECK:%.*]] 131; CHECK: vector.memcheck: 132; CHECK-NEXT: [[TMP11:%.*]] = add i32 [[N]], -1 133; CHECK-NEXT: [[TMP12:%.*]] = zext i32 [[TMP11]] to i64 134; CHECK-NEXT: [[TMP13:%.*]] = add nuw nsw i64 [[TMP12]], 1 135; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr i32, i32* [[B]], i64 [[TMP13]] 136; CHECK-NEXT: [[SCEVGEP2:%.*]] = bitcast i32* [[SCEVGEP]] to i8* 137; CHECK-NEXT: [[SCEVGEP4:%.*]] = getelementptr i32, i32* [[A]], i64 [[TMP13]] 138; CHECK-NEXT: [[SCEVGEP45:%.*]] = bitcast i32* [[SCEVGEP4]] to i8* 139; CHECK-NEXT: [[BOUND0:%.*]] = icmp ult i8* [[B1]], [[SCEVGEP45]] 140; CHECK-NEXT: [[BOUND1:%.*]] = icmp ult i8* [[A3]], [[SCEVGEP2]] 141; CHECK-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]] 142; CHECK-NEXT: br i1 [[FOUND_CONFLICT]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 143; CHECK: vector.ph: 144; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 2 145; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]] 146; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 147; CHECK: vector.body: 148; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 149; CHECK-NEXT: [[TMP14:%.*]] = add i32 [[INDEX]], 0 150; CHECK-NEXT: [[TMP15:%.*]] = urem i32 [[TMP14]], 4 151; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, i32* [[B]], i32 [[TMP14]] 152; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds i32, i32* [[TMP16]], i32 0 153; CHECK-NEXT: [[TMP18:%.*]] = bitcast i32* [[TMP17]] to <2 x i32>* 154; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, <2 x i32>* [[TMP18]], align 4, !alias.scope !8, !noalias !11 155; CHECK-NEXT: [[TMP19:%.*]] = add <2 x i32> [[WIDE_LOAD]], <i32 10, i32 10> 156; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[TMP15]] 157; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds i32, i32* [[TMP20]], i32 0 158; CHECK-NEXT: [[TMP22:%.*]] = bitcast i32* [[TMP21]] to <2 x i32>* 159; CHECK-NEXT: store <2 x i32> [[TMP19]], <2 x i32>* [[TMP22]], align 4, !alias.scope !11 160; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2 161; CHECK-NEXT: [[TMP23:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 162; CHECK-NEXT: br i1 [[TMP23]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP13:![0-9]+]] 163; CHECK: middle.block: 164; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]] 165; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 166; CHECK: scalar.ph: 167; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ], [ 0, [[VECTOR_SCEVCHECK]] ], [ 0, [[VECTOR_MEMCHECK]] ] 168; CHECK-NEXT: br label [[LOOP:%.*]] 169; CHECK: loop: 170; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 171; CHECK-NEXT: [[CLAMPED_INDEX:%.*]] = urem i32 [[IV]], 4 172; CHECK-NEXT: [[GEP_B:%.*]] = getelementptr inbounds i32, i32* [[B]], i32 [[IV]] 173; CHECK-NEXT: [[LV:%.*]] = load i32, i32* [[GEP_B]], align 4 174; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV]], 10 175; 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: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1 178; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]] 179; CHECK-NEXT: br i1 [[COND]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP14:![0-9]+]] 180; CHECK: exit: 181; CHECK-NEXT: ret void 182; 183entry: 184 br label %loop 185 186loop: 187 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 188 %clamped.index = urem i32 %iv, 4 189 %gep.B = getelementptr inbounds i32, i32* %B, i32 %iv 190 %lv = load i32, i32* %gep.B 191 %add = add i32 %lv, 10 192 %gep.A = getelementptr inbounds i32, i32* %A, i32 %clamped.index 193 store i32 %add, i32* %gep.A 194 %iv.next = add nuw nsw i32 %iv, 1 195 %cond = icmp eq i32 %iv.next, %N 196 br i1 %cond, label %exit, label %loop 197 198exit: 199 ret void 200} 201 202define void @clamped_index_dependence_non_clamped(i32* %A, i32* %B, i32 %N) { 203; CHECK-LABEL: @clamped_index_dependence_non_clamped( 204; CHECK-NEXT: entry: 205; CHECK-NEXT: br label [[LOOP:%.