1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -passes=loop-vectorize -force-vector-width=4 -hoist-runtime-checks=false -S < %s | FileCheck %s --check-prefix=CHECK 3; RUN: opt -passes=loop-vectorize -force-vector-width=4 -S < %s | FileCheck %s --check-prefix=CHECK-HOIST 4 5; This is the test case from PR26314. 6; When we were retrying dependence checking with memchecks only, 7; the loop-invariant access in the inner loop was incorrectly determined to be wrapping 8; because it was not strided in the inner loop. 9; Improved wrapping detection allows vectorization in the following case. 10 11; #define Z 32 12; typedef struct s { 13; int v1[Z]; 14; int v2[Z]; 15; int v3[Z][Z]; 16; } s; 17; 18; void slow_function (s* const obj, int z) { 19; for (int j=0; j<Z; j++) { 20; for (int k=0; k<z; k++) { 21; int x = obj->v1[k] + obj->v2[j]; 22; obj->v3[j][k] += x; 23; } 24; } 25; } 26 27target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" 28 29%struct.s = type { [32 x i32], [32 x i32], [32 x [32 x i32]] } 30 31define void @Test(ptr nocapture %obj, i64 %z) #0 { 32; CHECK-LABEL: @Test( 33; CHECK-NEXT: [[TMP1:%.*]] = shl i64 [[Z:%.*]], 2 34; CHECK-NEXT: [[TMP2:%.*]] = add i64 [[TMP1]], 256 35; CHECK-NEXT: [[SCEVGEP2:%.*]] = getelementptr i8, ptr [[OBJ:%.*]], i64 [[TMP1]] 36; CHECK-NEXT: br label [[DOTOUTER_PREHEADER:%.*]] 37; CHECK: .outer.preheader: 38; CHECK-NEXT: [[I:%.*]] = phi i64 [ 0, [[TMP0:%.*]] ], [ [[I_NEXT:%.*]], [[DOTOUTER:%.*]] ] 39; CHECK-NEXT: [[TMP3:%.*]] = shl nuw nsw i64 [[I]], 7 40; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[TMP3]], 256 41; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr nuw i8, ptr [[OBJ]], i64 [[TMP4]] 42; CHECK-NEXT: [[TMP5:%.*]] = add i64 [[TMP2]], [[TMP3]] 43; CHECK-NEXT: [[SCEVGEP1:%.*]] = getelementptr i8, ptr [[OBJ]], i64 [[TMP5]] 44; CHECK-NEXT: [[TMP6:%.*]] = shl nuw nsw i64 [[I]], 2 45; CHECK-NEXT: [[TMP7:%.*]] = add i64 [[TMP6]], 128 46; CHECK-NEXT: [[SCEVGEP3:%.*]] = getelementptr nuw i8, ptr [[OBJ]], i64 [[TMP7]] 47; CHECK-NEXT: [[TMP8:%.*]] = add i64 [[TMP6]], 132 48; CHECK-NEXT: [[SCEVGEP4:%.*]] = getelementptr i8, ptr [[OBJ]], i64 [[TMP8]] 49; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], ptr [[OBJ]], i64 0, i32 1, i64 [[I]] 50; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[Z]], 4 51; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_MEMCHECK:%.*]] 52; CHECK: vector.memcheck: 53; CHECK-NEXT: [[BOUND0:%.*]] = icmp ult ptr [[SCEVGEP]], [[SCEVGEP2]] 54; CHECK-NEXT: [[BOUND1:%.*]] = icmp ult ptr [[OBJ]], [[SCEVGEP1]] 55; CHECK-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]] 56; CHECK-NEXT: [[BOUND05:%.*]] = icmp ult ptr [[SCEVGEP]], [[SCEVGEP4]] 57; CHECK-NEXT: [[BOUND16:%.*]] = icmp ult ptr [[SCEVGEP3]], [[SCEVGEP1]] 58; CHECK-NEXT: [[FOUND_CONFLICT7:%.*]] = and i1 [[BOUND05]], [[BOUND16]] 59; CHECK-NEXT: [[CONFLICT_RDX:%.