200 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=loop-vectorize -force-vector-width=4 -hoist-runtime-checks=false -S < %s | FileCheck %s --check-prefix=CHECK3; RUN: opt -passes=loop-vectorize -force-vector-width=4 -S < %s | FileCheck %s --check-prefix=CHECK-HOIST4 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 wrapping8; because it was not strided in the inner loop.9; Improved wrapping detection allows vectorization in the following case.10 11; #define Z 3212; 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:%.*]], 234; CHECK-NEXT: [[TMP2:%.*]] = add i64 [[TMP1]], 25635; 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]], 740; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[TMP3]], 25641; 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]], 245; CHECK-NEXT: [[TMP7:%.*]] = add i64 [[TMP6]], 12846; CHECK-NEXT: [[SCEVGEP3:%.*]] = getelementptr nuw i8, ptr [[OBJ]], i64 [[TMP7]]47; CHECK-NEXT: [[TMP8:%.*]] = add i64 [[TMP6]], 13248; 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]], 451; 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]], 463; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[Z]], [[N_MOD_VF]]64; CHECK-NEXT: [[TMP10:%.*]] = load i32, ptr [[TMP9]], align 4, !alias.scope [[META0:![0-9]+]]65; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TMP10]], i64 066; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer67; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]68; CHECK: vector.body:69; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]70; CHECK-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 0, i64 [[INDEX]]71; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP11]], align 4, !alias.scope [[META3:![0-9]+]]72; CHECK-NEXT: [[TMP14:%.*]] = add nsw <4 x i32> [[BROADCAST_SPLAT]], [[WIDE_LOAD]]73; CHECK-NEXT: [[TMP15:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[INDEX]]74; CHECK-NEXT: [[WIDE_LOAD8:%.*]] = load <4 x i32>, ptr [[TMP15]], align 4, !alias.scope [[META5:![0-9]+]], !noalias [[META7:![0-9]+]]75; CHECK-NEXT: [[TMP17:%.*]] = add nsw <4 x i32> [[TMP14]], [[WIDE_LOAD8]]76; CHECK-NEXT: store <4 x i32> [[TMP17]], ptr [[TMP15]], align 4, !alias.scope [[META5]], !noalias [[META7]]77; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 478; CHECK-NEXT: [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]79; CHECK-NEXT: br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]80; CHECK: middle.block:81; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[Z]], [[N_VEC]]82; CHECK-NEXT: br i1 [[CMP_N]], label [[DOTOUTER]], label [[SCALAR_PH]]83; CHECK: scalar.ph:84; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[DOTOUTER_PREHEADER]] ], [ 0, [[VECTOR_MEMCHECK]] ]85; CHECK-NEXT: br label [[DOTINNER:%.*]]86; CHECK: .exit:87; CHECK-NEXT: ret void88; CHECK: .outer:89; CHECK-NEXT: [[I_NEXT]] = add nuw nsw i64 [[I]], 190; CHECK-NEXT: [[EXITCOND_OUTER:%.*]] = icmp eq i64 [[I_NEXT]], 3291; CHECK-NEXT: br i1 [[EXITCOND_OUTER]], label [[DOTEXIT:%.*]], label [[DOTOUTER_PREHEADER]]92; CHECK: .inner:93; CHECK-NEXT: [[J:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[J_NEXT:%.