Comment

@Oleg:

I removed my macros:
#define __tzcnt_u32(v) ((v) ? (_tzcnt_u32(v)) : (32))
#define __lzcnt32(v) ((v) ? (_lzcnt_u32(v)) : (32))
replacing them with simple inline code


Util.h

//#include // not including all
#include // just include what's needed
#include // just include what's needed
#include // just include what's needed

#if 1 // use safe localtime
struct tm buf; // use safe localtime
auto err = localtime_s(&buf, &cur_time); // use safe localtime
return std::put_time(&buf, "%F %T"); // use safe localtime
#else // use safe localtime
return std::put_time(std::localtime(&cur_time), "%F %T");
#endif // use safe localtime


State.h

changed _mm_set_epi8(0x80 to -0x80 to stop warnings

inline replacement:
//int i = __tzcnt_u32(mask); // for no BMI; without zero test, as not needed here
int i = _tzcnt_u32(mask); // for no BMI; without zero test, as not needed here

inline replacement:
//int last_idx_before_mid = 31 - __lzcnt32(off_mask); // for no BMI; without zero test, as not needed here
int last_idx_before_mid = _lzcnt_u32(off_mask); // for no BMI; without zero test, as not needed here


Solver.h

inline replacement:
//return ini.size(); // to stop warning
return (int)ini.size(); // to stop warning

inline replacement:
//const int dim = __tzcnt_u32(mask); // for no BMI; without zero test, as not needed here
const int dim = _tzcnt_u32(mask); // for no BMI; without zero test, as not needed here


I tried '9' runs: with asserts: 10:31, without: 10:18 (saved 2%)
A minute slower than the faulty version, but still not too bad for a 2013 (Q3) CPU :)