*]] = or i1 [[FOUND_CONFLICT]], [[FOUND_CONFLICT7]] 60; CHECK-NEXT: br i1 [[CONFLICT_RDX]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 61; CHECK: vector.ph: 62; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[Z]], 4 63; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[Z]], [[N_MOD_VF]] 64; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 65; CHECK: vector.body: 66; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 67; CHECK-NEXT: [[TMP10:%.*]] = add i64 [[INDEX]], 0 68; CHECK-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 0, i64 [[TMP10]] 69; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds i32, ptr [[TMP11]], i32 0 70; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP12]], align 4, !alias.scope [[META0:![0-9]+]] 71; CHECK-NEXT: [[TMP13:%.*]] = load i32, ptr [[TMP9]], align 4, !alias.scope [[META3:![0-9]+]] 72; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TMP13]], i64 0 73; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer 74; CHECK-NEXT: [[TMP14:%.*]] = add nsw <4 x i32> [[BROADCAST_SPLAT]], [[WIDE_LOAD]] 75; CHECK-NEXT: [[TMP15:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[TMP10]] 76; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, ptr [[TMP15]], i32 0 77; CHECK-NEXT: [[WIDE_LOAD8:%.*]] = load <4 x i32>, ptr [[TMP16]], align 4, !alias.scope [[META5:![0-9]+]], !noalias [[META7:![0-9]+]] 78; CHECK-NEXT: [[TMP17:%.*]] = add nsw <4 x i32> [[TMP14]], [[WIDE_LOAD8]] 79; CHECK-NEXT: store <4 x i32> [[TMP17]], ptr [[TMP16]], align 4, !alias.scope [[META5]], !noalias [[META7]] 80; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4 81; CHECK-NEXT: [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 82; CHECK-NEXT: br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]] 83; CHECK: middle.block: 84; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[Z]], [[N_VEC]] 85; CHECK-NEXT: br i1 [[CMP_N]], label [[DOTOUTER]], label [[SCALAR_PH]] 86; CHECK: scalar.ph: 87; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[VECTOR_MEMCHECK]] ], [ 0, [[DOTOUTER_PREHEADER]] ] 88; CHECK-NEXT: br label [[DOTINNER:%.*]] 89; CHECK: .exit: 90; CHECK-NEXT: ret void 91; CHECK: .outer: 92; CHECK-NEXT: [[I_NEXT]] = add nuw nsw i64 [[I]], 1 93; CHECK-NEXT: [[EXITCOND_OUTER:%.*]] = icmp eq i64 [[I_NEXT]], 32 94; CHECK-NEXT: br i1 [[EXITCOND_OUTER]], label [[DOTEXIT:%.*]], label [[DOTOUTER_PREHEADER]] 95; CHECK: .inner: 96; CHECK-NEXT: [[J:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[J_NEXT:%.*]], [[DOTINNER]] ] 97; CHECK-NEXT: [[TMP19:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 0, i64 [[J]] 98; CHECK-NEXT: [[TMP20:%.*]] = load i32, ptr [[TMP19]], align 4 99; CHECK-NEXT: [[TMP21:%.*]] = load i32, ptr [[TMP9]], align 4 100; CHECK-NEXT: [[TMP22:%.*]] = add nsw i32 [[TMP21]], [[TMP20]] 101; CHECK-NEXT: [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[J]] 102; CHECK-NEXT: [[TMP24:%.*]] = load i32, ptr [[TMP23]], align 4 103; CHECK-NEXT: [[TMP25:%.*]] = add nsw i32 [[TMP22]], [[TMP24]] 104; CHECK-NEXT: store i32 [[TMP25]], ptr [[TMP23]], align 4 105; CHECK-NEXT: [[J_NEXT]] = add nuw nsw i64 [[J]], 1 106; CHECK-NEXT: [[EXITCOND_INNER:%.