[akasakariko]

symbolize: fix legacy DWARF 5 reader for clang -gline-tables-only binaries

Fixes #255.

Problem

create_gcov aborts on any DWARF 5 binary built by clang (for example LLVM 19 with -gdwarf-5 -gline-tables-only -fdebug-info-for-profiling):

F... dwarf2reader.cc:754] Check failed: str_offsets_buffer_ != NULL

As noted in the issue, Google3Addr2line::Prepare() never loads .debug_str_offsets, so the first DW_FORM_strx* attribute fails the CHECK. GCC uses strp/line_strp forms, so it does not hit this.

Loading the section stops the crash, but the output is still wrong on the same binaries, because several DWARF 5 paths were never exercised before. This PR fixes them too:

• String tables in line programs: addr2line_inlinestack.cc and functioninfo.cc passed .debug_line_str to LineInfo as both .debug_str and .debug_str_offsets, and dereferenced end() when .debug_line_str was absent.
• Unit base attributes: clang emits DW_AT_str_offsets_base/DW_AT_addr_base after DW_AT_name (strx) and DW_AT_low_pc (addrx). Every unit after the first therefore used the first unit's tables. The result was a wrong unit base address and wrong DW_RLE_offset_pair inline ranges. Both bases are now read from the unit DIE first (CompilationUnit::ReadUnitBases), as LLVM does.
• DW_LNCT_MD5: clang emits it by default, and it made the v5 line header "malformed", so there was no line info. Content types that are not needed are now skipped (LineInfo::SkipForm).
• 0-based DWARF 5 file and directory indices: these were stored 1-based. clang's file 0 / DW_AT_call_file 0 were dropped (which then caused a NULL callsite_filename() dereference in GetInlineStack), and GCC DWARF 5 file names were off by one.
• Unit-level DW_AT_ranges with DW_FORM_rnglistx: C++, -ffunction-sections and ThinLTO builds hit a CHECK here.
• SkipAttribute: it did not handle DW_FORM_line_strp.
• Optional sections: strx attributes are now skipped instead of CHECK-failing when .debug_str/.debug_str_offsets is missing or the index is out of bounds.

I also checked the other DWARF 5 sections. .debug_addr, .debug_rnglists and .debug_line_str were already loaded. .debug_loclists is not used by the symbolizer, and DW_FORM_loclistx is already skipped.

Testing

New tests in addr2line_test.cc:

• Dwarf5ClangLineTablesOnlyBinary: new testdata/dwarf5_line_tables_only.bin (18 KB, 3 units). It was built by clang 19 from the committed testdata/dwarf5_line_tables_only_*.c sources; the command is in the test comment. The test checks full inline stacks at 5 addresses against values from llvm-symbolizer --inlining. Without this patch the test aborts with the Crashes with str_offsets_buffer_ = NULL #255 CHECK.
• Dwarf5StrxUsesLaterStrOffsetsBase, Dwarf5StrxWithoutStrOffsetsSectionIsSkipped, Dwarf5StrxOutOfRangeIndexIsSkipped: synthetic units. Without this patch they fail, abort and fail respectively.

Results:

• Builds: GCOV tool with clang 19.1.7 (Ninja, shared) and with GCC 15.2.0 (make, as in CI). In both, ctest passes 4/4 and addr2line_test passes 9/9. The tests also pass under ASan+UBSan.
• Inline stacks from the legacy symbolizer:
  • Match llvm-symbolizer exactly for every instruction of the new test binary, built with clang 19 and with clang 21.
  • Same as DWARF 4 for clang 19 DWARF 5 builds of Lua 5.4.7 (plain and -gdwarf64), the C++ googletest samples, and a GCC 15 DWARF 5 build of Lua.
• No regressions: output for DWARF 4 binaries is unchanged.
• Existing testdata/dwarf2_dwarf5.bin (GCC): wrong file names drop from 194,301 frames to 0.
• create_gcov run end to end (text profile) on a clang DWARF 5 binary: it used to abort and now writes a correct profile.

Not tested: the ENABLE_TOOL=LLVM configuration (llvm-project submodule not fetched locally) and real perf.data input.