Re: [PATCH v4 1/3] perf report: Support LLVM for addr2line()

[Namhyung Kim]

Hello,

Thanks a lot for the patches!

On Mon, May 20, 2024 at 1:31 AM Steinar H. Gunderson <sesse@xxxxxxxxxx> wrote:
>
> In addition to the existing support for libbfd and calling out to
> an external addr2line command, add support for using libllvm directly.
> This is both faster than libbfd, and can be enabled in distro builds
> (the LLVM license has an explicit provision for GPLv2 compatibility).
> Thus, it is set as the primary choice if available.
>
> As an example, running perf report on a medium-size profile with
> DWARF-based backtraces took 58 seconds with LLVM, 78 seconds with
> libbfd, 153 seconds with external llvm-addr2line, and I got tired
> and aborted the test after waiting for 55 minutes with external
> bfd addr2line (which is the default for perf as compiled by distributions
> today). Evidently, for this case, the bfd addr2line process needs
> 18 seconds (on a 5.2 GHz Zen 3) to load the .debug ELF in question,
> hits the 1-second timeout and gets killed during initialization,
> getting restarted anew every time. Having an in-process addr2line
> makes this much more robust.
>
> As future extensions, libllvm can be used in many other places where
> we currently use libbfd or other libraries:
>
>  - Symbol enumeration (in particular, for PE binaries).
>  - Demangling (including non-Itanium demangling, e.g. Microsoft
>    or Rust).
>  - Disassembling (perf annotate).

I think it should support other DWARF use cases like
unwinding and type info?

>
> However, these are much less pressing; most people don't profile
> PE binaries, and perf has non-bfd paths for ELF. The same with
> demangling; the default _cxa_demangle path works fine for most
> users. Disassembling is coming in a later patch in the series;
> however do note that while bfd objdump can be slow on large binaries,
> it is possible to use --objdump=llvm-objdump to get the speed benefits.

I remember bfd objdump is sometimes faster than llvm-objdump
especially when no line numbers are requested IIRC.


> (It appears LLVM-based demangling is very simple, should we want
> that.)
>
> Tested with LLVM 14, 15, 16, 18 and 19. For some reason, LLVM 12 was not
> correctly detected using feature_check, and thus was not tested.

Anyway, nice work.  Maybe we can implement other use cases
using LLVM and reduce the dependencies.

Thanks,
Namhyung

>
> Signed-off-by: Steinar H. Gunderson <sesse@xxxxxxxxxx>