*]], [[DOTINNER]] ]94; CHECK-NEXT: [[TMP19:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 0, i64 [[J]]95; CHECK-NEXT: [[TMP20:%.*]] = load i32, ptr [[TMP19]], align 496; CHECK-NEXT: [[TMP21:%.*]] = load i32, ptr [[TMP9]], align 497; CHECK-NEXT: [[TMP22:%.*]] = add nsw i32 [[TMP21]], [[TMP20]]98; CHECK-NEXT: [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[J]]99; CHECK-NEXT: [[TMP24:%.*]] = load i32, ptr [[TMP23]], align 4100; CHECK-NEXT: [[TMP25:%.*]] = add nsw i32 [[TMP22]], [[TMP24]]101; CHECK-NEXT: store i32 [[TMP25]], ptr [[TMP23]], align 4102; CHECK-NEXT: [[J_NEXT]] = add nuw nsw i64 [[J]], 1103; CHECK-NEXT: [[EXITCOND_INNER:%.*]] = icmp eq i64 [[J_NEXT]], [[Z]]104; CHECK-NEXT: br i1 [[EXITCOND_INNER]], label [[DOTOUTER]], label [[DOTINNER]], !llvm.loop [[LOOP11:![0-9]+]]105;106; CHECK-HOIST-LABEL: @Test(107; CHECK-HOIST-NEXT: [[SCEVGEP:%.*]] = getelementptr nuw i8, ptr [[OBJ:%.*]], i64 256108; CHECK-HOIST-NEXT: [[TMP1:%.*]] = shl i64 [[Z:%.*]], 2109; CHECK-HOIST-NEXT: [[TMP2:%.*]] = add i64 [[TMP1]], 4224110; CHECK-HOIST-NEXT: [[SCEVGEP1:%.*]] = getelementptr i8, ptr [[OBJ]], i64 [[TMP2]]111; CHECK-HOIST-NEXT: [[SCEVGEP2:%.*]] = getelementptr i8, ptr [[OBJ]], i64 [[TMP1]]112; CHECK-HOIST-NEXT: [[SCEVGEP3:%.*]] = getelementptr nuw i8, ptr [[OBJ]], i64 128113; CHECK-HOIST-NEXT: br label [[DOTOUTER_PREHEADER:%.*]]114; CHECK-HOIST: .outer.preheader:115; CHECK-HOIST-NEXT: [[I:%.*]] = phi i64 [ 0, [[TMP0:%.*]] ], [ [[I_NEXT:%.*]], [[DOTOUTER:%.*]] ]116; CHECK-HOIST-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], ptr [[OBJ]], i64 0, i32 1, i64 [[I]]117; CHECK-HOIST-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[Z]], 4118; CHECK-HOIST-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_MEMCHECK:%.*]]119; CHECK-HOIST: vector.memcheck:120; CHECK-HOIST-NEXT: [[BOUND0:%.*]] = icmp ult ptr [[SCEVGEP]], [[SCEVGEP2]]121; CHECK-HOIST-NEXT: [[BOUND1:%.*]] = icmp ult ptr [[OBJ]], [[SCEVGEP1]]122; CHECK-HOIST-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]123; CHECK-HOIST-NEXT: [[BOUND14:%.*]] = icmp ult ptr [[SCEVGEP3]], [[SCEVGEP1]]124; CHECK-HOIST-NEXT: br i1 [[FOUND_CONFLICT]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]125; CHECK-HOIST: vector.ph:126; CHECK-HOIST-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[Z]], 4127; CHECK-HOIST-NEXT: [[N_VEC:%.*]] = sub i64 [[Z]], [[N_MOD_VF]]128; CHECK-HOIST-NEXT: [[TMP4:%.*]] = load i32, ptr [[TMP3]], align 4, !alias.scope [[META0:![0-9]+]]129; CHECK-HOIST-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TMP4]], i64 0130; CHECK-HOIST-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer131; 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: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 0, i64 [[INDEX]]135; CHECK-HOIST-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP5]], align 4, !alias.scope [[META3:![0-9]+]]136; CHECK-HOIST-NEXT: [[TMP8:%.*]] = add nsw <4 x i32> [[BROADCAST_SPLAT]], [[WIDE_LOAD]]137; CHECK-HOIST-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[INDEX]]138; CHECK-HOIST-NEXT: [[WIDE_LOAD5:%.