xref: /llvm-project/llvm/test/Transforms/LoopVectorize/unroll_nonlatch.ll (revision eecb99c5f66c8491766628a2925587e20f3b1dbd)
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt < %s -S -passes=loop-vectorize -force-vector-interleave=2 | FileCheck %s
3
4; Demonstrate a case where we unroll a loop, but don't vectorize it.
5; The original loop runs stores in the latch block on iterations 0 to 1022,
6; and exits when %indvars.iv = 1023. (That is, it actually runs the stores
7; for an odd number of iterations.)  If we unroll by two in the "vector.body"
8; loop, we must exit to the epilogue on iteration with %indvars.iv = 1022 to
9; avoid an out of bounds access.
10
11define void @test(ptr %data) {
12; CHECK-LABEL: @test(
13; CHECK-NEXT:  entry:
14; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
15; CHECK:       vector.ph:
16; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
17; CHECK:       vector.body:
18; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
19; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
20; CHECK-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 1
21; CHECK-NEXT:    [[TMP2:%.*]] = shl nuw nsw i64 [[TMP0]], 1
22; CHECK-NEXT:    [[TMP3:%.*]] = shl nuw nsw i64 [[TMP1]], 1
23; CHECK-NEXT:    [[TMP4:%.*]] = or disjoint i64 [[TMP2]], 1
24; CHECK-NEXT:    [[TMP5:%.*]] = or disjoint i64 [[TMP3]], 1
25; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds double, ptr [[DATA:%.*]], i64 [[TMP4]]
26; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds double, ptr [[DATA]], i64 [[TMP5]]
27; CHECK-NEXT:    [[TMP8:%.*]] = load double, ptr [[TMP6]], align 8
28; CHECK-NEXT:    [[TMP9:%.*]] = load double, ptr [[TMP7]], align 8
29; CHECK-NEXT:    [[TMP10:%.*]] = fneg double [[TMP8]]
30; CHECK-NEXT:    [[TMP11:%.*]] = fneg double [[TMP9]]
31; CHECK-NEXT:    store double [[TMP10]], ptr [[TMP6]], align 8
32; CHECK-NEXT:    store double [[TMP11]], ptr [[TMP7]], align 8
33; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
34; CHECK-NEXT:    [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1022
35; CHECK-NEXT:    br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
36; CHECK:       middle.block:
37; CHECK-NEXT:    br label [[SCALAR_PH]]
38; CHECK:       scalar.ph:
39; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 1022, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
40; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
41; CHECK:       for.body:
42; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_LATCH:%.*]] ]
43; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
44; CHECK-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 1024
45; CHECK-NEXT:    br i1 [[EXITCOND_NOT]], label [[FOR_END:%.*]], label [[FOR_LATCH]]
46; CHECK:       for.latch:
47; CHECK-NEXT:    [[T15:%.*]] = shl nuw nsw i64 [[INDVARS_IV]], 1
48; CHECK-NEXT:    [[T16:%.*]] = or disjoint i64 [[T15]], 1
49; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds double, ptr [[DATA]], i64 [[T16]]
50; CHECK-NEXT:    [[T17:%.*]] = load double, ptr [[ARRAYIDX]], align 8
51; CHECK-NEXT:    [[FNEG:%.*]] = fneg double [[T17]]
52; CHECK-NEXT:    store double [[FNEG]], ptr [[ARRAYIDX]], align 8
53; CHECK-NEXT:    br label [[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
54; CHECK:       for.end:
55; CHECK-NEXT:    ret void
56;
57entry:
58  br label %for.body
59
60for.body:
61  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.latch ]
62  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
63  %exitcond.not = icmp eq i64 %indvars.iv.next, 1024
64  br i1 %exitcond.not, label %for.end, label %for.latch
65
66for.latch:
67  %t15 = shl nuw nsw i64 %indvars.iv, 1
68  %t16 = or disjoint i64 %t15, 1
69  %arrayidx = getelementptr inbounds double, ptr %data, i64 %t16
70  %t17 = load double, ptr %arrayidx, align 8
71  %fneg = fneg double %t17
72  store double %fneg, ptr %arrayidx, align 8
73  br label %for.body
74
75for.end:
76  ret void
77}
78