[all-commits] [llvm/llvm-project] a64d3c: [NFC][Lexer] Make Lexer::LangOpts const reference

[yurai007 via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: a64d3c602fb7e533855a64cb4d9fec77ea0a079e
      https://github.com/llvm/llvm-project/commit/a64d3c602fb7e533855a64cb4d9fec77ea0a079e
  Author: Dawid Jurczak <dawid_jurek at vp.pl>
  Date:   2022-02-28 (Mon, 28 Feb 2022)

  Changed paths:
    M clang/include/clang/Lex/Lexer.h
    M clang/lib/Format/FormatTokenLexer.cpp
    M clang/lib/Format/FormatTokenLexer.h
    M clang/lib/Lex/Lexer.cpp

  Log Message:
  -----------
  [NFC][Lexer] Make Lexer::LangOpts const reference

This change can be seen as code cleanup but motivation is more performance related.
While browsing perf reports captured during Linux build we can notice unusual portion of instructions executed in std::vector<std::string> copy constructor like:

0.59%     0.58%  clang-14    clang-14      [.] std::vector<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >,
                                                                std::allocator<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > >::vector

or even:

1.42%     0.26%  clang    clang-14             [.] clang::LangOptions::LangOptions
       |
        --1.16%--clang::LangOptions::LangOptions
                  |
                   --0.74%--std::vector<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >,
                            std::allocator<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > >::vector

After more digging we can see that relevant LangOptions std::vector members (*Files, ModuleFeatures and NoBuiltinFuncs)
are constructed when Lexer::LangOpts field is initialized on list:

Lexer::Lexer(..., const LangOptions &langOpts, ...)
            : ..., LangOpts(langOpts),

Since LangOptions copy constructor is called by Lexer(..., const LangOptions &LangOpts,...) and local Lexer objects are created thousands times
(in Lexer::getRawToken, Preprocessor::EnterSourceFile and more) during single module processing in frontend it makes std::vector copy constructors surprisingly hot.

Unfortunately even though in current Lexer implementation mentioned std::vector members are unused and most of time empty,
no compiler is smart enough to optimize their std::vector copy constructors out (take a look at test assembly): https://godbolt.org/z/hdoxPfMYY even with LTO enabled.
However there is simple way to fix this. Since Lexer doesn't access *Files, ModuleFeatures, NoBuiltinFuncs and any other LangOptions fields (but only LangOptionsBase)
we can simply get rid of redundant copy constructor assembly by changing LangOpts type to more appropriate const LangOptions reference: https://godbolt.org/z/fP7de9176

Additionally we need to store LineComment outside LangOpts because it's written in SkipLineComment function.
Also FormatTokenLexer need to be adjusted a bit to avoid lifetime issues related to passing local LangOpts reference to Lexer.

After this change I can see more than 1% speedup in some of my microbenchmarks when using Clang release binary built with LTO.
For Linux build gains are not so significant but still nice at the level of -0.4%/-0.5% instructions drop.

Differential Revision: https://reviews.llvm.org/D120334