[libc++] Optimize basic_string::operator=(const basic_string&) for SSO assignments
authorEric Fiselier <eric@efcs.ca>
Wed, 15 Jan 2020 22:27:10 +0000 (17:27 -0500)
committerEric Fiselier <eric@efcs.ca>
Wed, 15 Jan 2020 22:27:10 +0000 (17:27 -0500)
commit58c7fa5adeb218737f10ffeaeed0d9eb21f4a2f0
tree55b1309fbf3d24a3bcc255fe067d155c2f7fefa7
parent5aeb6798f268f74dc8b0bf5bb75b9c57ec938548
[libc++] Optimize basic_string::operator=(const basic_string&) for SSO assignments

This change optimizes the operator=() assignment for short strings by direcly
copying the raw data from the source into the current instance. This creates an
optimized / inlined mempcy up to over 2X faster for short string assignments.
With inlining enabled for operator=, performance is up to 6X faster.

Benchmarks 'as is':
name                                    old time/op   new time/op    delta
BM_StringAssignStr_Empty_Opaque         6.05ns ± 2%   3.59ns ± 0%  -40.67%
BM_StringAssignStr_Empty_Transparent    5.15ns ± 0%   3.08ns ± 0%  -40.12%
BM_StringAssignStr_Small_Opaque         7.71ns ± 0%   3.59ns ± 0%  -53.45%
BM_StringAssignStr_Small_Transparent    7.66ns ± 0%   3.09ns ± 0%  -59.66%
BM_StringAssignStr_Large_Opaque         24.1ns ± 0%   24.9ns ± 0%   +3.22%
BM_StringAssignStr_Large_Transparent    22.2ns ± 0%   22.8ns ± 0%   +2.77%
BM_StringAssignStr_Huge_Opaque           315ns ± 6%    320ns ± 5%     ~
BM_StringAssignStr_Huge_Transparent      318ns ± 5%    321ns ± 4%     ~

Benchmarks with partial inlining operator=():
name                                    old time/op   new time/op    delta
BM_StringAssignStr_Empty_Opaque         5.94ns ± 2%   1.95ns ± 0%  -67.21%
BM_StringAssignStr_Empty_Transparent    5.14ns ± 0%   1.04ns ± 1%  -79.73%
BM_StringAssignStr_Small_Opaque         7.69ns ± 0%   1.96ns ± 0%  -74.48%
BM_StringAssignStr_Small_Transparent    7.65ns ± 0%   1.04ns ± 0%  -86.40%
BM_StringAssignStr_Large_Opaque         24.1ns ± 0%   24.5ns ± 0%   +1.61%
BM_StringAssignStr_Large_Transparent    22.2ns ± 0%   21.1ns ± 0%   -4.70%
BM_StringAssignStr_Huge_Opaque           317ns ± 5%    323ns ± 4%     ~
BM_StringAssignStr_Huge_Transparent      318ns ± 5%    320ns ± 5%     ~

Patch by Martijn Vels (mvels@google.com)
Reviewed as https://reviews.llvm.org/D72704
libcxx/include/string