Hello Leonidas, by using the PKGBUILD directly it should work in all cases. The problem was really that rmlint was built on a system that has SSE 4.2 but was shipped to one as binary where it is not available. The build systems actually used to check the presence of a SSE 4.2 capable cpu, but that only works when you build it yourself. The auto detection is still there, but has to be enabled using `--with-sse` explicitly now (not recommended for packagers). It's probably best to remove it alltogether since rmlint also supports xxhash which looks a bit more stable. After all, SSE 4.2 support for rmlint is really very unimportant. Only one hash function has an option to use it (CityHash, not even the default) and I used to think "Well, why not pick it up if it's there". I don't have any metrics myself of how fast the normal variant is compared to the SSE variant, apparently I only own AMD/ARM hardware currently. Am Mi, 14. Okt, 2015 um 9:29 schrieb Leonidas Spyropoulos <artafinde@gmail.com>:
Hi Christopher,
On 14/10/15, christopher@ira-kunststoffe.de wrote:
Sorry for the inconvinience, it still crashes with "illegal instruction". This is most likely due to the fact that I didn't actually make --without-sse the default. Should be (hopefully!) corrected in v2.2.2. I shouldn't push changes late night. :)
Just tried it on AMD FX 8210 from the package and had the same Illegal hardware and code dump. I dowload the PKGBUILD of the package and bumped the version to 2.2.2. Compiled and worked.
Can't you make the code use the hardware SSE 4.2 CRC32C if the hardware is there or use a software implementation of it if not?
What's the difference in performance with and without SSE4.2 enabled - can we have some metrics?
-- Leonidas Spyropoulos
Compiled in Vim, sent through mutt.