On 12/6/22 14:02, Richard W.M. Jones wrote:
On Tue, Dec 06, 2022 at 05:52:16PM -0000, Andrii Nakryiko wrote:
>> On Tue, Dec 06, 2022 at 03:12:19AM +0000, Gary Buhrmaster wrote:
>>
>> Note that is not a fully equivalent scenario. The no-omit-frame-pointer
>> proposal was only offering a functional debugging benefit to a fairly
>> small number of users who are also developers, while adding a likely
>> performance hit to all users. There needs to be a high bar to justify
>> the performance hit when the benefit offered is narrow.
>
> First, frame pointers are not just for debugging benefit. It's not even it's
main benefit from POV of
https://pagure.io/fesco/issue/2817. Frame pointers are for
performance profiling and observability first and foremost, and that's most useful
under real-world conditions of production workloads. Not some custom re-built debug
versions of applications.
>
> Second, it might benefit a relatively small (but not tiny, it's at least
thousands of people doing performance profiling) fraction of users, but those users
(developers that care about performance) are the ones bringing benefits to very wide user
base.
Yes! I spent a frustrating time getting perf to record stack traces
properly until I recompiled the program with frame pointers. (I know
about --call-graph=dwarf but it doesn't seem to work most of the time.)
That is due to known limitations in perf, IIUC. Hence why at least I was
pushing so heavily to improve perf to not require frame pointers.
--
Sincerely,
Demi Marie Obenour (she/her/hers)