There is a request in the "C++ Standard Library Active Issues List" for The char_traits specializations should declare their length(), compare(), and find() members constexpr. Currently the following code fails to compile:
#include <string_view> // Compile time error: // > error: constexpr variable 'service' must be initialized by a constant expression // > constexpr string_view service = "HELLO WORD SERVICE"; // > ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ // > string_view:110:39: note: non-constexpr function 'length' cannot be used in a constant expression // constexpr string_view service = "HELLO WORD SERVICE";
This proposal attempts to solve the issue 2232 and make std::string_view
usable in constant expressions.
std::basic_string_view
uses from traits_type
only the static member functions length
, compare
and find
. Only those functions are affected by this proposal.
A proof of concept implementation with benchmark is available at: constexpr char_traits.
This proposal is a pure library extension. It proposes changes to the
existing header <string>
such that the changes do not break existing code
and do not degrade performance. It does not require any changes in the core
language and it can be (and has been) implemented in standard C++.
std::char_traits
implementations.libstdc++ and libc++ have implementations of std::char_traits<char16_t>
and std::char_traits<char32_t>
that do not use C functions or intrinsics. Those two std::char_traits
could be made constexpr without any impact on performance.
libstdc++ uses __builtin_strlen
, __builtin_memcmp
and __builtin_memchr
intrinsics for std::char_traits<char>
. Those extensions are already usable in constant expressions.
libstdc++ uses C functions for std::char_traits<wchar_t>
; libc++ uses C functions for std::char_traits<char>
and std::char_traits<wchar_t>
.
Implementations of standard library, C library and compilers may differ a lot, but according to the performance measures from "Appendix" following conclusions could be made for eglibc-2.19:
char
related intrinsics outperform C callschar
related intrinsics significantly outperform hand-made naïve implementationwchar_t
stringsAssuming that the situation in mostly the same on different platforms, it seems reasonable to:
std::char_traits<char>
constexpr. Typically this brings performance improvement.std::char_traits<wchar_t>
or leave it non-constexpr. Two wordings are provided in this paper: one that does not mark std::char_traits<wchar_t>
functions with constexpr
and another one that does.Note: Only one of the wordings may be accepted, not both.
std::char_traits
including std::char_traits<wchar_t>
Note for editor: Add constexpr in all the 21.2.3.* [char.traits.specializations.*]:
static constexpr int compare(const char_type* s1, const char_type* s2, size_t n); static constexpr size_t length(const char_type* s); static constexpr const char_type* find(const char_type* s, size_t n, const char_type& a);
std::char_traits
except std::char_traits<wchar_t>
Note for editor: Add constexpr in all the 21.2.3.* [char.traits.specializations.*] except 21.2.3.4 [char.traits.specializations.wchar.t]:
static constexpr int compare(const char_type* s1, const char_type* s2, size_t n); static constexpr size_t length(const char_type* s); static constexpr const char_type* find(const char_type* s, size_t n, const char_type& a);
For the purposes of SG10, we recommend the feature-testing macro name __cpp_lib_constexpr_char_traits
.
Revision 0:
[N4604] C++17 CD Ballot Document. Available online at www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/n4604.pdf.
[Antony Polukhin] Proof of concept implementation. Available online at https://github.com/apolukhin/apolukhin.github.io/tree/master/papers/constexpr_char_traits/.
[LWG2232] The char_traits specializations should declare their length(), compare(), and find() members constexpr. Available online at http://www.open-std.org/jtc1/sc22/wg21/docs/lwg-active.html#2232.
Walter E. Brown provided corrections and suggestions for this proposal.
Below are results of performance measures for different implementations of char_traits. Full output of benchmark and source codes are available online at https://github.com/apolukhin/apolukhin.github.io/tree/master/papers/constexpr_char_traits/.
First column in output describes the length of each char array that was given to the trait's function. Second and third columns show the median value of 500 runs of trait's function on 10000 char arrays in microseconds. Column "Relation" is equal to the second column divided on the third column. Last column is equal to max(mean absolute diviation of first trait's function runs, mean absolute diviation of second trait's function runs).