*]] 206; CHECK: loop: 207; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 208; CHECK-NEXT: [[GEP_B:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i32 [[IV]] 209; CHECK-NEXT: [[LV:%.*]] = load i32, i32* [[GEP_B]], align 4 210; CHECK-NEXT: [[GEP_A_1:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i32 [[IV]] 211; CHECK-NEXT: [[LV_A:%.*]] = load i32, i32* [[GEP_A_1]], align 4 212; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV]], [[LV_A]] 213; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1 214; CHECK-NEXT: [[CLAMPED_INDEX:%.*]] = urem i32 [[IV_NEXT]], 4 215; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[CLAMPED_INDEX]] 216; CHECK-NEXT: store i32 [[ADD]], i32* [[GEP_A]], align 4 217; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV_NEXT]], [[N:%.*]] 218; CHECK-NEXT: br i1 [[COND]], label [[EXIT:%.*]], label [[LOOP]] 219; CHECK: exit: 220; CHECK-NEXT: ret void 221; 222entry: 223 br label %loop 224 225loop: 226 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 227 %gep.B = getelementptr inbounds i32, i32* %B, i32 %iv 228 %lv = load i32, i32* %gep.B 229 %gep.A.1 = getelementptr inbounds i32, i32* %A, i32 %iv 230 %lv.A = load i32, i32* %gep.A.1 231 %add = add i32 %lv, %lv.A 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_dependence_clamped_index(i32* %A, i32* %B, i32 %N) { 245; CHECK-LABEL: @clamped_index_dependence_clamped_index( 246; CHECK-NEXT: entry: 247; CHECK-NEXT: br label [[LOOP:%.*]] 248; CHECK: loop: 249; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 250; CHECK-NEXT: [[CLAMPED_INDEX_1:%.*]] = urem i32 [[IV]], 4 251; CHECK-NEXT: [[GEP_A_1:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i32 [[CLAMPED_INDEX_1]] 252; CHECK-NEXT: [[LV_A:%.*]] = load i32, i32* [[GEP_A_1]], align 4 253; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV_A]], 10 254; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1 255; CHECK-NEXT: [[CLAMPED_INDEX:%.*]] = urem i32 [[IV_NEXT]], 4 256; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[CLAMPED_INDEX]] 257; CHECK-NEXT: store i32 [[ADD]], i32* [[GEP_A]], align 4 258; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV_NEXT]], [[N:%.*]] 259; CHECK-NEXT: br i1 [[COND]], label [[EXIT:%.*]], label [[LOOP]] 260; CHECK: exit: 261; CHECK-NEXT: ret void 262; 263entry: 264 br label %loop 265 266loop: 267 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 268 %clamped.index.1 = urem i32 %iv, 4 269 %gep.A.1 = getelementptr inbounds i32, i32* %A, i32 %clamped.index.1 270 %lv.A = load i32, i32* %gep.A.1 271 %add = add i32 %lv.A, 10 272 273 %iv.next = add nuw nsw i32 %iv, 1 274 %clamped.index = urem i32 %iv.next, 4 275 %gep.A = getelementptr inbounds i32, i32* %A, i32 %clamped.index 276 store i32 %add, i32* %gep.A 277 %cond = icmp eq i32 %iv.next, %N 278 br i1 %cond, label %exit, label %loop 279 280exit: 281 ret void 282} 283 284define void @clamped_index_equal_dependence(i32* %A, i32* %B, i32 %N) { 285; CHECK-LABEL: @clamped_index_equal_dependence( 286; CHECK-NEXT: entry: 287; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N:%.*]], 2 288; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]] 289; CHECK: vector.scevcheck: 290; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1 291; CHECK-NEXT: [[TMP1:%.