Files
why3/examples/fenwick.mlw
2018-06-15 16:45:58 +02:00

121 lines
4.3 KiB
Plaintext

(* Fenwick trees (or binary indexed tree) for prefix/interval sums.
Represent an integer array over interval [0;n[ such that the following
operations are both efficient:
. incrementation of individual cell (O(log(n)))
. Query sum of elements over interval [a;b[ (O(log(b-a)))
Author: Martin Clochard (Université Paris Sud) *)
(* Array-based implementation with i->(2i+1,2i+2) node encoding:
. Integer represent nodes
. 0 is the root
. childs of node n are 2n+1 and 2n+2
. Leaves represent the model array cells. For n > 0 elements in the model,
they are represented by the cells over the range [n-1;2n-1[
The structure manage queries by keeping for each node the sum of the
values of all descendant leaves, which we call here 'node summary' *)
module Fenwick
use int.Int
use int.ComputerDivision
use ref.Ref
use int.Sum
use array.Array
(* Encode fenwick trees within an array. The leaves field represent
the actual number of element within the model. *)
type fenwick = {
t : array int;
ghost leaves : int;
}
(* Invariant *)
predicate valid (f:fenwick) =
f.leaves >= 0 /\
f.t.length = (if f.leaves = 0 then 0 else 2 * f.leaves - 1) /\
forall i. 0 <= i /\ i < f.leaves - 1 ->
f.t[i] = f.t[2*i+1] + f.t[2*i+2]
(* Get the i-th elements of the model. *)
function get (f:fenwick) (i:int) : int = f.t[i+f.leaves-1]
(* Get the sum of elements over range [a;b[ *)
function rget (f:fenwick) (a b:int) : int = sum (get f) a b
(* Create a Fenwick tree initialized at 0 *)
let make (lv:int) : fenwick
requires { lv >= 0 }
ensures { valid result }
ensures { forall i. 0 <= i < lv -> get result i = 0 }
ensures { result.leaves = lv }
= { t = if lv = 0 then make 0 0 else make (2*lv-1) 0;
leaves = lv }
(* Add x to the l-th cell *)
let add (f:fenwick) (l:int) (x:int) : unit
requires { 0 <= l < f.leaves /\ valid f }
ensures { valid f }
ensures { forall i. 0 <= i < f.leaves /\ i <> l ->
get f i = get (old f) i }
ensures { get f l = get (old f) l + x }
= let lc = ref (div f.t.length 2 + l) in
f.t[!lc] <- f.t[!lc] + x;
(* Update node summaries for all elements on the path
from the updated leaf to the root. *)
label I in
while !lc <> 0 do
invariant { 0 <= !lc < f.t.length }
invariant { forall i. 0 <= i /\ i < f.leaves - 1 ->
f.t[i] = f.t[2*i+1] + f.t[2*i+2] -
if 2*i+1 <= !lc <= 2*i+2 then x else 0 }
invariant { forall i. f.leaves - 1 <= i < f.t.length ->
f.t[i] = (f at I).t[i] }
variant { !lc }
lc := div (!lc - 1) 2;
f.t[!lc] <- f.t[!lc] + x
done
(* Lemma to shift dum indices. *)
let rec ghost sum_dec (a b c:int) : unit
requires { a <= b }
ensures { forall f g. (forall i. a <= i < b -> f i = g (i+c)) ->
sum f a b = sum g (a+c) (b+c) }
variant { b - a }
= if a < b then sum_dec (a+1) b c
(* Crucial lemma for the query routine: Summing the node summaries
over range [2a+1;2b+1[ is equivalent to summing node summaries
over range [a;b[. This is because the elements of range [2a+1;2b+1[
are exactly the childs of elements of range [a;b[. *)
let rec ghost fen_compact (f:fenwick) (a b:int) : unit
requires { 0 <= a <= b /\ 2 * b < f.t.length /\ valid f }
ensures { sum (([]) f.t) a b = sum (([]) f.t) (2*a+1) (2*b+1) }
variant { b - a }
= if a < b then fen_compact f (a+1) b
(* Query sum of elements over interval [a,b[. *)
let query (f:fenwick) (a b:int) : int
requires { 0 <= a <= b <= f.leaves /\ valid f }
ensures { result = rget f a b }
= let lv = div f.t.length 2 in
let ra = ref (a + lv) in let rb = ref (b + lv) in
let acc = ref 0 in
ghost sum_dec a b lv;
(* If ra = rb, the sum is 0.
Otherwise, adjust the range to odd boundaries in constant time
and use compaction lemma to halve interval size. *)
while !ra <> !rb do
invariant { 0 <= !ra <= !rb <= f.t.length }
invariant { !acc + sum (([]) f.t) !ra !rb = rget f a b }
variant { !rb - !ra }
if mod !ra 2 = 0 then acc := !acc + f.t[!ra];
ra := div !ra 2;
rb := !rb - 1;
if mod !rb 2 <> 0 then acc := !acc + f.t[!rb];
rb := div !rb 2;
ghost fen_compact f !ra !rb
done;
!acc
end