[all-commits] [llvm/llvm-project] 403d61: [lld-macho] Enable EH frame relocation / pruning

[Jez Ng via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 403d61aeddec87fd210ee1425658f22367fc088a
      https://github.com/llvm/llvm-project/commit/403d61aeddec87fd210ee1425658f22367fc088a
  Author: Jez Ng <jezng at fb.com>
  Date:   2022-07-13 (Wed, 13 Jul 2022)

  Changed paths:
    M lld/MachO/Config.h
    M lld/MachO/Driver.cpp
    M lld/MachO/InputFiles.cpp
    M lld/docs/ReleaseNotes.rst

  Log Message:
  -----------
  [lld-macho] Enable EH frame relocation / pruning

This just removes the code that gates the logic. The main issue here is
perf impact: without {D122258}, LLD takes a significant perf hit because
it now has to do a lot more work in the input parsing phase. But with
that change to eliminate unnecessary EH frames from input object files,
the perf overhead here is minimal. Concretely, here are the numbers for
some builds as measured on my 16-core Mac Pro:

**chromium_framework**

This is without the use of `-femit-dwarf-unwind=no-compact-unwind`:

             base           diff           difference (95% CI)
  sys_time   1.826 ± 0.019  1.962 ± 0.034  [  +6.5% ..   +8.4%]
  user_time  9.306 ± 0.054  9.926 ± 0.082  [  +6.2% ..   +7.1%]
  wall_time  8.225 ± 0.068  8.947 ± 0.128  [  +8.0% ..   +9.6%]
  samples    15             22

With that flag enabled, the regression mostly disappears, as hoped:

             base           diff           difference (95% CI)
  sys_time   1.839 ± 0.062  1.866 ± 0.068  [  -0.9% ..   +3.8%]
  user_time  9.452 ± 0.068  9.490 ± 0.067  [  -0.1% ..   +0.9%]
  wall_time  8.383 ± 0.127  8.452 ± 0.114  [  -0.1% ..   +1.8%]
  samples    17             21

**Unnamed internal app**

Without `-femit-dwarf-unwind`, this is the perf hit:

             base           diff           difference (95% CI)
  sys_time   1.372 ± 0.029  1.317 ± 0.024  [  -4.6% ..   -3.5%]
  user_time  2.835 ± 0.028  2.980 ± 0.027  [  +4.8% ..   +5.4%]
  wall_time  3.205 ± 0.079  3.383 ± 0.066  [  +4.9% ..   +6.2%]
  samples    102            83

With `-femit-dwarf-unwind`, the perf hit almost disappears:

             base           diff           difference (95% CI)
  sys_time   1.274 ± 0.026  1.270 ± 0.025  [  -0.9% ..   +0.3%]
  user_time  2.812 ± 0.023  2.822 ± 0.035  [  +0.1% ..   +0.7%]
  wall_time  3.166 ± 0.047  3.174 ± 0.059  [  -0.2% ..   +0.7%]
  samples    95             97

Just for fun, I measured the impact of `-femit-dwarf-unwind` on ld64
(`base` has the extra DWARF unwind info in the input object files,
`diff` doesn't):

             base           diff           difference (95% CI)
  sys_time   1.128 ± 0.010  1.124 ± 0.023  [  -1.3% ..   +0.6%]
  user_time  7.176 ± 0.030  7.106 ± 0.094  [  -1.5% ..   -0.4%]
  wall_time  7.874 ± 0.041  7.795 ± 0.121  [  -1.7% ..   -0.3%]
  samples    16             25

And for LLD:

             base           diff           difference (95% CI)
  sys_time   1.315 ± 0.019  1.280 ± 0.019  [  -3.2% ..   -2.0%]
  user_time  2.980 ± 0.022  2.822 ± 0.016  [  -5.5% ..   -5.0%]
  wall_time  3.369 ± 0.038  3.175 ± 0.033  [  -6.2% ..   -5.3%]
  samples    47             47

So parsing the extra EH frames is a lot more expensive for us than for
ld64. But given that we are quite a lot faster than ld64 to begin with,
I guess this isn't entirely unexpected...

Reviewed By: #lld-macho, oontvoo

Differential Revision: https://reviews.llvm.org/D129540