Commit Graph

4 Commits

Author SHA1 Message Date
Sami Tolvanen
10e9510a6d gendwarfksyms: Add a separate pass to resolve FQNs
Using dwarf_getscopes_die to resolve fully-qualified names turns out to
be rather slow, and also results in duplicate scopes being processed,
which doesn't help. Simply adding an extra pass to resolve names for all
DIEs before processing exports is noticeably faster.

For the object files with the most exports in a defconfig+Rust build,
the performance improvement is consistently >50%:

rust/bindings.o: 1038 exports
    before: 9.5980 +- 0.0183 seconds time elapsed  ( +-  0.19% )
     after: 4.3116 +- 0.0287 seconds time elapsed  ( +-  0.67% )

rust/core.o: 424 exports
    before: 5.3584 +- 0.0204 seconds time elapsed  ( +-  0.38% )
     after: 0.05348 +- 0.00129 seconds time elapsed  ( +-  2.42% )
            ^ Not a mistake.

net/core/dev.o: 190 exports
    before: 9.0507 +- 0.0297 seconds time elapsed  ( +-  0.33% )
     after: 3.2882 +- 0.0165 seconds time elapsed  ( +-  0.50% )

rust/kernel.o: 129 exports
    before: 6.8571 +- 0.0317 seconds time elapsed  ( +-  0.46% )
     after: 2.9096 +- 0.0316 seconds time elapsed  ( +-  1.09% )

net/core/skbuff.o: 120 exports
    before: 5.4805 +- 0.0291 seconds time elapsed  ( +-  0.53% )
     after: 2.0339 +- 0.0231 seconds time elapsed  ( +-  1.14% )

drivers/gpu/drm/display/drm_dp_helper.o: 101 exports
    before: 1.7877 +- 0.0187 seconds time elapsed  ( +-  1.05% )
     after: 0.69245 +- 0.00994 seconds time elapsed  ( +-  1.44% )

net/core/sock.o: 97 exports
    before: 5.8327 +- 0.0653 seconds time elapsed  ( +-  1.12% )
     after: 2.0784 +- 0.0291 seconds time elapsed  ( +-  1.40% )

drivers/net/phy/phy_device.o: 95 exports
    before: 3.0671 +- 0.0371 seconds time elapsed  ( +-  1.21% )
     after: 1.2127 +- 0.0207 seconds time elapsed  ( +-  1.70% )

drivers/pci/pci.o: 93 exports
    before: 1.1130 +- 0.0113 seconds time elapsed  ( +-  1.01% )
     after: 0.4848 +- 0.0127 seconds time elapsed  ( +-  2.63% )

kernel/sched/core.o: 83 exports
    before: 3.5092 +- 0.0223 seconds time elapsed  ( +-  0.64% )
     after: 1.1231 +- 0.0145 seconds time elapsed  ( +-  1.29% )

Overall, a defconfig+DWARF5 build with gendwarfksyms and Rust is 14.8%
faster with this patch applied on my test system. Without Rust, there's
still a 10.4% improvement in build time when gendwarfksyms is used.

Note that symbol versions are unchanged with this patch.

Suggested-by: Giuliano Procida <gprocida@google.com>
Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
2025-03-15 21:16:11 +09:00
Sami Tolvanen
ab4439981f gendwarfksyms: Add symtypes output
Add support for producing genksyms-style symtypes files. Process
die_map to find the longest expansions for each type, and use symtypes
references in type definitions. The basic file format is similar to
genksyms, with two notable exceptions:

  1. Type names with spaces (common with Rust) in references are
     wrapped in single quotes. E.g.:

     s#'core::result::Result<u8, core::num::error::ParseIntError>'

  2. The actual type definition is the simple parsed DWARF format we
     output with --dump-dies, not the preprocessed C-style format
     genksyms produces.

Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
Reviewed-by: Petr Pavlu <petr.pavlu@suse.com>
Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
2025-01-11 01:25:25 +09:00
Sami Tolvanen
06b8b036ab gendwarfksyms: Expand type modifiers and typedefs
Add support for expanding DWARF type modifiers, such as pointers,
const values etc., and typedefs. These types all have DW_AT_type
attribute pointing to the underlying type, and thus produce similar
output.

Also add linebreaks and indentation to debugging output to make it
more readable.

Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
Reviewed-by: Petr Pavlu <petr.pavlu@suse.com>
Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
2025-01-11 01:25:25 +09:00
Sami Tolvanen
0c1c76274e gendwarfksyms: Add a cache for processed DIEs
Basic types in DWARF repeat frequently and traversing the DIEs using
libdw is relatively slow. Add a simple hashtable based cache for the
processed DIEs.

Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
Reviewed-by: Petr Pavlu <petr.pavlu@suse.com>
Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
2025-01-11 01:25:25 +09:00