|
| 1 | +; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4 |
| 2 | +; RUN: opt -S -passes=loop-idiom < %s | FileCheck %s |
| 3 | + |
| 4 | +; FIXME: The poison flags should be preserved, as no transform takes place. |
| 5 | +define void @test(ptr %p.end, ptr %p.start) { |
| 6 | +; CHECK-LABEL: define void @test( |
| 7 | +; CHECK-SAME: ptr [[P_END:%.*]], ptr [[P_START:%.*]]) { |
| 8 | +; CHECK-NEXT: entry: |
| 9 | +; CHECK-NEXT: [[P_END_INT:%.*]] = ptrtoint ptr [[P_END]] to i64 |
| 10 | +; CHECK-NEXT: [[P_START_INT:%.*]] = ptrtoint ptr [[P_START]] to i64 |
| 11 | +; CHECK-NEXT: [[DIST:%.*]] = sub i64 [[P_END_INT]], [[P_START_INT]] |
| 12 | +; CHECK-NEXT: [[LEN:%.*]] = lshr i64 [[DIST]], 5 |
| 13 | +; CHECK-NEXT: [[CMP:%.*]] = icmp eq ptr [[P_END]], [[P_START]] |
| 14 | +; CHECK-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[PREHEADER:%.*]] |
| 15 | +; CHECK: preheader: |
| 16 | +; CHECK-NEXT: br label [[LOOP:%.*]] |
| 17 | +; CHECK: loop: |
| 18 | +; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[LEN]], [[PREHEADER]] ], [ [[IV_DEC:%.*]], [[LOOP]] ] |
| 19 | +; CHECK-NEXT: [[IV_NEG:%.*]] = sub nsw i64 0, [[IV]] |
| 20 | +; CHECK-NEXT: [[SRC:%.*]] = getelementptr inbounds [32 x i8], ptr [[P_END]], i64 [[IV_NEG]] |
| 21 | +; CHECK-NEXT: [[V:%.*]] = load <4 x i64>, ptr [[SRC]], align 8 |
| 22 | +; CHECK-NEXT: [[DST:%.*]] = getelementptr inbounds [32 x i8], ptr [[P_START]], i64 [[IV_NEG]] |
| 23 | +; CHECK-NEXT: store <4 x i64> [[V]], ptr [[DST]], align 8 |
| 24 | +; CHECK-NEXT: [[IV_DEC]] = add i64 [[IV]], -1 |
| 25 | +; CHECK-NEXT: [[CMP2:%.*]] = icmp eq i64 [[IV_DEC]], 0 |
| 26 | +; CHECK-NEXT: br i1 [[CMP2]], label [[LOOP_EXIT:%.*]], label [[LOOP]] |
| 27 | +; CHECK: loop.exit: |
| 28 | +; CHECK-NEXT: br label [[EXIT]] |
| 29 | +; CHECK: exit: |
| 30 | +; CHECK-NEXT: ret void |
| 31 | +; |
| 32 | +entry: |
| 33 | + %p.end.int = ptrtoint ptr %p.end to i64 |
| 34 | + %p.start.int = ptrtoint ptr %p.start to i64 |
| 35 | + %dist = sub nuw i64 %p.end.int, %p.start.int |
| 36 | + %len = lshr exact i64 %dist, 5 |
| 37 | + %cmp = icmp eq ptr %p.end, %p.start |
| 38 | + br i1 %cmp, label %exit, label %preheader |
| 39 | + |
| 40 | +preheader: |
| 41 | + br label %loop |
| 42 | + |
| 43 | +loop: |
| 44 | + %iv = phi i64 [ %len, %preheader ], [ %iv.dec, %loop ] |
| 45 | + %iv.neg = sub nsw i64 0, %iv |
| 46 | + %src = getelementptr inbounds [32 x i8], ptr %p.end, i64 %iv.neg |
| 47 | + %v = load <4 x i64>, ptr %src, align 8 |
| 48 | + %dst = getelementptr inbounds [32 x i8], ptr %p.start, i64 %iv.neg |
| 49 | + store <4 x i64> %v, ptr %dst, align 8 |
| 50 | + %iv.dec = add i64 %iv, -1 |
| 51 | + %cmp2 = icmp eq i64 %iv.dec, 0 |
| 52 | + br i1 %cmp2, label %loop.exit, label %loop |
| 53 | + |
| 54 | +loop.exit: |
| 55 | + br label %exit |
| 56 | + |
| 57 | +exit: |
| 58 | + ret void |
| 59 | +} |
0 commit comments