[PATCH v2 00/42] Share DWARF partial symtabs between objfiles
Simon Marchi
simark@simark.ca
Wed May 13 14:52:41 GMT 2020
On 2020-05-12 5:08 p.m., Simon Marchi via Gdb-patches wrote:
> This is version 2 of this patchset originally written by Tom:
>
> https://sourceware.org/legacy-ml/gdb-patches/2020-02/msg00594.html
>
> The current patchset therefore contains a mix of Tom's patches that I modified,
> plus a few new patches. I won't repeat everything that is said in the cover
> letter of that patchset, please read that to understand the motivation behind
> this change..
>
> The big difference between v1 and v2 is how the split is done. In v1, the
> object structure looks like this:
>
> objfile -> dwarf2_per_objfile -> dwarf2_unshareable
>
> where `dwarf2_per_objfile` is actually shared between objfiles (there is one
> copy per BFD), and `dwarf2_unshareable` is not (there is one copy per objfile).
> The dwarf2_per_objfile -> dwarf2_unshareable link is updated at all entry
> points of the DWARF code based on what is the current objfile.
>
> The current patchset (v2) proposes this structure instead, which in my opinion
> is more natural:
>
> objfile -> dwarf2_per_objfile -> dwarf2_per_bfd
>
> Where `dwarf2_per_objfile` is really per-objfile (there is one copy per
> objfile) and `dwarf2_per_bfd` is per-bfd (there is one copy per-bfd). There is
> no link to update, whatever is shared between objfiles is put in
> `dwarf2_per_bfd` and each `dwarf2_per_objfile` always points to one
> `dwarf2_per_bfd`.
>
> Since there were performance numbers with the original patchset, I made my own
> here to see if the speedups were still there. Like in the original patchset, I
> set "maintenance time 1" and did "add-inferior -exec ./gdb" after "./gdb gdb".
> I disabled any auto-loading to avoid loading gdb-gdb.{gdb,py}. GDB is built
> with -O2 and gcc 9, and it is running on an AMD EPYC 7351P.
>
> Before: Command execution time: 1.213480 (cpu), 1.214268 (wall)
> After: Command execution time: 0.001697 (cpu), 0.001694 (wall)
>
> These are the times of one run, but all runs I did are very close. The "after"
> time is surpringly low compared to the original numbers, I hope it doesn't mean
> I messed up something and we are skipping something important...
>
> At some point I tested with all these boards and did not see any regression
> (they helped very much to find some issues in the process though):
>
> unix cc-with-debug-names cc-with-dwz cc-with-dwz-m cc-with-gdb-index dwarf4-gdb-index fission-dwp fission readnow
>
> I have since rebased the patchset and resolved conflicts, it's therefore
> possible I messed something up.
>
> Simon Marchi (35):
> Split dwarf2_per_objfile into dwarf2_per_objfile and dwarf2_per_bfd
> Remove symtab links from dwarf2_psymtab and dwarf2_per_cu_quick_data
> Move die_type_hash to dwarf2_per_objfile
> Add dwarf2_per_objfile field to dwarf2_cu
> Remove reference to dwarf2_per_cu_data::dwarf2_per_objfile in
> dw2_do_instantiate_symtab
> Remove dwarf2_cu->per_cu->dwarf2_per_objfile references
> Add dwarf2_per_bfd field to dwarf2_per_cu_data
> Make dwarf2_get_dwz_file take a dwarf2_per_bfd
> Use bfd_get_filename instead of objfile_name in lookup_dwo_unit
> Add dwarf2_per_objfile parameter to cutu_reader's constructors
> Make queue_and_load_dwo_tu receive a dwarf2_cu
> Remove dwarf2_per_cu_data::dwarf2_per_objfile reference in
> cutu_reader::keep
> Add dwarf2_per_objfile parameter to create_partial_symtab
> Add dwarf2_per_objfile parameter to recursively_compute_inclusions
> Add dwarf2_per_objfile parameter to process_full_{comp,type}_unit
> Pass dwarf2_cu objects to dwo-related functions, instead of
> dwarf2_per_cu_data
> Remove reference to dwarf2_per_cu_data::dwarf2_per_objfile in
> queue_and_load_all_dwo_tus
> Move int type methods out of dwarf2_per_cu_data
> Add dwarf2_per_objfile to dwarf_expr_context and dwarf2_frame_cache
> Remove dwarf2_per_cu_data::text_offset
> Add dwarf2_per_objfile parameter to dwarf2_read_addr_index
> Add dwarf2_per_objfile parameter to allocate_piece_closure
> Add dwarf2_per_objfile parameters to dwarf2_fetch_* functions
> Remove dwarf2_per_cu_data::objfile ()
> Add dwarf2_per_objfile parameter to free_one_cached_comp_unit
> Add dwarf2_per_objfile parameter to get_die_type_at_offset
> Remove leftover references to dwarf2_per_cu_data::dwarf2_per_objfile
> Remove dwarf2_per_cu_data::dwarf2_per_objfile
> Pass dwarf2_per_bfd instead of dwarf2_per_objfile to some
> index-related functions
> Pass dwarf2_cu to process_full_{comp,type}_unit
> Make load_cu return the loaded dwarf2_cu
> Add comp_unit_head to dwarf2_per_cu_data
> Pass existing_cu object to cutu_reader
> Replace dwarf2_per_cu_data::cu backlink with per-objfile map
> Make mapped_debug_names independent of objfile
>
> Tom Tromey (7):
> Introduce dwarf2_per_objfile::obstack
> Add "objfile" parameter to two partial_symtab methods
> Add dwarf2_per_cu_data::index
> Add dwarf2_per_objfile member to DWARF batons
> Split type_unit_group
> Move signatured_type::type to unshareable object
> Share DWARF partial symtabs
>
> gdb/compile/compile-loc2c.c | 20 +-
> gdb/compile/compile.h | 13 +-
> gdb/dwarf2/expr.c | 11 +-
> gdb/dwarf2/expr.h | 11 +-
> gdb/dwarf2/frame.c | 67 +-
> gdb/dwarf2/index-cache.c | 2 +-
> gdb/dwarf2/index-write.c | 51 +-
> gdb/dwarf2/loc.c | 289 ++--
> gdb/dwarf2/loc.h | 11 +-
> gdb/dwarf2/macro.c | 11 +-
> gdb/dwarf2/read.c | 2490 ++++++++++++++++++-----------------
> gdb/dwarf2/read.h | 298 +++--
> gdb/gdbtypes.h | 8 +-
> gdb/objfiles.h | 2 +-
> gdb/psympriv.h | 19 +-
> gdb/psymtab.c | 44 +-
> 16 files changed, 1848 insertions(+), 1499 deletions(-)
>
> --
> 2.26.2
>
For convenience, I pushed this series on user branch `users/simark/submit/share-dwarf-partial-symtabs-v2`:
https://sourceware.org/git/?p=binutils-gdb.git;a=shortlog;h=refs/heads/users/simark/submit/share-dwarf-partial-symtabs-v2
Simon
More information about the Gdb-patches
mailing list