Test lldb backtraces with machine function splitter

[tmsriram]

Summary

clang supports option -fsplit-machine-functions and this test checks if the backtraces are sane when functions are split

With -fsplit-machine-functions, a function with profiles can get split into 2 parts, the original function containing hot code and a cold part as determined by the profile info and the cold cutoff threshold.. The cold part gets the ".cold" suffix to disambiguate its symbol from the hot part and can be placed arbitrarily in the address space.

This test checks if the back-trace looks correct when the cold part is executed.

[snehasish]

lldb/test/Shell/Unwind/split-machine-functions.test

2

typo "produces"

17

Do we need a symbol ordering file here? Without it the split part should go to .text.split and the other functions would go to .text effectively creating the gap which you want here.

[tmsriram]

lldb/test/Shell/Unwind/split-machine-functions.test

17

In the final binary, in the default case, .text (after merging the individual .text sections) and .text.split are put next to each other. If foo is the last function in .text, foo will be placed next to foo.cold. In this case, I do see main between foo and foo.cold so there is a gap in the default case but the function ordering from the compiler is arbitrary and is not necessarily the same as the source order. Since MFS does not rely on any symbol ordering, I can just delete it but just want to mention that the gap is not guaranteed just because they are in separate sections.

[snehasish]

lldb/test/Shell/Unwind/Inputs/split-machine-functions.ll

3

Is it useful to include the source here? The generated IR is fairly straightforward and could just be edited by hand if necessary.

64

Instead of having a (auto generated) comment with the attrs would making these actual attributes make the test more robust?
(here and above)

66–67

These two lines can be removed I think? I guess they are here since you used -O0 to generate the bitcode.

lldb/test/Shell/Unwind/split-machine-functions.test

17

SGTM, can you add a comment here saying that a symbol ordering file is used to ensure that there is at least one symbol between foo and foo.cold since ordering is arbitrary?

Also would it make things cleaner if you used split-file (https://reviews.llvm.org/rGbcea3a7a288) for the symbol ordering parts instead of "RUN: echo ..." for the ordering file?

[tmsriram]

Address the following reviewer comments:

• Use split-file for sym order
• Cleanup the .ll file further
• Typo fixes.

lldb/test/Shell/Unwind/Inputs/split-machine-functions.ll

3

I was told in prior tests to include a reproducible if it exists. I agree this is not complicated .ll but it doesn't hurt? I am fine either way.

64

I intentionally removed the metadata for the attrs as the principle has been to to keep the tests trim. Removing these attrs too. Alright?

lldb/test/Shell/Unwind/split-machine-functions.test

17

Nice suggestion. I could use split-file for the ir too but since we already have an Inputs directory keeping it separate makes the file look less cluttered. Alright?

[snehasish]

lgtm, thanks for adding the test.

lldb/test/Shell/Unwind/Inputs/split-machine-functions.ll

3

Sure, lets keep if thats the convention here.

