Sorting library
Olin Shivers (edited by John Cowan)
2-4. (closed)
This SRFI describes the API for a full-featured sort toolkit. The spec comes with 1200 lines of high-quality reference code: tightly written, highly commented, portable code, available for free.
What happened to SRFI 32? It's been twelve years since it was withdrawn. R7RS-large needs a sorting package, and this is a substantial one. I've made as few adjustments as practical in order to bring the specification into 2014. (In this section, "I" means John Cowan; elsewhere, "I" means Olin Shivers.)
There are only three functional changes. The most important one is that the comparison predicate now precedes the data rather than following it. Most of the SRFI 32 commenters thought that this is the better order, despite the widespread precedent, and I've made the change. The second point is that the procedures accept SRFI 114 comparators rather than predicates. The third point is that the algorithm-specific procedures have been removed from the public API, though not from the sample implementation.
Editorially, I have reorganized the text and removed the massive redundancies. In accordance with R5RS and R7RS, I have specified that procedures returning an unspecified value actually do return one value. I have removed the multiple packages, most with only a few procedures, into which SRFI 32 is divided. I have added SRFI 95 and R6RS equivalents.
The implementation is Olin's, with a few bug fixes from Scheme 48. My changes were essentially mechanical.
I have designed and written a fairly comprehensive sorting and merging toolkit. It is very portable and freely reusable.
The code includes:
The code is tightly bummed. It is clearly written, and commented in my usual voluminous style. This includes notes on porting and implementation-specific optimisations.
In SRFI 32 there were two different interfaces: "what" (simple) and "how" (detailed).
However, the difficulty with naming specific algorithms is that the local optima in the search space of algorithms changes over time. For example, the "qsort" in musl C library is actually smooth sort, not quick sort; Python and Java have switched from quick sort to timsort; some implementations of the C++ STL use introsort rather than quicksort. Having procedure names that imply e.g. quick sort, that do not actually implement quick sort, is confusing.
Therefore, only the "what" interface has been retained.
Procedures share common signatures wherever possible, to facilitate switching a given call from one procedure to another.
These procedures uniformly observe the following parameter order: the comparison predicate comes before the data to be sorted. In SRFI 32, the reverse convention was used, because it was consistent with all the implementations examined in 2002, with the sole exception of Chez Scheme. However, R6RS adopted the reverse convention, which is more consistent with other Scheme libraries that put the procedure first — the "operation currying" convention, as in for-each or map or find. This SRFI uses the R6RS order throughout.
These routines take a SRFI 114 comparator, whereas their SRFI 32 counterparts took a < predicate, or in a few cases a three-valued comparison function. These are still extracted from the comparator and used internally by the sample implementation.
A "stable" sort is one that preserves the pre-existing order of equal elements. Suppose, for example, that we sort a list of numbers by comparing their absolute values using abs-comparator (defined below). If we sort a list that contains both 3 and -3, then a stable sort is an algorithm that will not swap the order of these two elements, that is, the answer will look like ... 3 -3 ..., not ... -3 3 ....
A "stable" merge is one that reliably favors one of its data sets when equal items appear in both data sets. All merge operations in this library are stable, breaking ties between data sets in favor of the first data set — elements of the first list come before equal elements in the second list.
So, if we are merging two lists of numbers ordered by absolute value using the stable merge operation list-merge and abs-comparator compares numbers by their absolute values, then
(list-merge abs-comparator '(0 -2 4 8 -10) '(-1 3 -4 7))reliably places the 4 of the first list before the equal-comparing -4 of the second list: (0 -1 -2 4 -4 7 8 -10).
