diff mbox

Remove redundant call to lookup_static_symbol.

Message ID m361di0xur.fsf@seba.sebabeach.org
State New
Headers show

Commit Message

Doug Evans Dec. 11, 2014, 7:50 p.m. UTC

Anytime you can remove a symbol lookup that loops over all objfiles
is A Good Thing.

The call to lookup_static_symbol in valops.c:value_maybe_namespace_elt
is redundant with this call in cp_lookup_nested_symbol:

	/* Now search all static file-level symbols.  We have to do this
	   for things like typedefs in the class.  We do not try to
	   guess any imported namespace as even the fully specified
	   namespace search is already not C++ compliant and more
	   assumptions could make it too magic.  */

	size = strlen (parent_name) + 2 + strlen (nested_name) + 1;
	concatenated_name = alloca (size);
	xsnprintf (concatenated_name, size, "%s::%s",
		 parent_name, nested_name);
	sym = lookup_static_symbol (concatenated_name, VAR_DOMAIN);
	if (sym != NULL)
	  return sym;

Earlier in value_maybe_namespace_elt we do this:

  sym = cp_lookup_symbol_namespace (namespace_name, name,
				    get_selected_block (0), VAR_DOMAIN);

That call goes like:

-> cp_lookup_symbol_namespace
-> cp_lookup_symbol_in_namespace
-> lookup_symbol_file
-> cp_lookup_nested_symbol
-> lookup_static_symbol

The call was added in commit 41f62f3939b1c69e68ef5652feb44fef90eb85c9.
With a part 2 here:

At the time the call to lookup_static_symbol (spelled
lookup_static_symbol_aux at the time) was needed.

However, this patch, 8dea366bbed7986295681c101dcfbd35aeb6dfc4,
augmented lookup_symbol_file to call cp_lookup_nested_symbol
and introduced the redundancy.
[I'm not discounting the possibility that there's some subtle
edge case I'm missing, with GDB that's always a possibility.
But I can't find one, and we've got to make forward progress
at cleaning gdb up.]

It's kinda buried, so it's totally not unexpected that this happened.
Part of what I want to do is make the symbol code cleaner so that it's
easier to avoid such regressions in the future.
That, plus spiffing up the performance testsuite to help catch these.
[Though a symbol cache will prevent such regressions, but I don't
want the cache to hide problems, so we'll want testing with/without it.]

Regression tested on amd64-linux.

Plus, tested by
- removing both calls to lookup_static_symbol
  (value_maybe_namespace_elt and cp_lookup_nested_symbol),
- seeing the testsuite failures (e.g., derivation.exp, namespace.exp,
- putting the call in value_maybe_namespace_elt back,
  seeing some (but not all - not unexpectedly) pass again

Also tested by turning on "set debug symbol-lookup 2"
(see https://sourceware.org/ml/gdb-patches/2014-12/msg00255.html),
and comparing the output.

2014-12-10  Doug Evans  <xdje42@gmail.com>

	* valops.c (value_maybe_namespace_elt): Remove redundant call to


Doug Evans Dec. 17, 2014, 8:35 a.m. UTC | #1
On Thu, Dec 11, 2014 at 11:50 AM, Doug Evans <xdje42@gmail.com> wrote:
> Hi.
> Anytime you can remove a symbol lookup that loops over all objfiles
> is A Good Thing.
> [...]
> 2014-12-10  Doug Evans  <xdje42@gmail.com>
>         * valops.c (value_maybe_namespace_elt): Remove redundant call to
>         lookup_static_symbol.

diff mbox


diff --git a/gdb/valops.c b/gdb/valops.c
index 4125fc0..2e3c9a1 100644
--- a/gdb/valops.c
+++ b/gdb/valops.c
@@ -3570,15 +3570,6 @@  value_maybe_namespace_elt (const struct type *curtype,
 				    get_selected_block (0), VAR_DOMAIN);
   if (sym == NULL)
-    {
-      char *concatenated_name = alloca (strlen (namespace_name) + 2
-					+ strlen (name) + 1);
-      sprintf (concatenated_name, "%s::%s", namespace_name, name);
-      sym = lookup_static_symbol (concatenated_name, VAR_DOMAIN);
-    }
-  if (sym == NULL)
     return NULL;
   else if ((noside == EVAL_AVOID_SIDE_EFFECTS)
 	   && (SYMBOL_CLASS (sym) == LOC_TYPEDEF))