*]] = load <4 x i32>, ptr [[TMP9]], align 4, !alias.scope [[META5:![0-9]+]], !noalias [[META7:![0-9]+]]139; CHECK-HOIST-NEXT: [[TMP11:%.*]] = add nsw <4 x i32> [[TMP8]], [[WIDE_LOAD5]]140; CHECK-HOIST-NEXT: store <4 x i32> [[TMP11]], ptr [[TMP9]], align 4, !alias.scope [[META5]], !noalias [[META7]]141; CHECK-HOIST-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4142; CHECK-HOIST-NEXT: [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]143; CHECK-HOIST-NEXT: br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]144; CHECK-HOIST: middle.block:145; CHECK-HOIST-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[Z]], [[N_VEC]]146; CHECK-HOIST-NEXT: br i1 [[CMP_N]], label [[DOTOUTER]], label [[SCALAR_PH]]147; CHECK-HOIST: scalar.ph:148; CHECK-HOIST-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[DOTOUTER_PREHEADER]] ], [ 0, [[VECTOR_MEMCHECK]] ]149; CHECK-HOIST-NEXT: br label [[DOTINNER:%.*]]150; CHECK-HOIST: .exit:151; CHECK-HOIST-NEXT: ret void152; CHECK-HOIST: .outer:153; CHECK-HOIST-NEXT: [[I_NEXT]] = add nuw nsw i64 [[I]], 1154; CHECK-HOIST-NEXT: [[EXITCOND_OUTER:%.*]] = icmp eq i64 [[I_NEXT]], 32155; CHECK-HOIST-NEXT: br i1 [[EXITCOND_OUTER]], label [[DOTEXIT:%.*]], label [[DOTOUTER_PREHEADER]]156; CHECK-HOIST: .inner:157; CHECK-HOIST-NEXT: [[J:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[J_NEXT:%.*]], [[DOTINNER]] ]158; CHECK-HOIST-NEXT: [[TMP13:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 0, i64 [[J]]159; CHECK-HOIST-NEXT: [[TMP14:%.*]] = load i32, ptr [[TMP13]], align 4160; CHECK-HOIST-NEXT: [[TMP15:%.*]] = load i32, ptr [[TMP3]], align 4161; CHECK-HOIST-NEXT: [[TMP16:%.*]] = add nsw i32 [[TMP15]], [[TMP14]]162; CHECK-HOIST-NEXT: [[TMP17:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[J]]163; CHECK-HOIST-NEXT: [[TMP18:%.*]] = load i32, ptr [[TMP17]], align 4164; CHECK-HOIST-NEXT: [[TMP19:%.*]] = add nsw i32 [[TMP16]], [[TMP18]]165; CHECK-HOIST-NEXT: store i32 [[TMP19]], ptr [[TMP17]], align 4166; CHECK-HOIST-NEXT: [[J_NEXT]] = add nuw nsw i64 [[J]], 1167; CHECK-HOIST-NEXT: [[EXITCOND_INNER:%.*]] = icmp eq i64 [[J_NEXT]], [[Z]]168; CHECK-HOIST-NEXT: br i1 [[EXITCOND_INNER]], label [[DOTOUTER]], label [[DOTINNER]], !llvm.loop [[LOOP11:![0-9]+]]169;170 br label %.outer.preheader171 172 173.outer.preheader:174 %i = phi i64 [ 0, %0 ], [ %i.next, %.outer ]175 %1 = getelementptr inbounds %struct.s, ptr %obj, i64 0, i32 1, i64 %i176 br label %.inner177 178.exit:179 ret void180 181.outer:182 %i.next = add nuw nsw i64 %i, 1183 %exitcond.outer = icmp eq i64 %i.next, 32184 br i1 %exitcond.outer, label %.exit, label %.outer.preheader185 186.inner:187 %j = phi i64 [ 0, %.outer.preheader ], [ %j.next, %.inner ]188 %2 = getelementptr inbounds %struct.s, ptr %obj, i64 0, i32 0, i64 %j189 %3 = load i32, ptr %2190 %4 = load i32, ptr %1191 %5 = add nsw i32 %4, %3192 %6 = getelementptr inbounds %struct.s, ptr %obj, i64 0, i32 2, i64 %i, i64 %j193 %7 = load i32, ptr %6194 %8 = add nsw i32 %5, %7195 store i32 %8, ptr %6196 %j.next = add nuw nsw i64 %j, 1197 %exitcond.inner = icmp eq i64 %j.next, %z198 br i1 %exitcond.inner, label %.outer, label %.inner199}200