On Mon, Apr 3, 2023 at 8:03 AM Lennart Sorensen <lsore...@csclub.uwaterloo.ca> wrote: > > On Sun, Apr 02, 2023 at 09:51:23PM -0400, Alan Corey wrote: > > I know I can but it will be twice as slow, which is why I want armhf. > > Under 64 bit both the data and pointers will be twice as big. With > > unlimited memory that would be OK but a Pi CPU can only access 1 GB. > > I've tried 64 bit. > > It's certainly a balance trade off. The pointers will be twice as large. > The data will be whatever size the code asked for. Only in the case that > the code asked to use a long will be be 32 bit in one case and 64 bit > in the other case. Most code isn't that sloppy about their data types. > > In terms of actual code, apparently the A53 core runs 64 bit code about > 20% faster than 32 bit code, so it comes down to whether you are doing > execution heavy or data heavy work.
There's one exception to 64-bit speedups that I am aware of. On A-57's in 64 bit mode, ASIMD (formerly NEON) shifts used to be awful slow. It was measurable. I'm not sure if it is still the case. See, for example, http://github.com/weidai11/cryptopp/issues/367 . It is better to use the integer units instead of the SIMD unit for shifts in 64-bit mode. Jeff