1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -loop-vectorize -force-vector-width=4 -S < %s | FileCheck %s 3 4; This is the test case from PR26314. 5; When we were retrying dependence checking with memchecks only, 6; the loop-invariant access in the inner loop was incorrectly determined to be wrapping 7; because it was not strided in the inner loop. 8; Improved wrapping detection allows vectorization in the following case. 9 10; #define Z 32 11; typedef struct s { 12; int v1[Z]; 13; int v2[Z]; 14; int v3[Z][Z]; 15; } s; 16; 17; void slow_function (s* const obj, int z) { 18; for (int j=0; j<Z; j++) { 19; for (int k=0; k<z; k++) { 20; int x = obj->v1[k] + obj->v2[j]; 21; obj->v3[j][k] += x; 22; } 23; } 24; } 25 26target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" 27 28%struct.s = type { [32 x i32], [32 x i32], [32 x [32 x i32]] } 29 30define void @Test(%struct.s* nocapture %obj, i64 %z) #0 { 31; CHECK-LABEL: @Test( 32; CHECK-NEXT: [[OBJ4:%.*]] = bitcast %struct.s* [[OBJ:%.*]] to i8* 33; CHECK-NEXT: [[SCEVGEP5:%.*]] = getelementptr [[STRUCT_S:%.*]], %struct.s* [[OBJ]], i64 0, i32 0, i64 [[Z:%.*]] 34; CHECK-NEXT: [[SCEVGEP56:%.*]] = bitcast i32* [[SCEVGEP5]] to i8* 35; CHECK-NEXT: br label [[DOTOUTER_PREHEADER:%.*]] 36; CHECK: .outer.preheader: 37; CHECK-NEXT: [[I:%.*]] = phi i64 [ 0, [[TMP0:%.*]] ], [ [[I_NEXT:%.*]], [[DOTOUTER:%.*]] ] 38; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 2, i64 [[I]], i64 0 39; CHECK-NEXT: [[SCEVGEP1:%.*]] = bitcast i32* [[SCEVGEP]] to i8* 40; CHECK-NEXT: [[SCEVGEP2:%.*]] = getelementptr [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[Z]] 41; CHECK-NEXT: [[SCEVGEP23:%.*]] = bitcast i32* [[SCEVGEP2]] to i8* 42; CHECK-NEXT: [[SCEVGEP7:%.*]] = getelementptr [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 1, i64 [[I]] 43; CHECK-NEXT: [[SCEVGEP78:%.*]] = bitcast i32* [[SCEVGEP7]] to i8* 44; CHECK-NEXT: [[UGLYGEP:%.*]] = getelementptr i8, i8* [[SCEVGEP78]], i64 1 45; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 1, i64 [[I]] 46; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[Z]], 4 47; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_MEMCHECK:%.*]] 48; CHECK: vector.memcheck: 49; CHECK-NEXT: [[BOUND0:%.*]] = icmp ult i8* [[SCEVGEP1]], [[SCEVGEP56]] 50; CHECK-NEXT: [[BOUND1:%.*]] = icmp ult i8* [[OBJ4]], [[SCEVGEP23]] 51; CHECK-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]] 52; CHECK-NEXT: [[BC:%.*]] = bitcast i32* [[TMP1]] to i8* 53; CHECK-NEXT: [[BOUND09:%.*]] = icmp ult i8* [[SCEVGEP1]], [[UGLYGEP]] 54; CHECK-NEXT: [[BOUND110:%.*]] = icmp ult i8* [[BC]], [[SCEVGEP23]] 55; CHECK-NEXT: [[FOUND_CONFLICT11:%.*]] = and i1 [[BOUND09]], [[BOUND110]] 56; CHECK-NEXT: [[CONFLICT_RDX:%.*]] = or i1 [[FOUND_CONFLICT]], [[FOUND_CONFLICT11]] 57; CHECK-NEXT: [[MEMCHECK_CONFLICT:%.*]] = and i1 [[CONFLICT_RDX]], true 58; CHECK-NEXT: br i1 [[MEMCHECK_CONFLICT]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 59; CHECK: vector.ph: 60; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[Z]], 4 61; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[Z]], [[N_MOD_VF]] 62; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 63; CHECK: vector.body: 64; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 65; CHECK-NEXT: [[TMP2:%.*]] = add i64 [[INDEX]], 0 66; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 0, i64 [[TMP2]] 67; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds i32, i32* [[TMP3]], i32 0 68; CHECK-NEXT: [[TMP5:%.*]] = bitcast i32* [[TMP4]] to <4 x i32>* 69; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP5]], align 4, !alias.scope !0 70; CHECK-NEXT: [[TMP6:%.*]] = load i32, i32* [[TMP1]], align 4, !alias.scope !3 71; CHECK-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP1]], align 4, !alias.scope !3 72; CHECK-NEXT: [[TMP8:%.