GitLab

LockImpl::Unlock() is inlined, but not LockImpl::Lock(). Separately,
Lock() calls Try() before locking, to increase uncontended
performance. This code is duplicated between POSIX and Windows. This
commit:
- Merges the two paths
- Inlines Lock() and Try(), partly for symmetry, and partly for
  performance

On a Linux Xeon "Haswell" workstation (E5-2690v4), this improves
uncontended acquire/release pairs by 8%. These numbers are collected
from a non-PGO build, and the impact on a PGO one is likely smaller.

This was assessed by running the performance test 50 times with:
$ out/Release-desktop/base_perftests \
    --gtest_filter="LockPerfTest.Simple" --gtest_repeat=50 \
    | grep RESULT | sed -e 's/.*= //;s/ .*//'

Results are summarized below (numbers are runs/s, higher is better):
trunk:   Mean = 4.534e+07  Standard Deviation = 6.234e+05
This CL: Mean = 4.900e+07  Standard Deviation = 6.681e+05

That is, this patch makes it 8.08% faster, and given the standard
deviation, the difference is likely significant.

Bug: 1061437
Change-Id: Ibfb3f1f07598c22f1c9446c02e16a883bf328d20
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/2386738Reviewed-by: Wez <wez@chromium.org>
Reviewed-by: Bruce Dawson <brucedawson@chromium.org>
Commit-Queue: Benoit L <lizeb@chromium.org>
Cr-Commit-Position: refs/heads/master@{#804963}