*]] = trunc i32 [[TMP0]] to i2 292; CHECK-NEXT: [[MUL:%.*]] = call { i2, i1 } @llvm.umul.with.overflow.i2(i2 1, i2 [[TMP1]]) 293; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i2, i1 } [[MUL]], 0 294; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i2, i1 } [[MUL]], 1 295; CHECK-NEXT: [[TMP2:%.*]] = add i2 0, [[MUL_RESULT]] 296; CHECK-NEXT: [[TMP3:%.*]] = sub i2 0, [[MUL_RESULT]] 297; CHECK-NEXT: [[TMP4:%.*]] = icmp ugt i2 [[TMP3]], 0 298; CHECK-NEXT: [[TMP5:%.*]] = icmp ult i2 [[TMP2]], 0 299; CHECK-NEXT: [[TMP6:%.*]] = select i1 false, i1 [[TMP4]], i1 [[TMP5]] 300; CHECK-NEXT: [[TMP7:%.*]] = icmp ugt i32 [[TMP0]], 3 301; CHECK-NEXT: [[TMP8:%.*]] = or i1 [[TMP6]], [[TMP7]] 302; CHECK-NEXT: [[TMP9:%.*]] = or i1 [[TMP8]], [[MUL_OVERFLOW]] 303; CHECK-NEXT: [[TMP10:%.*]] = or i1 false, [[TMP9]] 304; CHECK-NEXT: br i1 [[TMP10]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 305; CHECK: vector.ph: 306; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N]], 2 307; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[N]], [[N_MOD_VF]] 308; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 309; CHECK: vector.body: 310; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 311; CHECK-NEXT: [[TMP11:%.*]] = add i32 [[INDEX]], 0 312; CHECK-NEXT: [[TMP12:%.*]] = urem i32 [[TMP11]], 4 313; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i32 [[TMP12]] 314; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds i32, i32* [[TMP13]], i32 0 315; CHECK-NEXT: [[TMP15:%.*]] = bitcast i32* [[TMP14]] to <2 x i32>* 316; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, <2 x i32>* [[TMP15]], align 4 317; CHECK-NEXT: [[TMP16:%.*]] = add <2 x i32> [[WIDE_LOAD]], <i32 10, i32 10> 318; CHECK-NEXT: [[TMP17:%.*]] = bitcast i32* [[TMP14]] to <2 x i32>* 319; CHECK-NEXT: store <2 x i32> [[TMP16]], <2 x i32>* [[TMP17]], align 4 320; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2 321; CHECK-NEXT: [[TMP18:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 322; CHECK-NEXT: br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP15:![0-9]+]] 323; CHECK: middle.block: 324; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]] 325; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 326; CHECK: scalar.ph: 327; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ], [ 0, [[VECTOR_SCEVCHECK]] ] 328; CHECK-NEXT: br label [[LOOP:%.*]] 329; CHECK: loop: 330; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 331; CHECK-NEXT: [[CLAMPED_INDEX:%.*]] = urem i32 [[IV]], 4 332; CHECK-NEXT: [[GEP_A:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 [[CLAMPED_INDEX]] 333; CHECK-NEXT: [[LV_A:%.*]] = load i32, i32* [[GEP_A]], align 4 334; CHECK-NEXT: [[ADD:%.*]] = add i32 [[LV_A]], 10 335; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1 336; CHECK-NEXT: store i32 [[ADD]], i32* [[GEP_A]], align 4 337; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]] 338; CHECK-NEXT: br i1 [[COND]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP16:![0-9]+]] 339; CHECK: exit: 340; CHECK-NEXT: ret void 341; 342entry: 343 br label %loop 344 345loop: 346 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 347 %clamped.index = urem i32 %iv, 4 348 %gep.A = getelementptr inbounds i32, i32* %A, i32 %clamped.index 349 %lv.A = load i32, i32* %gep.A 350 %add = add i32 %lv.A, 10 351 352 %iv.next = add nuw nsw i32 %iv, 1 353 store i32 %add, i32* %gep.A 354 %cond = icmp eq i32 %iv.next, %N 355 br i1 %cond, label %exit, label %loop 356 357exit: 358 ret void 359} 360