*]] = load i32, i32* [[TMP1]], align 4, !alias.scope !3 73; CHECK-NEXT: [[TMP9:%.*]] = load i32, i32* [[TMP1]], align 4, !alias.scope !3 74; CHECK-NEXT: [[TMP10:%.*]] = insertelement <4 x i32> undef, i32 [[TMP6]], i32 0 75; CHECK-NEXT: [[TMP11:%.*]] = insertelement <4 x i32> [[TMP10]], i32 [[TMP7]], i32 1 76; CHECK-NEXT: [[TMP12:%.*]] = insertelement <4 x i32> [[TMP11]], i32 [[TMP8]], i32 2 77; CHECK-NEXT: [[TMP13:%.*]] = insertelement <4 x i32> [[TMP12]], i32 [[TMP9]], i32 3 78; CHECK-NEXT: [[TMP14:%.*]] = add nsw <4 x i32> [[TMP13]], [[WIDE_LOAD]] 79; CHECK-NEXT: [[TMP15:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[TMP2]] 80; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, i32* [[TMP15]], i32 0 81; CHECK-NEXT: [[TMP17:%.*]] = bitcast i32* [[TMP16]] to <4 x i32>* 82; CHECK-NEXT: [[WIDE_LOAD12:%.*]] = load <4 x i32>, <4 x i32>* [[TMP17]], align 4, !alias.scope !5, !noalias !7 83; CHECK-NEXT: [[TMP18:%.*]] = add nsw <4 x i32> [[TMP14]], [[WIDE_LOAD12]] 84; CHECK-NEXT: [[TMP19:%.*]] = bitcast i32* [[TMP16]] to <4 x i32>* 85; CHECK-NEXT: store <4 x i32> [[TMP18]], <4 x i32>* [[TMP19]], align 4, !alias.scope !5, !noalias !7 86; CHECK-NEXT: [[INDEX_NEXT]] = add i64 [[INDEX]], 4 87; CHECK-NEXT: [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 88; CHECK-NEXT: br i1 [[TMP20]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !8 89; CHECK: middle.block: 90; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[Z]], [[N_VEC]] 91; CHECK-NEXT: br i1 [[CMP_N]], label [[DOTOUTER]], label [[SCALAR_PH]] 92; CHECK: scalar.ph: 93; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[DOTOUTER_PREHEADER]] ], [ 0, [[VECTOR_MEMCHECK]] ] 94; CHECK-NEXT: br label [[DOTINNER:%.*]] 95; CHECK: .exit: 96; CHECK-NEXT: ret void 97; CHECK: .outer: 98; CHECK-NEXT: [[I_NEXT]] = add nuw nsw i64 [[I]], 1 99; CHECK-NEXT: [[EXITCOND_OUTER:%.*]] = icmp eq i64 [[I_NEXT]], 32 100; CHECK-NEXT: br i1 [[EXITCOND_OUTER]], label [[DOTEXIT:%.*]], label [[DOTOUTER_PREHEADER]] 101; CHECK: .inner: 102; CHECK-NEXT: [[J:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[J_NEXT:%.*]], [[DOTINNER]] ] 103; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 0, i64 [[J]] 104; CHECK-NEXT: [[TMP22:%.*]] = load i32, i32* [[TMP21]], align 4 105; CHECK-NEXT: [[TMP23:%.*]] = load i32, i32* [[TMP1]], align 4 106; CHECK-NEXT: [[TMP24:%.*]] = add nsw i32 [[TMP23]], [[TMP22]] 107; CHECK-NEXT: [[TMP25:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[J]] 108; CHECK-NEXT: [[TMP26:%.*]] = load i32, i32* [[TMP25]], align 4 109; CHECK-NEXT: [[TMP27:%.*]] = add nsw i32 [[TMP24]], [[TMP26]] 110; CHECK-NEXT: store i32 [[TMP27]], i32* [[TMP25]] 111; CHECK-NEXT: [[J_NEXT]] = add nuw nsw i64 [[J]], 1 112; CHECK-NEXT: [[EXITCOND_INNER:%.*]] = icmp eq i64 [[J_NEXT]], [[Z]] 113; CHECK-NEXT: br i1 [[EXITCOND_INNER]], label [[DOTOUTER]], label [[DOTINNER]], !llvm.loop !10 114; 115 br label %.outer.preheader 116 117 118.outer.preheader: 119 %i = phi i64 [ 0, %0 ], [ %i.next, %.outer ] 120 %1 = getelementptr inbounds %struct.s, %struct.s* %obj, i64 0, i32 1, i64 %i 121 br label %.inner 122 123.exit: 124 ret void 125 126.outer: 127 %i.next = add nuw nsw i64 %i, 1 128 %exitcond.outer = icmp eq i64 %i.next, 32 129 br i1 %exitcond.outer, label %.exit, label %.outer.preheader 130 131.inner: 132 %j = phi i64 [ 0, %.outer.preheader ], [ %j.next, %.inner ] 133 %2 = getelementptr inbounds %struct.s, %struct.s* %obj, i64 0, i32 0, i64 %j 134 %3 = load i32, i32* %2 135 %4 = load i32, i32* %1 136 %5 = add nsw i32 %4, %3 137 %6 = getelementptr inbounds %struct.s, %struct.s* %obj, i64 0, i32 2, i64 %i, i64 %j 138 %7 = load i32, i32* %6 139 %8 = add nsw i32 %5, %7 140 store i32 %8, i32* %6 141 %j.next = add nuw nsw i64 %j, 1 142 %exitcond.inner = icmp eq i64 %j.next, %z 143 br i1 %exitcond.inner, label %.outer, label %.inner 144} 145