*]] = icmp eq i64 [[J_NEXT]], [[Z]] 107; CHECK-NEXT: br i1 [[EXITCOND_INNER]], label [[DOTOUTER]], label [[DOTINNER]], !llvm.loop [[LOOP11:![0-9]+]] 108; 109; CHECK-HOIST-LABEL: @Test( 110; CHECK-HOIST-NEXT: [[SCEVGEP:%.*]] = getelementptr nuw i8, ptr [[OBJ:%.*]], i64 256 111; CHECK-HOIST-NEXT: [[TMP1:%.*]] = shl i64 [[Z:%.*]], 2 112; CHECK-HOIST-NEXT: [[TMP2:%.*]] = add i64 [[TMP1]], 4224 113; CHECK-HOIST-NEXT: [[SCEVGEP1:%.*]] = getelementptr i8, ptr [[OBJ]], i64 [[TMP2]] 114; CHECK-HOIST-NEXT: [[SCEVGEP2:%.*]] = getelementptr i8, ptr [[OBJ]], i64 [[TMP1]] 115; CHECK-HOIST-NEXT: [[SCEVGEP3:%.*]] = getelementptr nuw i8, ptr [[OBJ]], i64 128 116; CHECK-HOIST-NEXT: br label [[DOTOUTER_PREHEADER:%.*]] 117; CHECK-HOIST: .outer.preheader: 118; CHECK-HOIST-NEXT: [[I:%.*]] = phi i64 [ 0, [[TMP0:%.*]] ], [ [[I_NEXT:%.*]], [[DOTOUTER:%.*]] ] 119; CHECK-HOIST-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], ptr [[OBJ]], i64 0, i32 1, i64 [[I]] 120; CHECK-HOIST-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[Z]], 4 121; CHECK-HOIST-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_MEMCHECK:%.*]] 122; CHECK-HOIST: vector.memcheck: 123; CHECK-HOIST-NEXT: [[BOUND0:%.*]] = icmp ult ptr [[SCEVGEP]], [[SCEVGEP2]] 124; CHECK-HOIST-NEXT: [[BOUND1:%.*]] = icmp ult ptr [[OBJ]], [[SCEVGEP1]] 125; CHECK-HOIST-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]] 126; CHECK-HOIST-NEXT: [[BOUND14:%.*]] = icmp ult ptr [[SCEVGEP3]], [[SCEVGEP1]] 127; CHECK-HOIST-NEXT: br i1 [[FOUND_CONFLICT]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 128; CHECK-HOIST: vector.ph: 129; CHECK-HOIST-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[Z]], 4 130; CHECK-HOIST-NEXT: [[N_VEC:%.*]] = sub i64 [[Z]], [[N_MOD_VF]] 131; CHECK-HOIST-NEXT: br label [[VECTOR_BODY:%.*]] 132; CHECK-HOIST: vector.body: 133; CHECK-HOIST-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 134; CHECK-HOIST-NEXT: [[TMP4:%.*]] = add i64 [[INDEX]], 0 135; CHECK-HOIST-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 0, i64 [[TMP4]] 136; CHECK-HOIST-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[TMP5]], i32 0 137; CHECK-HOIST-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP6]], align 4, !alias.scope [[META0:![0-9]+]] 138; CHECK-HOIST-NEXT: [[TMP7:%.*]] = load i32, ptr [[TMP3]], align 4, !alias.scope [[META3:![0-9]+]] 139; CHECK-HOIST-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TMP7]], i64 0 140; CHECK-HOIST-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer 141; CHECK-HOIST-NEXT: [[TMP8:%.*]] = add nsw <4 x i32> [[BROADCAST_SPLAT]], [[WIDE_LOAD]] 142; CHECK-HOIST-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[TMP4]] 143; CHECK-HOIST-NEXT: [[TMP10:%.*]] = getelementptr inbounds i32, ptr [[TMP9]], i32 0 144; CHECK-HOIST-NEXT: [[WIDE_LOAD5:%.*]] = load <4 x i32>, ptr [[TMP10]], align 4, !alias.scope [[META5:![0-9]+]], !noalias [[META7:![0-9]+]] 145; CHECK-HOIST-NEXT: [[TMP11:%.