Here's a definition of abs-comparator, introduced just for examples (it is not part of this SRFI):
(define abs-comparator (make-comparator number? #t (lambda (x y) (- (abs x) (abs y))) #f)) === All vector operations accept optional subrange parameters === The vector operations specified below all take optional ''start'' and ''end'' arguments indicating a selected subrange of a vector's elements. === Required vs. allowed side-effects === `list-sort!` and `list-stable-sort!` are allowed, but not required, to alter their arguments' cons cells to construct the result list. This is consistent with the what-not-how character of the group of procedures to which they belong. The `list-delete-neighbor-dups!`, `list-merge!` and `list-sort!` procedures, on the other hand, provide specific algorithms, and, as such, explicitly commit to the use of side-effects on their input lists in order to guarantee their key algorithmic properties (e.g., linear-time operation, constant-space stack use). === Sorting lists === Note that sorting lists involves chasing pointers through memory, which can be a loser on modern machine architectures because of poor cache and page locality. Pointer ''writing'', which is what the `set-cdr!`s of a destructive list-sort algorithm do, is even worse, especially if your Scheme has a generational GC — the writes will thrash the write-barrier. Sorting vectors has inherently better locality. The reference implementation's destructive list merge and merge sort implementations are opportunistic — they avoid redundant `set-cdr!`s, and try to take long already-ordered runs of list structure as-is when doing the merges. == Specification == === Procedure naming and functionality === Almost all of the procedures described below are variants of two basic operations: sorting and merging. These procedures are consistently named by composing a set of basic lexemes to indicate what they do. ||Lexeme||Meaning ||`vector`||The procedure operates upon vectors.|| ||`list`||The procedure operates upon lists.|| ||`stable`||This lexeme indicates that the sort is a stable one.|| ||`sort`||The procedure sorts its input data set by some comparator.|| ||`merge`||The procedure merges two ordered data sets into a single ordered result.|| ||`!`||Procedures that end in `!` are allowed, and sometimes required, to reuse their input storage to construct their answer.|| === Types of parameters and return values === In the procedures specified below, * A ''lis'' parameter is a list; * A ''v'' parameter is a vector; * A ''comparator'' parameter is a [http://srfi.schemers.org/srfi-114/srfi-114.html SRFI 114] comparator. * A ''start'' parameter or ''start'' and ''end'' parameter pair are exact non-negative integers such that 0 <= ''start'' <= ''end'' <= `(vector-length `''v''`)`, where ''v'' is the related vector parameter. If not specified, they default to 0 and the length of the vector, respectively. They are interpreted to select the range [''start'', ''end''), that is, all elements from index ''start'' (inclusive) up to, but not including, index ''end''. Passing values to procedures with these parameters that do not satisfy these types is an error. If a procedure is said to return "an unspecified value", this means that nothing at all is said about what the procedure returns, except that it returns one value. === Predicates === `(list-sorted? `''comparator lis''`)` `(vector-sorted? `''comparator v'' [''start'' [ ''end'' ] ] These procedures return true iff their input list or vector is in sorted order, as determined by ''comparator''. Specifically, they return `#f` iff there is an adjacent pair ... X Y ... in the input list or vector such that Y < X in the sense of ''comparator''. The optional ''start'' and ''end'' range arguments restrict `vector-sorted?` to examining the indicated subvector. These procedures are equivalent to the SRFI 95 `sorted?` procedure when applied to lists or vectors respectively, except that they do not accept a key procedure. === General sort procedures === These procedures provide basic sorting and merging functionality suitable for general programming. The procedures are named by their semantic properties, i.e., what they do to the data (sort, stable sort, and so forth). `(list-sort ` ''comparator lis''`)` `(list-stable-sort ` ''comparator lis''`)` These procedures do not alter their inputs and are allowed to return a value that shares a common tail with a list argument. The `list-stable-sort` procedure is equivalent to the R6RS `list-sort` procedure. It is also equivalent to the SRFI 95 `sort` procedure when applied to lists, except that it does not accept a key procedure. `(list-sort! ` ''comparator lis''`)` `(list-stable-sort! ` ''comparator lis''`)` These procedures are "linear update" operators — they are allowed, but not required, to alter the cons cells of their arguments to produce their results. The `list-stable-sort!