File gdb-symtab-fix-slow-symbol-lookup-with-dwz.patch of Package gdb

From 8b5dc3da60d6c07991821c447f183288c4aac69e Mon Sep 17 00:00:00 2001
From: Tom de Vries <tdevries@suse.de>
Date: Tue, 10 Feb 2026 10:30:59 +0100
Subject: [PATCH] [gdb/symtab] Fix slow symbol lookup with dwz

Consider the following setup on openSUSE Leap 16.0 x86_64:
...
$ sudo zypper install qt6-tools-assistant qt6-tools-assistant-debuginfo \
    libQt6Core6-debuginfo
$ gdb -q \
    -iex "set debuginfod enabled off" \
    -iex "set pagination off" \
    /usr/bin/assistant6 \
    -ex "break BookmarkManager::setupFinished" \
    -ex run
Reading symbols from /usr/bin/assistant6...
Reading symbols from /usr/lib/debug/usr/lib64/qt6/bin/assistant.debug...
   ...
Thread 1 "assistant6" hit Breakpoint 1, BookmarkManager::setupFinished (...)
    at .../bookmarkmanager.cpp:297
297	{
(gdb) maint time 1
(gdb)
...

[ Note that we need packages:
- qt6-tools-assistant for /usr/bin/assistant6
- qt6-tools-assistant-debuginfo to be able to set the breakpoint, and
- libQt6Core6-debuginfo to have debuginfo for shared library
  /lib64/libQt6Core.so.6. ]

This lookup is slow with an -O2 build of gdb-17-branch:
...
(gdb) print *bookmarkWidget
  ...
Command execution time: 17.712828 (cpu), 17.712880 (wall)
(gdb)
...

[ Note that a second lookup takes roughly the same time, so the slowness is
not due to symtab expansion.

Also note that the slowness is gone:
...
Command execution time: 0.091105 (cpu), 0.091100 (wall)
...
if we de-install libQt6Core6-debuginfo. ]

This is PR symtab/33825, originally reported for gdb 16.3 on openSUSE
Tumbleweed x86_64.

With current trunk, we have instead:
...
Command execution time: 0.009536 (cpu), 0.009509 (wall)
...

The slowness is gone at commit ce889924a72 ("Simplify
basic_lookup_transparent_type"), but the commit is part of a patch series of
31 patches, so the commit cannot be trivially backported to gdb-17-branch.

The slowness with the gdb-17-branch is caused by the following.

In basic_lookup_transparent_type_1, we have this bit of code:
...
   for (compunit_symtab *cust : objfile->compunits ())
     {
       bv = cust->blockvector ();
       block = bv->block (block_index);
       sym = block_find_symbol (block, name, flags, nullptr);
...
and via block_find_symbol -> block_iterator_range -> block_iterator_first ->
initialize_block_iterator we get to this code:
...
  compunit_symtab *cu = block->global_block ()->compunit ();

  /* If this is an included symtab, find the canonical includer and
     use it instead.  */
  while (cu->user != NULL)
    cu = cu->user;
...

Say we have an objfile with two cus, CU1 and CU2, where:
- CU1 represents a DWARF compilation unit, and
- CU2 represents a DWARF partial unit imported by CU1.

In symtab terms, CU2 is an "included symtab", and CU1 is its "canonical
includer", so we have:
- CU1->user == nullptr, and
- CU2->user == CU1.

In the for loop shown above, we have subsequently:
- cust == CU1: during block_find_symbol, we visit both CU1 and included symtab
  CU2.
- cust == CU2: during block_find symbol, we walk up from CU2 to its canonical
  includer CU1, and again visit CU1 and included symtab CU2.
So, instead of visiting CU1 and CU2 each once, we visited them twice.

Fix this by skipping included symtabs in the for loop, getting us instead the
faster:
...
Command execution time: 0.293385 (cpu), 0.293386 (wall)
...

The fix can be trivially backported to the gdb-16-branch.

Tested on x86_64-linux, with target boards unix, cc-with-dwz and
cc-with-dwz-m.

Approved-By: Tom Tromey <tom@tromey.com>

Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=33825
(cherry picked from commit 36356c77b8969b684566c96b01cc97ace6a449bf)
---
 gdb/symtab.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/gdb/symtab.c b/gdb/symtab.c
index 106e540ca48..569fd18e5e4 100644
--- a/gdb/symtab.c
+++ b/gdb/symtab.c
@@ -2785,6 +2785,11 @@ basic_lookup_transparent_type_1 (struct objfile *objfile,
 
   for (compunit_symtab *cust : objfile->compunits ())
     {
+      /* Skip included compunits, to avoid searching them and their canonical
+	 includers more than once.  */
+      if (cust->user != nullptr)
+	continue;
+
       bv = cust->blockvector ();
       block = bv->block (block_index);
       sym = block_find_symbol (block, name, flags, nullptr);

base-commit: 6da3492819000aa16136905bab03c312b290ab67
-- 
2.51.0