1; RUN: opt < %s -pgo-instr-gen -S | FileCheck %s --check-prefix=GEN
2; RUN: llvm-profdata merge %S/Inputs/loop2.proftext -o %t.profdata
3; RUN: opt < %s -pgo-instr-use -pgo-test-profile-file=%t.profdata -S | FileCheck %s --check-prefix=USE
4target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
5target triple = "x86_64-unknown-linux-gnu"
6
7; GEN: $__llvm_profile_raw_version = comdat any
8; GEN: @__llvm_profile_raw_version = constant i64 72057594037927939, comdat
9; GEN: @__profn_test_nested_for = private constant [15 x i8] c"test_nested_for"
10
11define i32 @test_nested_for(i32 %r, i32 %s) {
12entry:
13; GEN: entry:
14; GEN-NOT: call void @llvm.instrprof.increment
15  br label %for.cond.outer
16
17for.cond.outer:
18; GEN: for.cond.outer:
19; GEN-NOT: call void @llvm.instrprof.increment
20  %i.0 = phi i32 [ 0, %entry ], [ %inc.2, %for.inc.outer ]
21  %sum.0 = phi i32 [ 1, %entry ], [ %sum.1, %for.inc.outer ]
22  %cmp = icmp slt i32 %i.0, %r
23  br i1 %cmp, label %for.body.outer, label %for.end.outer
24; USE: br i1 %cmp, label %for.body.outer, label %for.end.outer
25; USE-SAME: !prof ![[BW_FOR_COND_OUTER:[0-9]+]]
26
27for.body.outer:
28; GEN: for.body.outer:
29; GEN-NOT: call void @llvm.instrprof.increment
30  br label %for.cond.inner
31
32for.cond.inner:
33; GEN: for.cond.inner:
34; GEN-NOT: call void @llvm.instrprof.increment
35  %j.0 = phi i32 [ 0, %for.body.outer ], [ %inc.1, %for.inc.inner ]
36  %sum.1 = phi i32 [ %sum.0, %for.body.outer ], [ %inc, %for.inc.inner ]
37  %cmp2 = icmp slt i32 %j.0, %s
38  br i1 %cmp2, label %for.body.inner, label %for.end.inner
39; USE: br i1 %cmp2, label %for.body.inner, label %for.end.inner
40; USE-SAME: !prof ![[BW_FOR_COND_INNER:[0-9]+]]
41
42for.body.inner:
43; GEN: for.body.inner:
44; GEN-NOT: call void @llvm.instrprof.increment
45  %inc = add nsw i32 %sum.1, 1
46  br label %for.inc.inner
47
48for.inc.inner:
49; GEN: for.inc.inner:
50; GEN: call void @llvm.instrprof.increment(i8* getelementptr inbounds ([15 x i8], [15 x i8]* @__profn_test_nested_for, i32 0, i32 0), i64 53929068288, i32 3, i32 0)
51  %inc.1 = add nsw i32 %j.0, 1
52  br label %for.cond.inner
53
54for.end.inner:
55; GEN: for.end.inner:
56  br label %for.inc.outer
57
58for.inc.outer:
59; GEN: for.inc.outer:
60; GEN: call void @llvm.instrprof.increment(i8* getelementptr inbounds ([15 x i8], [15 x i8]* @__profn_test_nested_for, i32 0, i32 0), i64 53929068288, i32 3, i32 1)
61  %inc.2 = add nsw i32 %i.0, 1
62  br label %for.cond.outer
63
64for.end.outer:
65; GEN: for.end.outer:
66; GEN: call void @llvm.instrprof.increment(i8* getelementptr inbounds ([15 x i8], [15 x i8]* @__profn_test_nested_for, i32 0, i32 0), i64 53929068288, i32 3, i32 2)
67  ret i32 %sum.0
68}
69
70; USE-DAG: ![[BW_FOR_COND_OUTER]] = !{!"branch_weights", i32 10, i32 6}
71; USE-DAG: ![[BW_FOR_COND_INNER]] = !{!"branch_weights", i32 33, i32 10}
72
73