1 ; RUN: opt < %s -pgo-instr-gen -S | FileCheck %s --check-prefix=GEN 2 ; RUN: opt < %s -passes=pgo-instr-gen -S | FileCheck %s --check-prefix=GEN 3 ; RUN: llvm-profdata merge %S/Inputs/loop2.proftext -o %t.profdata 4 ; RUN: opt < %s -pgo-instr-use -pgo-test-profile-file=%t.profdata -S | FileCheck %s --check-prefix=USE 5 ; RUN: opt < %s -passes=pgo-instr-use -pgo-test-profile-file=%t.profdata -S | FileCheck %s --check-prefix=USE 6 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" 7 target triple = "x86_64-unknown-linux-gnu" 8 9 ; GEN: $__llvm_profile_raw_version = comdat any 10 ; GEN: @__llvm_profile_raw_version = constant i64 {{[0-9]+}}, comdat 11 ; GEN: @__profn_test_nested_for = private constant [15 x i8] c"test_nested_for" 12 13 define i32 @test_nested_for(i32 %r, i32 %s) { 14 entry: 15 ; GEN: entry: 16 ; GEN-NOT: call void @llvm.instrprof.increment 17 br label %for.cond.outer 18 19 for.cond.outer: 20 ; GEN: for.cond.outer: 21 ; GEN-NOT: call void @llvm.instrprof.increment 22 %i.0 = phi i32 [ 0, %entry ], [ %inc.2, %for.inc.outer ] 23 %sum.0 = phi i32 [ 1, %entry ], [ %sum.1, %for.inc.outer ] 24 %cmp = icmp slt i32 %i.0, %r 25 br i1 %cmp, label %for.body.outer, label %for.end.outer 26 ; USE: br i1 %cmp, label %for.body.outer, label %for.end.outer 27 ; USE-SAME: !prof ![[BW_FOR_COND_OUTER:[0-9]+]] 28 29 for.body.outer: 30 ; GEN: for.body.outer: 31 ; GEN-NOT: call void @llvm.instrprof.increment 32 br label %for.cond.inner 33 34 for.cond.inner: 35 ; GEN: for.cond.inner: 36 ; GEN-NOT: call void @llvm.instrprof.increment 37 %j.0 = phi i32 [ 0, %for.body.outer ], [ %inc.1, %for.inc.inner ] 38 %sum.1 = phi i32 [ %sum.0, %for.body.outer ], [ %inc, %for.inc.inner ] 39 %cmp2 = icmp slt i32 %j.0, %s 40 br i1 %cmp2, label %for.body.inner, label %for.end.inner 41 ; USE: br i1 %cmp2, label %for.body.inner, label %for.end.inner 42 ; USE-SAME: !prof ![[BW_FOR_COND_INNER:[0-9]+]] 43 44 for.body.inner: 45 ; GEN: for.body.inner: 46 ; GEN-NOT: call void @llvm.instrprof.increment 47 %inc = add nsw i32 %sum.1, 1 48 br label %for.inc.inner 49 50 for.inc.inner: 51 ; GEN: for.inc.inner: 52 ; 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) 53 %inc.1 = add nsw i32 %j.0, 1 54 br label %for.cond.inner 55 56 for.end.inner: 57 ; GEN: for.end.inner: 58 br label %for.inc.outer 59 60 for.inc.outer: 61 ; GEN: for.inc.outer: 62 ; 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) 63 %inc.2 = add nsw i32 %i.0, 1 64 br label %for.cond.outer 65 66 for.end.outer: 67 ; GEN: for.end.outer: 68 ; 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) 69 ret i32 %sum.0 70 } 71 72 ; USE-DAG: ![[BW_FOR_COND_OUTER]] = !{!"branch_weights", i32 10, i32 6} 73 ; USE-DAG: ![[BW_FOR_COND_INNER]] = !{!"branch_weights", i32 33, i32 10} 74 75
