On Fri, Sep 25, 2020 at 12:02:38AM +0200, René Scharfe wrote: > Older versions of gcc and clang didn't see through the shifting > put_be32() implementation. If you go back further there are also > versions that didn't optimize the shifting get_be32(). And the latest > icc still can't do that. > > gcc 10.2 just optimizes all functions to a bswap and a mov. Can't do > any better than that, can you? > > But why do we then see a difference in our benchmark results? Not sure, > but https://www.godbolt.org/z/7xh8ao shows that gcc is shuffling some > instructions around depending on the implementation. Switch to clang if > you want to see more vigorous shuffling. We do redefine ntohl(), etc in compat/bswap.h. Looking at them, I'd think the result would end up as a bswap instruction, though. And indeed, trying to feed that to godbolt results in the same output you got. It does sound like older compilers were benefiting from the unaligned versions. Building with gcc-4.8 (from debian jessie in a docker container on the same machine), I get ~6.25s with the unaligned load versus ~6.6s with the bit-shifting code. So that's the opposite of how the newer compiler behaves. Benchmarking clang-8 (which your results showed doesn't handle the shifting version well). It likewise is just _slightly_ slower after my patch (11.47s versus 11.57s). Given that newer compilers behave the opposite way, and the overall small magnitude of the impact, I'm still comfortable with the change. It's nice to have a better understanding of how the compiler is impacting it (even if I am still confused how anything at all changes on the newer compilers). Thanks for digging. -Peff
