xref: /llvm-project/llvm/test/Transforms/LoopUnroll/guard-cost-for-unrolling.ll (revision d12af65d4626ac69faf8f56afd30971b8a2fc3b3) 
1; RUN: opt -S -disable-output -unroll-partial-threshold=16 -debug-only=loop-unroll -passes='loop-unroll<runtime>' < %s 2>&1 | FileCheck %s
2
3; REQUIRES: asserts
4
5; This test is needed to make sure that the guard cost remains the same,
6; independently on guard representation form (either intrinsic call or branch with
7; widenable condition).
8
9define void @test_guard_as_intrinsic(ptr %arr, i64 %n, i64 %bound) {
10; CHECK-LABEL: Loop Unroll: F[test_guard_as_intrinsic] Loop %loop
11; CHECK-NEXT:    Loop Size = 8
12; CHECK-NEXT:    runtime unrolling with count: 2
13entry:
14  br label %loop
15
16loop:
17  %iv = phi i64 [0, %entry], [%iv.next, %loop]
18  %gep = getelementptr i64, ptr %arr, i64 %iv
19  %bound_check = icmp ult i64 %iv, %bound
20  call void(i1, ...) @llvm.experimental.guard(i1 %bound_check) [ "deopt"() ]
21  store i64 %iv, ptr %gep, align 8
22  store i64 %iv, ptr %gep, align 8
23  %iv.next = add i64 %iv, 1
24  %loop_cond = icmp ult i64 %iv, %n
25  br i1 %loop_cond, label %loop, label %exit
26
27exit:
28  ret void
29}
30
31define void @test_guard_as_branch(ptr %arr, i64 %n, i64 %bound) {
32; CHECK-LABEL: Loop Unroll: F[test_guard_as_branch] Loop %loop
33; CHECK-NEXT:    Loop Size = 8
34; CHECK-NEXT:    runtime unrolling with count: 2
35entry:
36  br label %loop
37
38loop:                                             ; preds = %guarded, %entry
39  %iv = phi i64 [ 0, %entry ], [ %iv.next, %guarded ]
40  %gep = getelementptr i64, ptr %arr, i64 %iv
41  %bound_check = icmp ult i64 %iv, %bound
42  %widenable_cond = call i1 @llvm.experimental.widenable.condition()
43  %exiplicit_guard_cond = and i1 %bound_check, %widenable_cond
44  br i1 %exiplicit_guard_cond, label %guarded, label %deopt, !prof !0
45
46deopt:                                            ; preds = %loop
47  call void (...) @llvm.experimental.deoptimize.isVoid() [ "deopt"() ]
48  ret void
49
50guarded:                                          ; preds = %loop
51  store i64 %iv, ptr %gep, align 8
52  store i64 %iv, ptr %gep, align 8
53  %iv.next = add i64 %iv, 1
54  %loop_cond = icmp ult i64 %iv, %n
55  br i1 %loop_cond, label %loop, label %exit
56
57exit:                                             ; preds = %guarded
58  ret void
59}
60
61; Function Attrs: nocallback nofree nosync willreturn
62declare void @llvm.experimental.guard(i1, ...) #0
63
64declare void @llvm.experimental.deoptimize.isVoid(...)
65
66; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(inaccessiblemem: readwrite)
67declare i1 @llvm.experimental.widenable.condition() #1
68
69!0 = !{!"branch_weights", i32 1048576, i32 1}
70