Clang results:
TypeTrait::length String length intrinsics<char> c_calls<char> Relation +/-deviation 10 40 48 1.20 +/-0.00 20 64 67 1.05 +/-0.00 30 64 66 1.03 +/-0.00 40 64 66 1.03 +/-0.00 50 65 69 1.06 +/-0.00 500 234 233 1.00 +/-0.01 5000 285 285 1.00 +/-0.01 50000 335 333 0.99 +/-0.02 String length intrinsics<char> hand-made<char> Relation +/-deviation 10 40 68 1.70 +/-0.00 20 63 131 2.08 +/-0.00 30 64 199 3.11 +/-0.00 40 64 274 4.28 +/-0.00 50 65 309 4.75 +/-0.00 500 226 651 2.88 +/-0.01 5000 287 713 2.48 +/-0.02 50000 342 781 2.28 +/-0.02 String length hand-made<wchar_t> std::char_traits<wchar_t> Relation +/-deviation 10 67 68 1.01 +/-0.00 20 129 92 0.71 +/-0.00 30 200 118 0.59 +/-0.00 40 272 147 0.54 +/-0.00 50 307 177 0.58 +/-0.01 500 911 796 0.87 +/-0.03 5000 1108 1000 0.90 +/-0.02 50000 1742 1526 0.88 +/-0.01 TypeTrait::find String length intrinsics<char> c_calls<char> Relation +/-deviation 10 32 49 1.53 +/-0.00 20 94 86 0.91 +/-0.00 30 94 87 0.93 +/-0.00 40 98 92 0.94 +/-0.00 50 103 100 0.97 +/-0.00 500 278 278 1.00 +/-0.00 5000 308 307 1.00 +/-0.02 50000 355 356 1.00 +/-0.03 String length intrinsics<char> hand-made<char> Relation +/-deviation 10 32 85 2.66 +/-0.00 20 94 161 1.71 +/-0.00 30 94 236 2.51 +/-0.00 40 98 321 3.28 +/-0.00 50 103 363 3.52 +/-0.00 500 272 784 2.88 +/-0.01 5000 296 781 2.64 +/-0.03 50000 356 872 2.45 +/-0.04 String length hand-made<wchar_t> std::char_traits<wchar_t> Relation +/-deviation 10 86 120 1.40 +/-0.00 20 164 154 0.94 +/-0.00 30 241 198 0.82 +/-0.00 40 334 230 0.69 +/-0.00 50 377 277 0.73 +/-0.00 500 1126 937 0.83 +/-0.01 5000 1305 1102 0.84 +/-0.02 50000 1488 1305 0.88 +/-0.00 TypeTrait::find2 String length intrinsics<char> c_calls<char> Relation +/-deviation 10 46 43 0.93 +/-0.00 20 79 82 1.04 +/-0.00 30 79 82 1.04 +/-0.00 40 84 88 1.05 +/-0.00 50 94 93 0.99 +/-0.00 500 251 252 1.00 +/-0.00 5000 331 331 1.00 +/-0.02 50000 379 378 1.00 +/-0.02 String length intrinsics<char> hand-made<char> Relation +/-deviation 10 46 87 1.89 +/-0.00 20 79 161 2.04 +/-0.00 30 79 237 3.00 +/-0.00 40 84 343 4.08 +/-0.00 50 94 386 4.11 +/-0.00 500 250 784 3.14 +/-0.00 5000 323 793 2.46 +/-0.02 50000 388 886 2.28 +/-0.02 String length hand-made<wchar_t> std::char_traits<wchar_t> Relation +/-deviation 10 86 123 1.43 +/-0.00 20 163 160 0.98 +/-0.00 30 240 203 0.85 +/-0.00 40 350 234 0.67 +/-0.00 50 393 282 0.72 +/-0.00 500 1095 908 0.83 +/-0.01 5000 1248 1092 0.88 +/-0.03 50000 1476 1326 0.90 +/-0.01 TypeTrait::compare String length intrinsics<char> c_calls<char> Relation +/-deviation 10 59 40 0.68 +/-0.00 20 48 42 0.88 +/-0.00 30 51 45 0.88 +/-0.00 40 62 55 0.89 +/-0.00 50 72 61 0.85 +/-0.00 500 317 317 1.00 +/-0.00 5000 3569 3570 1.00 +/-0.00 50000 28745 28679 1.00 +/-0.00 String length intrinsics<char> hand-made<char> Relation +/-deviation 10 59 95 1.61 +/-0.00 20 48 181 3.77 +/-0.00 30 51 267 5.24 +/-0.00 40 62 355 5.73 +/-0.00 50 72 441 6.12 +/-0.00 500 317 4487 14.15 +/-0.01 5000 3572 44655 12.50 +/-0.00 50000 28724 461979 16.08 +/-0.00 String length hand-made<wchar_t> std::char_traits<wchar_t> Relation +/-deviation 10 92 64 0.70 +/-0.00 20 168 122 0.73 +/-0.00 30 248 125 0.50 +/-0.00 40 332 154 0.46 +/-0.00 50 418 172 0.41 +/-0.00 500 4407 1572 0.36 +/-0.01 5000 43248 14056 0.33 +/-0.00 50000 431685 112438 0.26 +/-0.00
