Fix assertion failure in datetime.stopwatch when compiling with -O2

[ibuclaw]

The benchmark could take no time at all because the benchmark is inlined, and the loop is removed.

[PetarKirov]

@ibuclaw can you help forge upstream an intrinsic (or a set of) for use in std.benchmark which would prevent similar compiler optimizations? Would the GDC syntax equivalent of the inline assembly shown in the CppCon 2015 talk https://stackoverflow.com/questions/33975479/escape-and-clobber-equivalent-in-msvc Just Work tm?

[ibuclaw]

In this case its not so much a problem that the first benchmark takes no time, the unittest here is only making sure that the second function took longer.

[jpf91]

For this specific case it probably doesn't matter, but I think std.benchmark should prevent such optimizations in general. So we should have a look at the solutions presented in that talk.

[jpf91]

OK, but why version(GNU)? LDC could be affected as well, or maybe will be in the future.

[ibuclaw]

@jpf91 - Oh, I'm just going to push the >= condition upstream, so when we (eventually) merge it in, the version condition can be removed.

Edit, actually, I was beaten. Upstream PR is here: dlang/phobos#6265

[ibuclaw]

@ZombineDev - As for the assembler, it should work to have:

void escape(void* p)
{
    asm { "" : : "g"(p) : "memory"; }
}

void clobber()
{
    asm { "" : : : "memory"; }
}

All asm statements are marked volatile, unless they are asm pure { }.

[ibuclaw]

Granted though, I don't think there's anything peculiar about these primitives that they can't just be intrinsics also.

[wilzbach]

For this specific case it probably doesn't matter, but I think std.benchmark should prevent such optimizations in general.

See also: dlang/phobos#5416