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
4 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
5 target triple = "x86_64-unknown-linux-gnu"
7 ; GEN: @__profn_test_nested_for = private constant [15 x i8] c"test_nested_for"
9 define i32 @test_nested_for(i32 %r, i32 %s) {
12 ; GEN-NOT: call void @llvm.instrprof.increment
13 br label %for.cond.outer
16 ; GEN: for.cond.outer:
17 ; GEN-NOT: call void @llvm.instrprof.increment
18 %i.0 = phi i32 [ 0, %entry ], [ %inc.2, %for.inc.outer ]
19 %sum.0 = phi i32 [ 1, %entry ], [ %sum.1, %for.inc.outer ]
20 %cmp = icmp slt i32 %i.0, %r
21 br i1 %cmp, label %for.body.outer, label %for.end.outer
22 ; USE: br i1 %cmp, label %for.body.outer, label %for.end.outer
23 ; USE-SAME: !prof ![[BW_FOR_COND_OUTER:[0-9]+]]
26 ; GEN: for.body.outer:
27 ; GEN-NOT: call void @llvm.instrprof.increment
28 br label %for.cond.inner
31 ; GEN: for.cond.inner:
32 ; GEN-NOT: call void @llvm.instrprof.increment
33 %j.0 = phi i32 [ 0, %for.body.outer ], [ %inc.1, %for.inc.inner ]
34 %sum.1 = phi i32 [ %sum.0, %for.body.outer ], [ %inc, %for.inc.inner ]
35 %cmp2 = icmp slt i32 %j.0, %s
36 br i1 %cmp2, label %for.body.inner, label %for.end.inner
37 ; USE: br i1 %cmp2, label %for.body.inner, label %for.end.inner
38 ; USE-SAME: !prof ![[BW_FOR_COND_INNER:[0-9]+]]
41 ; GEN: for.body.inner:
42 ; GEN-NOT: call void @llvm.instrprof.increment
43 %inc = add nsw i32 %sum.1, 1
44 br label %for.inc.inner
48 ; 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)
49 %inc.1 = add nsw i32 %j.0, 1
50 br label %for.cond.inner
54 br label %for.inc.outer
58 ; 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)
59 %inc.2 = add nsw i32 %i.0, 1
60 br label %for.cond.outer
64 ; 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)
68 ; USE-DAG: ![[BW_FOR_COND_OUTER]] = !{!"branch_weights", i32 10, i32 6}
69 ; USE-DAG: ![[BW_FOR_COND_INNER]] = !{!"branch_weights", i32 33, i32 10}