` procedure is equivalent to the SRFI 95 `sort!` procedure when applied to lists, except that it does not accept a key procedure. `(vector-sort ` ''comparator v'' [ ''start'' [ ''end'' ] ]`)` `(vector-stable-sort ` ''comparator v'' [ ''start'' [ ''end'' ] ]`)` These procedures do not alter their inputs, but allocate a fresh vector for their result, of length ''end'' - ''start''. The `vector-stable-sort` procedure with no optional arguments is equivalent to the R6RS `vector-sort` procedure. It is also equivalent to the SRFI 95 `sort` procedure when applied to vectors, except that it does not accept a key procedure. `(vector-sort! ` ''comparator v'' [ ''start'' [ ''end'' ] ]`)` `(vector-stable-sort! ` ''comparator v'' [ ''start'' [ ''end'' ] ]`)` These procedures sort their data in-place. (But note that `vector-stable-sort!` may allocate temporary storage proportional to the size of the input — I am not aware of O(n lg n) stable vector sorting algorithms that run in constant space.) They return an unspecified value. The `vector-sort!` procedure with no optional arguments is equivalent to the R6RS `vector-sort!` procedure. === Merge procedures === All four merge operations are stable: an element of the initial list ''lis1'' or vector ''v1'' will come before an equal-comparing element in the second list ''lis2'' or vector ''v2'' in the result. `(list-merge `''comparator lis1 lis2''`)` This procedure does not alter its inputs, and is allowed to return a value that shares a common tail with a list argument. This procedure is equivalent to the SRFI 95 `merge` procedure when applied to lists, except that it does not accept a key procedure. `(list-merge! `''comparator lis1 lis2''`)` This procedure makes only a single, iterative, linear-time pass over its argument lists, using `set-cdr!`s to rearrange the cells of the lists into the list that is returned — they work "in place." Hence, any cons cell appearing in the result must have originally appeared in an input. Additionally, `list-merge!` is iterative, not recursive — it can operate on arguments of arbitrary size without requiring an unbounded amount of stack space. The intent of this iterative-algorithm commitment is to allow the programmer to be sure that if, for example, `list-merge!` is asked to merge two ten-million-element lists, the operation will complete without performing some extremely (possibly twenty-million) deep recursion. This procedure is equivalent to the SRFI 95 `merge!` procedure when applied to lists, except that it does not accept a key procedure. `(vector-merge` ''comparator v1 v2'' [ ''start1'' [ ''end1'' [ ''start2'' [ ''end2'' ] ] ] ]`)` This procedure does not alter its inputs, and returns a newly allocated vector of length (''end1'' - ''start1'') + (''end2'' - ''start2''`)`. This procedure is equivalent to the SRFI 95 `merge` procedure when applied to vectors, except that it does not accept a key procedure. `(vector-merge!` ''comparator to from1 from2'' [ ''at'' [ ''start1'' [ ''end1'' [ ''start2'' [ ''end2'' ] ] ] ] ]`)` This procedure writes its result into vector ''to'', beginning at index ''at'', for indices less than ''end'' = ''start'' + (''end1'' - ''start1'') + (''end2'' - ''start2''`)`. The target subvector ''to''[''start'', ''end'') may not overlap either source subvector, ''from1''[''start1'', ''end1'') or ''from2''[''start2'', ''end2''). It returns an unspecified value. This procedure is equivalent to the SRFI 95 `merge!