Why do not demangle at link time?

Dmitry Antipov dantipov@nvidia.com
Fri Dec 1 11:24:00 GMT 2017

For a huge C++ binaries, it's quite typical to have a bottleneck in demangling while
loading symbols; I've seen >30% of time spent in d_xxx and friends, for example:

     12.19%  gdb      gdb                   [.] d_print_comp_inner
     10.11%  gdb      gdb                   [.] cp_find_first_component_aux
      6.21%  gdb      gdb                   [.] minimal_symbol_reader::install
      5.11%  gdb      gdb                   [.] htab_hash_string
      4.00%  gdb      gdb                   [.] d_print_comp
      3.56%  gdb      gdb                   [.] d_count_templates_scopes
      2.66%  gdb      libc-2.25.so          [.] __strcmp_sse2_unaligned
      2.58%  gdb      libc-2.25.so          [.] _int_malloc
      1.86%  gdb      gdb                   [.] cp_search_name_hash
      1.69%  gdb      libc-2.25.so          [.] strlen
      1.65%  gdb      gdb                   [.] d_name
      1.52%  gdb      gdb                   [.] cplus_demangle_type
      1.52%  gdb      gdb                   [.] compare_minimal_symbols
      1.51%  gdb      gdb                   [.] htab_find_slot_with_hash
      1.31%  gdb      gdb                   [.] d_make_comp

(observed on Chrome binary with .gdb_index embedded at link time). OK, .gdb_index
was designed to load symbols faster, and both gold and lld can emit it when linking;
so why do not (optionally) offload all demangling to linkers and extend .gdb_index
to add a kind of mapping from mangled to demangled symbol names?


More information about the Gdb mailing list