llvm-symbolizer failed to get inline functions with split dwarf in dwarf 5

[ZequanWu]

a.cpp:

extern "C" void __attribute__((noreturn)) abort();

void bar();

__attribute__((always_inline)) void func() {
    static volatile int gv_func;
    ++gv_func;
}
__attribute__((always_inline)) void foo() {
    static volatile int gv_foo;
    ++gv_foo;
    if (gv_foo)
        abort();
    bar();
}

void bar() {
    static volatile int gv_bar;
    ++gv_bar;
    func();
}


int main() {
    foo();
    return 0;
}

commands:

$ clang++ -fuse-ld=lld a.cpp -O2 -g -gsplit-dwarf -fdebug-compilation-dir=. -o a.split
$ llvm-dwarfdump a.split
0x00000014: DW_TAG_skeleton_unit
              DW_AT_stmt_list   (0x00000000)
              DW_AT_str_offsets_base    (0x00000008)
              DW_AT_comp_dir    (".")
              DW_AT_GNU_pubnames        (true)
              DW_AT_dwo_name    ("a.split-a.dwo")
              DW_AT_low_pc      (0x0000000000001760)
              DW_AT_high_pc     (0x00000000000017d4)
              DW_AT_addr_base   (0x00000008)
$ llvm-dwarfdump a.split-a.dwo
...
0x0000003b:   DW_TAG_subprogram
                DW_AT_ranges    (indexed (0x0) rangelist = 0x00000030
                   [0x0000000000000010, 0x0000000000000032))
                DW_AT_frame_base        (DW_OP_reg7 RSP)
                DW_AT_call_all_calls    (true)
                DW_AT_abstract_origin   (0x00000093 "_Z3foov")

0x00000043:     DW_TAG_variable
                  DW_AT_name    ("gv_foo")
                  DW_AT_type    (0x00000032 "volatile int")
                  DW_AT_decl_file       (0x00)
                  DW_AT_decl_line       (10)
                  DW_AT_location        (DW_OP_addrx 0x1)

0x0000004e:     DW_TAG_inlined_subroutine
                  DW_AT_abstract_origin (0x00000085 "_Z3barv")
                  DW_AT_ranges  (indexed (0x1) rangelist = 0x00000034
                     [0x000000000000001f, 0x000000000000002b))
                  DW_AT_call_file       (0x00)
                  DW_AT_call_line       (14)
                  DW_AT_call_column     (0x05)
...
$ llvm-symbolizer --obj=a.split -a 0x177f
0x177f
foo()
./a.cpp:19:5
$ lldb a.split -o "image lookup -a 0x177f" -o "exit"
...
(lldb) image lookup -a 0x177f
      Address: a.split[0x000000000000177f] (a.split.PT_LOAD[1]..text + 271)
      Summary: a.split`foo() + 15 [inlined] bar() at a.cpp:19:5
               a.split`foo() + 15 at a.cpp:14:5
...

Expected behaviour is that it should show inlined function bar.

[llvmbot]

@llvm/issue-subscribers-tools-llvm-symbolizer

[dwblaikie]

Ah - probably doesn't come up much in practice, but this seems to happen if the CU has a genuine relocated base address (ie: if all the code is only in one section - which is rare in practice).

But the libDebugInfoDWARF code tries to find the base address, but is inspecting the split CU instead of the skeleton CU, so it doesn't find a base address and assumes the ranges are all relative to zero, instead of the actual base address (the low_pc on the skeleton CU).

Looking further.

[dwblaikie]

Fix is:

diff --git a/llvm/lib/DebugInfo/DWARF/DWARFUnit.cpp b/llvm/lib/DebugInfo/DWARF/DWARFUnit.cpp
index 2fc26fd86ba3..e2f17ca2e135 100644
--- a/llvm/lib/DebugInfo/DWARF/DWARFUnit.cpp
+++ b/llvm/lib/DebugInfo/DWARF/DWARFUnit.cpp
@@ -1049,7 +1049,7 @@ std::optional<object::SectionedAddress> DWARFUnit::getBaseAddress() {
   if (BaseAddr)
     return BaseAddr;
 
-  DWARFDie UnitDie = getUnitDIE();
+  DWARFDie UnitDie = (SU ? SU : this)->getUnitDIE();
   std::optional<DWARFFormValue> PC =
       UnitDie.find({DW_AT_low_pc, DW_AT_entry_pc});
   BaseAddr = toSectionedAddress(PC);

I'll work on a test case and get this committed soon.

[llvmbot]

@llvm/issue-subscribers-debuginfo