Open orlp opened 5 years ago
I would like to re-open the discussion with the following observations, in the hope that @alexcrichton might change his mind:
To clarify, all previous cases are purely -O0 optimizations. If this doesn't optimize at higher optimization levels then we can of course change the code to fix that. If this is only a -O0 optimization then that's different.
@alexcrichton I just tested it and the Vec::truncate
operation also doesn't optimize away on opt-level=1
, only 2 and higher. So even that portion isn't purely -O0.
But even for -O0 I think we shouldn't leave O(n) -> O(1) reductions on the table when they cost only a couple lines of code. Debug performance matters.
+1: having a non-guaranteed O(1) truncate
pretty much defeats the purpose of having a deque in many cases. And I agree that a truncate_front
(O(1) as well) would come very handy to have built-in for VecDeque.
I was just bitten by this. Had assumed that truncate()
was an O(1) op for Vec
+1 for having a truncate_front
as well.
And somewhat similarly, it would be useful to have a prepend
method as the symmetric version of append
, as well as an ExtendFront
as a symmetric version of std::iter::Extend
. I'm not sure whether it makes sense to track those as separate issues.
however this does not happen for VecDeque::truncate.
That seems to be fixed since 1.41
Currently both
Vec
s andVecDeque
struncate
implementations just callpop_back()
in a loop. Whenstd::mem::needs_drop::<T>()
isfalse
this is pretty inefficient.Now LLVM does optimize away the loop on
-O3
forVec::truncate
, however this does not happen forVecDeque::truncate
.Despite previous complaints, this exact issue has been wont-fix'd under the unofficial policy that we do not include -O0 exclusive optimizations in rust.
I would like to re-open the discussion with the following observations, in the hope that @alexcrichton might change his mind:
No matter the optimization level, the optimization never happens for
VecDeque::truncate
.The optimization is not just a constant factor speedup, it's an asymptotic speedup changing an O(n) time operation into O(1) where O(1) could easily be guaranteed. The difference in speed could literally be several orders of magnitude.
The optimizer is always fickle and ever-changing. Someone designing an algorithm using default Rust building blocks needs to rely on at least asymptotic complexities not suddenly regressing due to the optimizer not catching some pattern. Updating your Rust compiler should never have a chance to turn an O(n) algorithm into an O(n^2) one.
It pushes people towards unsafe code. In the linked complaint above the user ended up using the unsafe
set_len
when the safetruncate
could've exactly done the job.In addition to the above,
VecDeque
hastruncate
, but is missingtruncate_front
. If we make these truncate operations O(1) time it should also get atruncate_front
operation.Without these changes it is impossible to implement certain algorithms using the default Rust collections for no particular good reason. For example, if I have a
v: VecDeque<f64>
that keeps its elements sorted, it is currently impossible to write a O(log n) time routine that removes all elements less thanx
inv
. We can find how many elements we need to remove from the front in O(log n) time, but removing them needlessly take O(n) time.