miloyip / itoa-benchmark Goto Github PK

Hello

Please have a look at the Sourceforge project:
https://sourceforge.net/projects/itoa/files/itoa/
http://itoa.cvs.sourceforge.net/itoa/itoa/

It gives interesting performance compared to the existing benches.

On top of that:

• It returns the address of the end of the string which is useful for concatenating strings.
• It is entirely portable, no specific instructions.
• It is written in C, not in C++.

Regards

Remi

https://gist.github.com/anonymous/e78c3535e72d54208529

folly conversion has been added to source, need to update the benchmark ?

Thanks for the inspiration, I think I have improved it a bit too (~30% speedup):

https://github.com/ultimatepp/ultimatepp/blob/1d018151f105f4e739045307577e3dd0bc8b856a/uppsrc/Core/Format.cpp#L28

This function depends on undefined behavior for std::numeric_limits<int32_t>::min():

void i32toa_sse2(int32_t value, char* buffer) {
if (value < 0) {
*buffer++ = '-';
value = -value;
}
u32toa_sse2(static_cast<uint32_t>(value), buffer);
}

value = -std::numeric_limits<int32_t>::min() is not defined in C++:
http://stackoverflow.com/questions/16188263/is-signed-integer-overflow-still-undefined-behavior-in-c

I've observed g++ 4.9.1 create code that segfaults: it incorrectly enters the if (v < 10000) branch then tries to lookup gDigitsLut[d1] because d1 is too big.

Recommend:
void i32toa_sse2(int32_t value, char* buffer) {
if (value < 0) {
*buffer++ = '-';
u32toa_sse2(~static_cast<uint32_t>(value) + 1, buffer);
} else {
u32toa_sse2(static_cast<uint32_t>(value), buffer);
}
}
Same for 64.

https://code.google.com/p/protobuf/source/browse/trunk/src/google/protobuf/stubs/strutil.cc#892
64-bit is also available.

Currently the time unit is millisecond per test.
It is not very intuitive as audiences need to know the test size.
I suggest to change it to nanosecond per operation. It is more intuitive and can be compared to other operation such as dtoa().