> -----Original Message-----
> From: Thomas Neumann <thomas.neum...@in.tum.de>
> Sent: Monday, November 21, 2022 11:23 AM
> To: Tamar Christina <tamar.christ...@arm.com>; gcc-patches@gcc.gnu.org;
> Jason Merrill <ja...@redhat.com>
> Cc: Florian Weimer <fwei...@redhat.com>; Jakub Jelinek
> <ja...@redhat.com>; Jonathan Wakely <jwakely....@gmail.com>
> Subject: Re: [PATCH v4] eliminate mutex in fast path of __register_frame
>
> Hi,
>
> > When dynamically linking a fast enough machine hides the latency, but
> > when Statically linking or on slower devices this change caused a 5x
> > increase in Instruction count and 2x increase in cycle count before getting
> to main.
> >
> > This has been quite noticeable on smaller devices. Is there a reason
> > the btree can't be initialized lazily? It seems a bit harsh to pay the
> > cost of unwinding at startup even when you don't throw exceptions..
>
> we cannot easily do that lazily because otherwise we need a mutex for lazy
> initialization, which is exactly what we wanted to get rid of.
>
> Having said that, I am surprised that you saw a noticeable difference.
> On most platforms there should not be dynamic frame registration at all, as
> the regular frames are directly read from the ELF data.
>
> Can you please send me an precise description on how to reproduce the
> issue? (Platform, tools, a VM if you have one would be great). I will then
> debug this to improve the startup time.
It's easy to reproduce on x86 as well.
As a testcase:
#include <cstdio>
int main(int argc, char** argv) {
return 0;
}
And just compile with: g++ -O1 hello.cpp -static -o hello.exe.
Before this change on x86 I got:
> perf stat -r 200 ./hello.exe
Performance counter stats for './hello.exe' (200 runs):
0.32 msec task-clock # 0.326 CPUs utilized
( +- 0.34% )
0 context-switches # 0.000 K/sec
0 cpu-migrations # 0.000 K/sec
22 page-faults # 0.070 M/sec
( +- 0.13% )
310,194 cycles # 0.984 GHz
( +- 0.33% )
317,310 instructions # 1.02 insn per cycle
( +- 0.18% )
58,885 branches # 186.710 M/sec
( +- 0.12% )
931 branch-misses # 1.58% of all branches
( +- 2.57% )
0.00096799 +- 0.00000374 seconds time elapsed ( +- 0.39% )
And after this change:
> perf stat -r 200 ./hello.exe
Performance counter stats for './hello.exe' (200 runs):
1.03 msec task-clock # 0.580 CPUs utilized
( +- 0.23% )
0 context-switches # 0.000 K/sec
0 cpu-migrations # 0.000 K/sec
27 page-faults # 0.026 M/sec
( +- 0.10% )
1,034,038 cycles # 1.002 GHz
( +- 0.11% )
2,485,983 instructions # 2.40 insn per cycle
( +- 0.02% )
557,567 branches # 540.215 M/sec
( +- 0.01% )
4,843 branch-misses # 0.87% of all branches
( +- 0.53% )
0.00178093 +- 0.00000456 seconds time elapsed ( +- 0.26% )
Regards,
Tamar
>
> Best
>
> Thomas
