iwyu suggest to include std::map for std::regex

[maflcko]

Steps to reproduce:

$ cat /tmp/1.cpp 
#include <regex>
#include <string>

std::regex Reg(const std::string &r) { return std::regex{r}; }

int main() {}

$ ./bin/include-what-you-use  /tmp/1.cpp 

/tmp/1.cpp should add these lines:
#include <map>      // for map
#include <sstream>  // for basic_stringbuf<>::int_type, basic_stringbuf<>::p...
#include <vector>   // for vector

/tmp/1.cpp should remove these lines:

The full include-list for /tmp/1.cpp:
#include <map>      // for map
#include <regex>    // for regex, _NFA
#include <sstream>  // for basic_stringbuf<>::int_type, basic_stringbuf<>::p...
#include <string>   // for string
#include <vector>   // for vector
---

$ ./bin/include-what-you-use  --version 
include-what-you-use 0.18 (git:abd5d2b) based on Ubuntu clang version 14.0.0-1ubuntu1

[kimgr]

Interesting. Does it behave the same if you:

• only default-construct std::regex
• construct std::regex with a pattern, but don't copy it

?

We have some issues with template symbols being attributed to instatiation location, so I'm looking for a pattern.

Also, is this with libc++ or libstdc++?

[maflcko]

Yeah, this was with libstdc++. ./bin/include-what-you-use -stdlib=libstdc++ /tmp/1.cpp gives the same result. libc++ gives:

$ ./bin/include-what-you-use -stdlib=libc++ /tmp/1.cpp 

/tmp/1.cpp should add these lines:
#include <iosfwd>  // for string

/tmp/1.cpp should remove these lines:
- #include <string>  // lines 2-2

The full include-list for /tmp/1.cpp:
#include <iosfwd>  // for string
#include <regex>   // for regex
---

Switching to default-construction is fine:

$ ./bin/include-what-you-use -stdlib=libstdc++ /tmp/1.cpp 

(/tmp/1.cpp has correct #includes/fwd-decls)

Removing the copy keeps the bug:

$ cat /tmp/1.cpp && ./bin/include-what-you-use -stdlib=libstdc++ /tmp/1.cpp 
#include <regex>
#include <string>
int main() {std::regex{std::string{}};}

/tmp/1.cpp should add these lines:
#include <map>      // for map
#include <sstream>  // for basic_stringbuf<>::int_type, basic_stringbuf<>::p...
#include <vector>   // for vector

/tmp/1.cpp should remove these lines:

The full include-list for /tmp/1.cpp:
#include <map>      // for map
#include <regex>    // for _NFA, regex
#include <sstream>  // for basic_stringbuf<>::int_type, basic_stringbuf<>::p...
#include <string>   // for string
#include <vector>   // for vector
---

[kimgr]

Thanks! I wonder what the string ctor is doing with the map. Will look into this in more detail as time permits.

[kimgr]

I wonder what the string ctor is doing with the map. Will look into this in more detail as time permits.

I've dug deep into the regex tree of headers, and found a clue. Any non-trivial constructor for std::regex calls __detail::__compile_nfa<_Rx_traits>(...), which I think eventually ends up in regex_automaton.tcc, _StateSeq<>::_M_clone(), which in turn uses std::map in its implementation.

This is not an explanation for why include-what-you-use behaves the way it does, it just establishes the link between the std::regex constructor and std::map.

The IWYU behavior is caused by a long-standing problem with how we misinterpret types in template instantiations. All related issues are tagged template-spill .

[kimgr]

Fixed on master as of ffe6c3f.