GCC results:
TypeTrait::length String length intrinsics<char> c_calls<char> Relation +/-deviation 10 34 40 1.18 +/-0.00 20 56 56 1.00 +/-0.00 30 56 56 1.00 +/-0.00 40 56 56 1.00 +/-0.00 50 60 60 1.00 +/-0.00 500 238 241 1.01 +/-0.00 5000 303 304 1.00 +/-0.01 50000 348 351 1.01 +/-0.03 String length intrinsics<char> hand-made<char> Relation +/-deviation 10 34 60 1.76 +/-0.00 20 57 126 2.21 +/-0.00 30 56 194 3.46 +/-0.00 40 56 272 4.86 +/-0.00 50 60 299 4.98 +/-0.00 500 238 596 2.50 +/-0.00 5000 304 632 2.08 +/-0.01 50000 380 708 1.86 +/-0.04 String length hand-made<wchar_t> std::char_traits<wchar_t> Relation +/-deviation 10 62 79 1.27 +/-0.00 20 128 103 0.80 +/-0.00 30 190 124 0.65 +/-0.01 40 273 152 0.56 +/-0.00 50 294 180 0.61 +/-0.01 500 822 839 1.02 +/-0.00 5000 1004 1000 1.00 +/-0.02 50000 1643 1538 0.94 +/-0.01 TypeTrait::find String length intrinsics<char> c_calls<char> Relation +/-deviation 10 49 42 0.86 +/-0.00 20 70 70 1.00 +/-0.00 30 70 70 1.00 +/-0.00 40 82 77 0.94 +/-0.00 50 94 96 1.02 +/-0.00 500 284 285 1.00 +/-0.00 5000 313 311 0.99 +/-0.03 50000 362 360 0.99 +/-0.04 String length intrinsics<char> hand-made<char> Relation +/-deviation 10 49 83 1.69 +/-0.00 20 70 145 2.07 +/-0.00 30 70 212 3.03 +/-0.00 40 82 267 3.26 +/-0.00 50 94 330 3.51 +/-0.00 500 281 772 2.75 +/-0.00 5000 307 777 2.53 +/-0.04 50000 369 893 2.42 +/-0.05 String length hand-made<wchar_t> std::char_traits<wchar_t> Relation +/-deviation 10 100 174 1.74 +/-0.00 20 180 208 1.16 +/-0.00 30 262 255 0.97 +/-0.00 40 342 295 0.86 +/-0.00 50 391 347 0.89 +/-0.00 500 1119 1103 0.99 +/-0.01 5000 1300 1335 1.03 +/-0.01 50000 1542 1582 1.03 +/-0.01 TypeTrait::find2 String length intrinsics<char> c_calls<char> Relation +/-deviation 10 33 39 1.18 +/-0.00 20 77 86 1.12 +/-0.00 30 77 86 1.12 +/-0.00 40 88 91 1.03 +/-0.00 50 92 97 1.05 +/-0.00 500 272 274 1.01 +/-0.00 5000 339 337 0.99 +/-0.02 50000 389 386 0.99 +/-0.02 String length intrinsics<char> hand-made<char> Relation +/-deviation 10 33 85 2.58 +/-0.00 20 77 161 2.09 +/-0.00 30 77 242 3.14 +/-0.00 40 88 321 3.65 +/-0.00 50 92 371 4.03 +/-0.00 500 264 801 3.03 +/-0.00 5000 340 811 2.39 +/-0.02 50000 384 869 2.26 +/-0.03 String length hand-made<wchar_t> std::char_traits<wchar_t> Relation +/-deviation 10 84 116 1.38 +/-0.00 20 162 150 0.93 +/-0.00 30 240 194 0.81 +/-0.00 40 322 225 0.70 +/-0.00 50 373 274 0.73 +/-0.00 500 1073 901 0.84 +/-0.01 5000 1233 1102 0.89 +/-0.01 50000 1481 1356 0.92 +/-0.02 TypeTrait::compare String length intrinsics<char> c_calls<char> Relation +/-deviation 10 54 40 0.74 +/-0.00 20 47 47 1.00 +/-0.00 30 50 43 0.86 +/-0.00 40 54 54 1.00 +/-0.00 50 70 56 0.80 +/-0.00 500 316 318 1.01 +/-0.00 5000 3589 3503 0.98 +/-0.01 50000 28598 28577 1.00 +/-0.00 String length intrinsics<char> hand-made<char> Relation +/-deviation 10 54 95 1.76 +/-0.00 20 47 184 3.91 +/-0.00 30 50 273 5.46 +/-0.00 40 54 399 7.39 +/-0.00 50 70 476 6.80 +/-0.00 500 314 3431 10.93 +/-0.00 5000 3590 33538 9.34 +/-0.00 50000 28582 333971 11.68 +/-0.00 String length hand-made<wchar_t> std::char_traits<wchar_t> Relation +/-deviation 10 113 57 0.50 +/-0.00 20 210 121 0.58 +/-0.00 30 307 119 0.39 +/-0.00 40 444 158 0.36 +/-0.00 50 509 174 0.34 +/-0.00 500 3832 1623 0.42 +/-0.00 5000 36665 14117 0.39 +/-0.00 50000 364886 112373 0.31 +/-0.00