* Midpoint obsession
@ 2020-09-06 16:24 namark
0 siblings, 0 replies; only message in thread
From: namark @ 2020-09-06 16:24 UTC (permalink / raw)
To: libstdc++
Hello, what do you think about this midpoint implementation for integers?
https://git.sr.ht/~namark/libsimple_support/tree/a35dc5f34bac6139cd8394311ef9d7dbffdd02b0/source/simple/support/algorithm.hpp#L361-371
The best way I can explain it is 2's complement with the carry as a sign and the unsigned integer as the lower bits. For the addition and subtraction the sign doesn't matter as it doesn't affect the lower bits (which we know can hold the result). The division on the other hand is sign sensitive, so in case of negative difference between a and b it is done in three steps (negation -> division -> negation).
Even though I wrote it with conditional move in mind, something about it makes gcc prefer to generate a branch on x86, and usually I wouldn't argue, but to validate my efforts I asked for -fno-guess-branch-probability, which proved it can produce shorter(better?) branchless assembly as well. Also got slight shortening on armv7-a. I believe the code doesn't make any more assumptions than the current implementation, though I haven't tested it on other platforms.
Any feedback is appreciated, and I'm happy to contribute if you think this is a meaningful improvement.
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2020-09-06 16:24 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-09-06 16:24 Midpoint obsession namark
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).