Tentative patch:
...
diff --git a/gdb/dwarf2/cooked-index.c b/gdb/dwarf2/cooked-index.c
index 25635d9b72e..8134f8dbbfc 100644
--- a/gdb/dwarf2/cooked-index.c
+++ b/gdb/dwarf2/cooked-index.c
@@ -68,6 +68,25 @@ language_requires_canonicalization (enum language lang)
 
 /* See cooked-index.h.  */
 
+cooked_index_entry::cooked_index_entry (sect_offset die_offset_,
+                                       enum dwarf_tag tag_,
+                                       cooked_index_flag flags_,
+                                       const char *name_,
+                                       const cooked_index_entry *parent_entry_,

+                                       dwarf2_per_cu_data *per_cu_)
+  : name (name_),
+    tag (tag_),
+    flags (flags_),
+    die_offset (die_offset_),
+    parent_entry (parent_entry_),
+    per_cu (per_cu_)
+{
+  if (!language_requires_canonicalization (per_cu->lang ()))
+    canonical = name_;
+}
+
+/* See cooked-index.h.  */
+
 int
 cooked_index_entry::compare (const char *stra, const char *strb,
                             comparison_mode mode)
@@ -343,8 +362,9 @@ cooked_index_shard::do_finalize ()
 
   for (cooked_index_entry *entry : m_entries)
     {
-      /* Note that this code must be kept in sync with
-        language_requires_canonicalization.  */
+      if (!language_requires_canonicalization (entry->per_cu->lang ()))
+       continue;
+
       gdb_assert (entry->canonical == nullptr);
       if ((entry->flags & IS_LINKAGE) != 0)
        entry->canonical = entry->name;
@@ -388,7 +408,7 @@ cooked_index_shard::do_finalize ()
            }
        }
       else
-       entry->canonical = entry->name;
+       gdb_assert_not_reached ("Unhandled canonicalization");
     }--
 
   m_names.shrink_to_fit ();
diff --git a/gdb/dwarf2/cooked-index.h b/gdb/dwarf2/cooked-index.h
index 0d6f3e5aa0e..10207305109 100644
--- a/gdb/dwarf2/cooked-index.h
+++ b/gdb/dwarf2/cooked-index.h
@@ -78,15 +78,7 @@ struct cooked_index_entry : public allocate_on_obstack
   cooked_index_entry (sect_offset die_offset_, enum dwarf_tag tag_,
                      cooked_index_flag flags_, const char *name_,
                      const cooked_index_entry *parent_entry_,
-                     dwarf2_per_cu_data *per_cu_)
-    : name (name_),
-      tag (tag_),
-      flags (flags_),
-      die_offset (die_offset_),
-      parent_entry (parent_entry_),
-      per_cu (per_cu_)
-  {
-  }
+                     dwarf2_per_cu_data *per_cu_);
 
   /* Return true if this entry matches SEARCH_FLAGS.  */
   bool matches (block_search_flags search_flags) const
...

Created attachment 15004 [details]
tmp.patch

Alternative approach: make canonical std::atomic.

I suspect this is a false positive.
What is the value of "for_main" in that call to write_scope?

Using atomic here seems like it would cause performance problems.

The problem here is as follows.

This bit of code:
...
  const char *local_name = for_main ? name : canonical;
...
is compiled by gcc at -O2 to load both name and canonical.

This speculative load is considered harmless by the compiler, because it's considered side-effect free.

The speculation however introduces a data race, which tsan reports.

(In reply to Tom Tromey from comment #3)
> I suspect this is a false positive.

Agreed.

> What is the value of "for_main" in that call to write_scope?
> 

I think it's false, but due to optimization gcc speculatively loads "canonical".

> Using atomic here seems like it would cause performance problems.

I tested that a bit, saw no impact, but yeah that could be the case.

The culprint in gcc seems to be -fhoist-adjacent-loads.

(In reply to Tom de Vries from comment #6)
> The culprint in gcc seems to be -fhoist-adjacent-loads.

I've filed https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110799 .

I'm continuing my excercise with -fsanitize=thread -O2 -fno-hoist-adjacent-loads.

For now, closing as resolved-moved.

*** Bug 30671 has been marked as a duplicate of this bug. ***