lldb/test/Shell/Unwind/Inputs/split-machine-functions.ll

+;; IR + Profile to split machine functions with -fsplit-machine-functions
+;; Source used to generate this file
+;; int k;
+
+;; int baz() {
+;;   return 0;
+;; }
+;;
+;; int bar() {
+;;   return 0;
+;; }
+;;
+;; int foo() {
+;;   if (k)
+;;     baz();
+;;   else
+;;     bar();
+;;   return 0;
+;; }
+;;
+;; int main() {
+;;   foo();
+;;   return 0;
+;; }
+
+;; Commands:
+;; clang -fprofile-generate -O0 source.cc
+;; ./a.out && llvm-profdata merge -o default.profdata *.profraw
+;; clang -fprofile-use -O0 source.cc -emit-llvm -S
+
+target triple = "x86_64-unknown-linux-gnu"
+
+@k = dso_local global i32 0, align 4
+
+define dso_local i32 @_Z3bazv() !prof !31 {
+  ret i32 0
+}
+
+define dso_local i32 @_Z3barv() !prof !32 {
+  ret i32 0
+}
+
+define dso_local i32 @_Z3foov() !prof !32 {
+  %1 = load i32, i32* @k, align 4
+  %2 = icmp ne i32 %1, 0
+  br i1 %2, label %3, label %5, !prof !33
+
+3:                                                ; preds = %0
+  %4 = call i32 @_Z3bazv()
+  br label %7
+
+5:                                                ; preds = %0
+  %6 = call i32 @_Z3barv()
+  br label %7
+
+7:                                                ; preds = %5, %3
+  ret i32 0
+}
+
+define dso_local i32 @main() !prof !32 {
+  %1 = call i32 @_Z3foov()
+  ret i32 0
+}
+
+!llvm.module.flags = !{!0, !1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 1, !"ProfileSummary", !2}
+!2 = !{!3, !4, !5, !6, !7, !8, !9, !10, !11, !12}
+!3 = !{!"ProfileFormat", !"InstrProf"}
+!4 = !{!"TotalCount", i64 3}
+!5 = !{!"MaxCount", i64 1}
+!6 = !{!"MaxInternalCount", i64 1}
+!7 = !{!"MaxFunctionCount", i64 1}
+!8 = !{!"NumCounts", i64 5}
+!9 = !{!"NumFunctions", i64 4}
+!10 = !{!"IsPartialProfile", i64 0}
+!11 = !{!"PartialProfileRatio", double 0.000000e+00}
+!12 = !{!"DetailedSummary", !13}
+!13 = !{!14, !15, !16, !17, !18, !19, !20, !21, !22, !23, !24, !25, !26, !27, !28, !29}
+!14 = !{i32 10000, i64 0, i32 0}
+!15 = !{i32 100000, i64 0, i32 0}
+!16 = !{i32 200000, i64 0, i32 0}
+!17 = !{i32 300000, i64 0, i32 0}
+!18 = !{i32 400000, i64 1, i32 3}
+!19 = !{i32 500000, i64 1, i32 3}
+!20 = !{i32 600000, i64 1, i32 3}
+!21 = !{i32 700000, i64 1, i32 3}
+!22 = !{i32 800000, i64 1, i32 3}
+!23 = !{i32 900000, i64 1, i32 3}
+!24 = !{i32 950000, i64 1, i32 3}
+!25 = !{i32 990000, i64 1, i32 3}
+!26 = !{i32 999000, i64 1, i32 3}
+!27 = !{i32 999900, i64 1, i32 3}
+!28 = !{i32 999990, i64 1, i32 3}
+!29 = !{i32 999999, i64 1, i32 3}
+!31 = !{!"function_entry_count", i64 0}
+!32 = !{!"function_entry_count", i64 1}
+!33 = !{!"branch_weights", i32 1, i32 0}

lldb/test/Shell/Unwind/split-machine-functions.test

+# Test to check if machine function splitter still produces the right backtraces
+# with lldb when a function is split into a hot and cold part and the cold part
+# is executed.  The cold part has the same function symbol name but with a
+# ".cold" suffix and this test checks that the back trace is clear.
+
+# UNSUPPORTED: system-darwin, system-windows
+# REQUIRES: target-x86_64
+# REQUIRES: lld
+
+# RUN: split-file %s %t.split
+#
+# RUN: %clang_host %p/Inputs/split-machine-functions.ll -o %t
+# RUN: %lldb %t -s %t.split/commands -o exit | FileCheck  %s --check-prefix=DEFAULT
+#
+# RUN: %clang_host %p/Inputs/split-machine-functions.ll -fsplit-machine-functions -o %t
+# RUN: %lldb %t -s %t.split/commands -o exit | FileCheck  %s --check-prefix=SPLIT
+#
+# Test a permutation where foo.cold is very far from foo.  The default ordering does not
+# ensure that there will be a gap between foo and foo.cold.  Using a symbol ordering
+# file guarantees this
+# RUN: %clang_host %p/Inputs/split-machine-functions.ll -fsplit-machine-functions -o %t -fuse-ld=lld -Wl,--symbol-ordering-file,%t.split/sym_order_1.txt -Wl,--warn-symbol-ordering -Wl,--fatal-warnings
+# RUN: %lldb %t -s %t.split/commands -o exit | FileCheck  %s --check-prefix=SPLIT
+
+#--- commands
+breakpoint set -n bar
+# DEFAULT: Breakpoint 1: where = {{.*}}`bar
+# SPLIT: Breakpoint 1: where = {{.*}}`bar
+
+process launch
+# DEFAULT: stop reason = breakpoint 1.1
+# SPLIT: stop reason = breakpoint 1.1
+
+thread backtrace
+# DEFAULT: frame #0: {{.*}}`bar
+# DEFAULT: frame #1: {{.*}}`foo() +
+# DEFAULT: frame #2: {{.*}}`main +
+# SPLIT: frame #0: {{.*}}`bar
+# SPLIT: frame #1: {{.*}}`foo() (.cold) +
+# SPLIT: frame #2: {{.*}}`main +
+
+#--- sym_order_1.txt
+_Z3foov
+main
+_Z3barv
+_Z3bazv
+_Z3foov.cold