News for hackers. Stuff that matters.
Re:Getting 4 bytes or a full cache line: same speed or not? by Const-me on Thursday August 02, 2018 @12:30PM (#17672793)
> you are comparing now not 1 iteration with 16 iteration

GCC made code with zero iterations, it unrolled the complete inner loop and produced 16 scalar add instructions.

MSVC made code with two iterations, each iteration containing two paddd SSE2 instructions, which stand for _mm_add_epi32.

> Turn off vectorization of code in msvc

I can’t. When I try, the compiler says “Command line warning D9002: ignoring unknown option '/arch:IA32'”

SSE2 is a requirement for AMD64 CPUs, that’s why the MS compiler doesn’t support scalar code anymore. And the data size is too large for 32 bit builds, it’s not enough address space for a 8GB vector.

Re:Getting 4 bytes or a full cache line: same speed or not? by ordu on Thursday August 02, 2018 @12:50PM (#17672999)
> GCC made code with zero iterations, it unrolled the complete inner loop and produced 16 scalar add instructions.

It doesn't matter really. It is amount of work that needed for experiment.

> I can’t. When I try, the compiler says “Command line warning D9002: ignoring unknown option '/arch:IA32'”

I have zero experience with msvc, but I believe you are doing it wrong. You need not to change arch, you need to control over optimizations. For example, gcc has options -ftree-vectorize, -ftree-loop-vectorize, and -fno-tree-vectorize and so on. You need to find docs on compiler optimization techniques and on options that allow to control them.

> MS compiler doesn’t support scalar code anymore.

Strange, but ok. If we return to the starting question "why the author didn't write the code portably", than we now know the answer. There are no way to write scalar code with MSVC. MSVC doesn't allow make this kind of experiments, so there are no reasons to make such tests portable to msvc.

Re:Getting 4 bytes or a full cache line: same speed or not? by Const-me on Thursday August 02, 2018 @01:48PM (#17673641)
> You need not to change arch

See the pics: http://const.me/tmp/vc/codegen.png http://const.me/tmp/vc/optimize.png

> You need to find docs on compiler optimization techniques

Indeed, #pragma loop(no_vector) in the source code before the inner loop did the trick. The results are very close to GCC.