*]] = add nsw <4 x i32> [[TMP8]], [[WIDE_LOAD5]] 146; CHECK-HOIST-NEXT: store <4 x i32> [[TMP11]], ptr [[TMP10]], align 4, !alias.scope [[META5]], !noalias [[META7]] 147; CHECK-HOIST-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4 148; CHECK-HOIST-NEXT: [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 149; CHECK-HOIST-NEXT: br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]] 150; CHECK-HOIST: middle.block: 151; CHECK-HOIST-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[Z]], [[N_VEC]] 152; CHECK-HOIST-NEXT: br i1 [[CMP_N]], label [[DOTOUTER]], label [[SCALAR_PH]] 153; CHECK-HOIST: scalar.ph: 154; CHECK-HOIST-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[VECTOR_MEMCHECK]] ], [ 0, [[DOTOUTER_PREHEADER]] ] 155; CHECK-HOIST-NEXT: br label [[DOTINNER:%.*]] 156; CHECK-HOIST: .exit: 157; CHECK-HOIST-NEXT: ret void 158; CHECK-HOIST: .outer: 159; CHECK-HOIST-NEXT: [[I_NEXT]] = add nuw nsw i64 [[I]], 1 160; CHECK-HOIST-NEXT: [[EXITCOND_OUTER:%.*]] = icmp eq i64 [[I_NEXT]], 32 161; CHECK-HOIST-NEXT: br i1 [[EXITCOND_OUTER]], label [[DOTEXIT:%.*]], label [[DOTOUTER_PREHEADER]] 162; CHECK-HOIST: .inner: 163; CHECK-HOIST-NEXT: [[J:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[J_NEXT:%.*]], [[DOTINNER]] ] 164; CHECK-HOIST-NEXT: [[TMP13:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 0, i64 [[J]] 165; CHECK-HOIST-NEXT: [[TMP14:%.*]] = load i32, ptr [[TMP13]], align 4 166; CHECK-HOIST-NEXT: [[TMP15:%.*]] = load i32, ptr [[TMP3]], align 4 167; CHECK-HOIST-NEXT: [[TMP16:%.*]] = add nsw i32 [[TMP15]], [[TMP14]] 168; CHECK-HOIST-NEXT: [[TMP17:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[J]] 169; CHECK-HOIST-NEXT: [[TMP18:%.*]] = load i32, ptr [[TMP17]], align 4 170; CHECK-HOIST-NEXT: [[TMP19:%.*]] = add nsw i32 [[TMP16]], [[TMP18]] 171; CHECK-HOIST-NEXT: store i32 [[TMP19]], ptr [[TMP17]], align 4 172; CHECK-HOIST-NEXT: [[J_NEXT]] = add nuw nsw i64 [[J]], 1 173; CHECK-HOIST-NEXT: [[EXITCOND_INNER:%.*]] = icmp eq i64 [[J_NEXT]], [[Z]] 174; CHECK-HOIST-NEXT: br i1 [[EXITCOND_INNER]], label [[DOTOUTER]], label [[DOTINNER]], !llvm.loop [[LOOP11:![0-9]+]] 175; 176 br label %.outer.preheader 177 178 179.outer.preheader: 180 %i = phi i64 [ 0, %0 ], [ %i.next, %.outer ] 181 %1 = getelementptr inbounds %struct.s, ptr %obj, i64 0, i32 1, i64 %i 182 br label %.inner 183 184.exit: 185 ret void 186 187.outer: 188 %i.next = add nuw nsw i64 %i, 1 189 %exitcond.outer = icmp eq i64 %i.next, 32 190 br i1 %exitcond.outer, label %.exit, label %.outer.preheader 191 192.inner: 193 %j = phi i64 [ 0, %.outer.preheader ], [ %j.next, %.inner ] 194 %2 = getelementptr inbounds %struct.s, ptr %obj, i64 0, i32 0, i64 %j 195 %3 = load i32, ptr %2 196 %4 = load i32, ptr %1 197 %5 = add nsw i32 %4, %3 198 %6 = getelementptr inbounds %struct.s, ptr %obj, i64 0, i32 2, i64 %i, i64 %j 199 %7 = load i32, ptr %6 200 %8 = add nsw i32 %5, %7 201 store i32 %8, ptr %6 202 %j.next = add nuw nsw i64 %j, 1 203 %exitcond.inner = icmp eq i64 %j.next, %z 204 br i1 %exitcond.inner, label %.outer, label %.inner 205} 206