|
|
|
|
|
|
by maleadt
4109 days ago
|
|
|
A colleague of mine mentioned how this has been the case for quite a bit longer than Haswell, see "Compiler mitigations for time attacks on modern x86 processors" (from 2012). Or Agner Fog's timings pages, where many DIV instructions list widely ranging latencies, with the comment "[l]ow values are for small results, high values for high results". |
|
|
What I remembered having seen advertised (presented as an improvement, of course) is new early-exit paths for the easier arguments to division. The marketing brochure emphasised the novelty, so I must have wrongly inferred that division took constant time before that.
Considering paragraph 2.2 in http://users.elis.ugent.be/~brdsutte/research/publications/2... , I may have been remembering a new division algorithm in Nehalem. I will cite that article from now on.