[Bug gdb/23042] New: typedef resolution traverses all CUs for MI var-create

[s at martinien dot de]

https://sourceware.org/bugzilla/show_bug.cgi?id=23042

Bug ID: 23042
Summary: typedef resolution traverses all CUs for MI var-create
Product: gdb
Version: 8.1
Status: UNCONFIRMED
Severity: minor
Priority: P2
Component: gdb
Assignee: unassigned at sourceware dot org
Reporter: s at martinien dot de
Target Milestone: ---

Hi,

when using gdb via MI (e.g. via Visual Studio Code or Studio) on a build
compiled with -gsplit-dwarf, debugging becomes unusably slow.

This can be traced down to Visual Studio's feature which displays values of
function arguments in the callstack. They do this via simple MI commands:

1-var-create - * "i"

It appears this command is very slow when i is a common typedef, like size_t.
With this example:

---
typedef unsigned long size_ttt;
typedef unsigned long size_t;

int main()
{
size_ttt fast = 0;
size_t slow = 0;
return 0;
}
---

1-var-create - * "fast" is fast, but
1-var-create - * "slow" is slow.

Note that in my example I have lots of other shared libraries linked into the
program.

Stepping through gdb for these two commands above, the main difference is that
for the fast case, dw2_symtab_iter_next() always returns 0x0. For the slow case
it returns CUs, and then dw2_instantiate_symtab() is called which seems to take
all the time.

In the code I saw this comment:

dwarf2read.c
4241 /* Work around gold/15646. */
4242 if (!is_static && iter->global_seen)

Looking at https://sourceware.org/bugzilla/show_bug.cgi?id=15646 this seems to
be related to the gdbindex of gold, which incidentally was enabled on my
binaries. Switching to indexing via gdb solved the problem.

Now some questions:
- Can this be workarounded in gdb also? The gold builtin gdbindexer is faster,
so I would prefer using that.
- Should the gold gdbindex be fixed (can it?) Unfortunately nothing has
happened in the other bug... I will also try lld.
- Why does this only trigger via MI? "p slow", "ptype slow" both return
quickly.

I'm attaching a simple reproducer that shows the behavior for the two
var-create commands. You can place a breakpoint in dw2_symtab_iter_next() and
compare the return values.

Thanks and Best regards,
Martin

[s at martinien dot de]

https://sourceware.org/bugzilla/show_bug.cgi?id=23042

--- Comment #1 from Martin Richtarsky <s at martinien dot de> ---
Created attachment 10937
--> https://sourceware.org/bugzilla/attachment.cgi?id=10937&action=edit
Reproducer

[tromey at sourceware dot org]

https://sourceware.org/bugzilla/show_bug.cgi?id=23042

Tom Tromey <tromey at sourceware dot org> changed:

What |Removed |Added
----------------------------------------------------------------------------
CC| |tromey at sourceware dot org
Component|gdb |mi

[s at martinien dot de]

https://sourceware.org/bugzilla/show_bug.cgi?id=23042

--- Comment #2 from Martin Richtarsky <s at martinien dot de> ---
Tested with binutils-2.29, same issue.