ned14 / stl-header-heft

Measures how parsing overweight the major STLs have become

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

STL Header Heft

(C) 2018 - 2020 Niall Douglas http://www.nedproductions.biz/

Herein lie some STL header token parsing benchmarks as according to my pure Python C99 preprocessor pcpp. Although written in Python, it scales linearly with tokens and has comparable performance dynamics to a preprocessor/tokeniser written in C or C++ (i.e. it will be mostly some fixed linear multiple slower).

C tokens per file actually has little correlation with header file size which was a surprise to me. Most C++ compilers will generate an internal DAG (often called an AST) from the tokens parsed from all the includes and source files. The generation of this in-memory representation is usually mostly linear to token count. Only when templates get instantiated and inline functions compiled does the compile time begin to vary significantly from the parse time, particularly if a lot of SFINAE or complex constexpr is being repeatedly executed by the compiler.

These benchmarks say nothing about how complex a header is to compile, only how complex it is to parse. Nevertheless, for million file builds, milliseconds of difference can add up quickly. Therefore it would be really useful to know:

  • What C++ STL headers are low token count to include?
    • Always?
    • Or depending on STL?
  • What C++ STL headers are high token count to include?
    • Always?
    • Or depending on STL?

libstdc++ 7, 8, 9 (2017-2020)

Detailed observation notes

VS2017, VS2019 (2018-2020)

Detailed observation notes

Low token count C++ 20 headers on both libstdc++ 9 and VS2019

Each of these is in the bottom quartile of token parse times for all STL headers in their respective STLs. You are therefore more likely to be safe if you include only these:

  • cassert
  • cctype
  • cerrno
  • cfenv
  • cfloat
  • cinttypes
  • ciso646
  • climits
  • clocale
  • compare
  • csetjmp
  • csignal
  • cstdalign
  • cstdarg
  • cstdbool
  • cstddef
  • cstdint
  • cstring
  • ctime
  • cuchar
  • cwctype
  • initializer_list
  • version

Some of the C headers like <cmath>, <cstdio> and <cstdlib> can be quite heavy in some STLs!

Highest token count C++ 20 headers on either libstdc++ 9 and VS2019

You may wish to avoid using these headers, especially in global interface files, if compile times across GCC and MSVC are very important to you:

  • algorithm
  • array
  • bitset
  • ccomplex
  • codecvt
  • complex
  • condition_variable
  • ctgmath
  • deque
  • execution
  • filesystem
  • forward_list
  • fstream
  • functional
  • future
  • iomanip
  • ios
  • iostream
  • istream
  • iterator
  • list
  • locale
  • map
  • memory
  • memory_resource
  • mutex
  • ostream
  • queue
  • random
  • ranges
  • regex
  • scoped_allocator
  • set
  • shared_mutex
  • sstream
  • stack
  • stdexcept
  • streambuf
  • string
  • string_view
  • strstream
  • system_error
  • thread
  • unordered_map
  • unordered_set
  • valarray
  • variant
  • vector

About

Measures how parsing overweight the major STLs have become


Languages

Language:Python 100.0%