[binutils-gdb] Fix TAB-completion + .gdb_index slowness (generalize filename_seen_cache)

[sergiodj+buildbot at sergiodj dot net]

*** TEST RESULTS FOR COMMIT bbf2f4dfaec5cf2e21b0935300b4921f0b5a8eb7 ***

Author: Pedro Alves <palves@redhat.com>
Branch: master
Commit: bbf2f4dfaec5cf2e21b0935300b4921f0b5a8eb7

Fix TAB-completion + .gdb_index slowness (generalize filename_seen_cache)

Tab completion when debugging a program binary that uses GDB index is
surprisingly much slower than when GDB uses psymtabs instead.  Around
1.5x/3x slower.  That's surprising, because the whole point of GDB
index is to speed things up...

For example, with:

 set pagination off
 set $count = 0
 while $count < 400
   complete b string_prin         # matches gdb's string_printf
   printf "count = %d\n", $count
   set $count = $count + 1
 end

 $ time ./gdb --batch -q  ./gdb-with-index -ex "source script.cmd"
 real    0m11.042s
 user    0m10.920s
 sys     0m0.042s

 $ time ./gdb --batch -q  ./gdb-without-index -ex "source script.cmd"
 real    0m4.635s
 user    0m4.590s
 sys     0m0.037s

Same but with:
 -   complete b string_prin
 +   complete b zzzzzz
to exercise the no-matches worst case, master currently gets you
something like:

 with index           without index
 real    0m11.971s    0m8.413s
 user    0m11.912s    0m8.355s
 sys     0m0.035s     0m0.035s

Running gdb under perf shows 80% spent inside
maybe_add_partial_symtab_filename, and 20% spent in the lbasename
inside that.

The problem that tab completion walks over all compunit symtabs, and
for each, walks the contained file symtabs.  And there a huge number
of file symtabs (each included system header, etc.) that appear in
each compunit symtab's file symtab list.  As in, when debugging GDB, I
have 367381 symtabs iterated, when of those only 5371 filenames are
unique...

This was a regression from the earlier (nice) split of symtabs in
compunit symtabs + file symtabs.

The fix here is to add a cache of unique filenames per objfile so that
the walk / uniquing is only done once.  There's already a abstraction
for this in symtab.c; this patch moves that code out to a separate
file and C++ifies it bit.

This makes the worst-case scenario above consistently drop to ~2.5s
(1.5s for the "string_prin" hit case), making it over 3.3x times
faster than psymtabs in this use case (7x in the "string_prin" hit
case).

gdb/ChangeLog:
2017-07-17  Pedro Alves  <palves@redhat.com>

	* Makefile.in (COMMON_OBS): Add filename-seen-cache.o.
	* dwarf2read.c: Include "filename-seen-cache.h".
	* dwarf2read.c (dwarf2_per_objfile) <filenames_cache>: New field.
	(dw2_map_symbol_filenames): Build and use a filenames_seen_cache.
	* filename-seen-cache.c: New file.
	* filename-seen-cache.h: New file.
	* symtab.c: Include "filename-seen-cache.h".
	(struct filename_seen_cache, INITIAL_FILENAME_SEEN_CACHE_SIZE)
	(create_filename_seen_cache, clear_filename_seen_cache)
	(delete_filename_seen_cache, filename_seen): Delete, parts moved
	to filename-seen-cache.h/filename-seen-cache.c.
	(output_source_filename, sources_info)
	(maybe_add_partial_symtab_filename)
	(make_source_files_completion_list): Adjust to use
	filename_seen_cache.