Re: [PATCH 0/5] Optimize number of comparisons for heap/heapsort implementaion

[Kent Overstreet]

On Mon, Jan 22, 2024 at 12:55:51AM +0800, Kuan-Wei Chiu wrote:
> On Sun, Jan 21, 2024 at 11:21:06AM -0500, Kent Overstreet wrote:
> > On Sun, Jan 21, 2024 at 11:36:44PM +0800, Kuan-Wei Chiu wrote:
> > > Hello,
> > > 
> > > The existing implementations of heap/heapsort follow the conventional
> > > textbook approach, where each heapify operation requires approximately
> > > 2*log2(n) comparisons. In this series, I introduce a bottom-up variant
> > > that reduces the number of comparisons during heapify operations to
> > > approximately log2(n), while maintaining the same number of swap
> > > operations.
> > > 
> > > Thanks,
> > > Kuan-Wei
> > > 
> > > Kuan-Wei Chiu (5):
> > >   bcachefs: Optimize eytzinger0_sort() using bottom-up heapsort
> > >   bcachefs: Introduce parent function for sort_cmp_size()
> > >   bcachefs: Optimize sort_cmp_size() using bottom-up heapsort
> > >   bcachefs: Optimize number of comparisons in heap_sift_down
> > >   bcache: Optimize number of comparisons in heap_sift
> > > 
> > >  drivers/md/bcache/util.h |  23 +++++----
> > >  fs/bcachefs/util.c       | 109 ++++++++++++++++++++++++++-------------
> > >  fs/bcachefs/util.h       |  23 +++++----
> > >  3 files changed, 98 insertions(+), 57 deletions(-)
> > 
> > Good stuff
> > 
> > While we're looking at this code, we should be doing some cleanup too -
> > there's no reason for the heap code to be duplicated in bcache and
> > bcachefs anymore, and it'd also be nice to get fs/bcachefs/eytzinger.h
> > moved to include/linux and bcache converted to use it.
> > 
> > I also would not be surprised if there's another heap implementation in
> > include/linux; we'll want to check for that and if there is decide which
> > is worth keeping.
> >
> Yes, we have 'min_heap.h' in include/linux.

So that has the advantage of more readable code - functions instead of
macros - whereas my version has the type safe interface.

We could combine the two approaches, and put a type-safe interface on
top of the min_heap.h code with some small macro wrappers - see
generic-radix-tree.h for an example of how that's done.

min_heap.h has only one user though? I don't think I can quite believe
that's the only other code in the kernel using a heap, there must be
more open coded out there...