82 lines · plain
1; RUN: opt < %s -passes=loop-interchange -cache-line-size=1 -pass-remarks-output=%t -disable-output \2; RUN: -verify-dom-info -verify-loop-info3; RUN: FileCheck -input-file %t %s4 5 6; Test that loop-interchange doesn't undo its own transoformation. This is7; the case when the cost computed by CacheCost is the same for the loop of `j`8; and `k`.9;10; #define N 411; int a[N*N][N*N][N*N];12; void f() {13; for (int i = 0; i < N; i++)14; for (int j = 1; j < 2*N; j++)15; for (int k = 1; k < 2*N; k++)16; a[i][k+1][j-1] -= a[i+N-1][k][j];17; }18 19; CHECK: --- !Passed20; CHECK-NEXT: Pass: loop-interchange21; CHECK-NEXT: Name: Interchanged22; CHECK-NEXT: Function: f23; CHECK-NEXT: Args:24; CHECK-NEXT: - String: Loop interchanged with enclosing loop.25; CHECK-NEXT: ...26; CHECK-NEXT: --- !Missed27; CHECK-NEXT: Pass: loop-interchange28; CHECK-NEXT: Name: Dependence29; CHECK-NEXT: Function: f30; CHECK-NEXT: Args:31; CHECK-NEXT: - String: Cannot interchange loops due to dependences.32; CHECK-NEXT: ...33; CHECK-NEXT: --- !Missed34; CHECK-NEXT: Pass: loop-interchange35; CHECK-NEXT: Name: InterchangeNotProfitable36; CHECK-NEXT: Function: f37; CHECK-NEXT: Args:38; CHECK-NEXT: - String: Interchanging loops is not considered to improve cache locality nor vectorization.39; CHECK-NEXT: ...40 41@a = dso_local local_unnamed_addr global [16 x [16 x [16 x i32]]] zeroinitializer, align 442 43define dso_local void @f() {44entry:45 br label %for.cond1.preheader46 47for.cond1.preheader:48 %indvars.iv46 = phi i64 [ 0, %entry ], [ %indvars.iv.next47, %for.cond.cleanup3 ]49 %0 = add nuw nsw i64 %indvars.iv46, 350 br label %for.cond5.preheader51 52for.cond5.preheader:53 %indvars.iv41 = phi i64 [ 1, %for.cond1.preheader ], [ %indvars.iv.next42, %for.cond.cleanup7 ]54 %1 = add nsw i64 %indvars.iv41, -155 br label %for.body856 57for.cond.cleanup3:58 %indvars.iv.next47 = add nuw nsw i64 %indvars.iv46, 159 %exitcond50 = icmp ne i64 %indvars.iv.next47, 460 br i1 %exitcond50, label %for.cond1.preheader, label %for.cond.cleanup61 62for.cond.cleanup7:63 %indvars.iv.next42 = add nuw nsw i64 %indvars.iv41, 164 %exitcond45 = icmp ne i64 %indvars.iv.next42, 865 br i1 %exitcond45, label %for.cond5.preheader, label %for.cond.cleanup366 67for.body8:68 %indvars.iv = phi i64 [ 1, %for.cond5.preheader ], [ %indvars.iv.next, %for.body8 ]69 %arrayidx12 = getelementptr inbounds nuw [16 x [16 x [16 x i32]]], ptr @a, i64 0, i64 %0, i64 %indvars.iv, i64 %indvars.iv4170 %2 = load i32, ptr %arrayidx12, align 471 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 172 %arrayidx20 = getelementptr inbounds [16 x [16 x [16 x i32]]], ptr @a, i64 0, i64 %indvars.iv46, i64 %indvars.iv.next, i64 %173 %3 = load i32, ptr %arrayidx20, align 474 %sub21 = sub nsw i32 %3, %275 store i32 %sub21, ptr %arrayidx20, align 476 %exitcond = icmp ne i64 %indvars.iv.next, 877 br i1 %exitcond, label %for.body8, label %for.cond.cleanup778 79for.cond.cleanup:80 ret void81}82