` procedure when applied to lists, except that it does not accept a key procedure. === Deleting duplicate neighbors === These procedures delete adjacent duplicate elements from a list or a vector, using a given element-equality procedure. The first/leftmost element of a run of equal elements is the one that survives. The list or vector is not otherwise disordered. These procedures are linear time — much faster than the O(n^2^) general duplicate-element deletion procedures that do not assume any "bunching" of elements (such as the ones provided by [http://srfi.schemers.org/srfi-1/srfi-1.html SRFI 1]). If you want to delete duplicate elements from a large list or vector, sort the elements to bring equal items together, then use one of these procedures, for a total time of O(n lg n). The equality procedure extracted from the comparator passed to these procedures is always applied as `(= x y)`, where ''x'' comes before ''y'' in the containing list or vector. `(list-delete-neighbor-dups `''comparator v''`)` This procedure does not alter its input list; its result may share storage with the input list. `(list-delete-neighbor-dups! ` ''comparator lis''`)` This procedure mutates its input list in order to construct its result. It makes only a single, iterative, linear-time pass over its argument, using `set-cdr!`s to rearrange the cells of the list into the final result — it works "in place." Hence, any cons cell appearing in the result must have originally appeared in the input. `(vector-delete-neighbor-dups ` ''comparator v'' [ ''start'' [ ''end'' ] ]`)` This procedure does not alter its input vector, but rather newly allocates and returns a vector to hold the result. `(vector-delete-neighbor-dups! ` ''comparator v'' [ ''start'' [ ''end'' ] ]`)` This procedure reuses its input vector to hold the answer, packing its answer into the index range [start,newend), where `newend` is the non-negative exact integer that is returned as its value. The vector is not altered outside the range [''start, newend''). Examples: {{{ (list-delete-neighbor-dups = '(1 1 2 7 7 7 0 -2 -2)) => (1 2 7 0 -2) (vector-delete-neighbor-dups = '#(1 1 2 7 7 7 0 -2 -2)) => #(1 2 7 0 -2) (vector-delete-neighbor-dups = '#(1 1 2 7 7 7 0 -2 -2) 3 7)) => #(7 0 -2) ;; Result left in v[3,9): (let ((v (vector 0 0 0 1 1 2 2 3 3 4 4 5 5 6 6))) (cons (vector-delete-neighbor-dups! = v 3) v)) => (9 . #(0 0 0 1 2 3 4 5 6 4 4 5 5 6 6)) }}} == Porting and optimisation == This package should be trivial to port. There are only four non-R4RS bits in the code: * Use of multiple-value return, with the R5RS `values` procedure, and the simple [http://srfi.schemers.org/srfi-8/srfi-8.html SRFI 8] `(receive (`''var'' ...`)` ''mv-exp body'' ... `)` multiple-value binding macro. * A `vector-copy` procedure. This is a tiny little procedure: `(vector-copy `''v'' [ ''start'' [ ''end'' ] ]`)`. It is part of R7RS-small. * The `let-optionals` macro from scsh to parse and default optional arguments to three routines. Again, easy to port the macro or rewrite the code to parse, default, and error check the args by hand. * An `error` procedure for complaining about bad arguments. It is part of R7RS-small. This code is tightly bummed, as far as I can go in portable Scheme. You could speed up the vector code a lot by error-checking the procedure parameters and then shifting over to fixnum-specific arithmetic and dangerous vector-indexing and vector-setting primitives. The comments in the code indicate where the initial error checks would have to be added. There are several `(quotient `''n''` 2)` calls that could be changed to a fixnum right-shift, as well, in both the list and vector code. The code is designed to enable this — each file usually exports one or two "safe" procedures that end up calling an internal "dangerous" primitive. The little exported cover procedures are where you move the error checks. This should provide ''big'' speedups. In fact, all the code bumming I've done pretty much disappears in the noise unless you have a good compiler and also can dump the vector-index checks and generic arithmetic — so I've really just set things up for you to exploit. The optional-arg parsing, defaulting, and error checking is done with a portable R4RS macro. But if your Scheme has a faster mechanism (e.g., Chez), you should definitely port over to it. Note that argument defaulting and error-checking are interleaved — you don't have to error-check defaulted ''start'' and ''end'' args to see if they are fixnums that are legal vector indices for the corresponding vector, etc. == Acknowledgements == I thank the authors of the open source I consulted when designing this library, particularly Richard O'Keefe, Donovan Kolbly and the MIT Scheme Team. == Copyright == === SRFI text === This document is copyright (C) Olin Shivers (1998, 1999). All Rights Reserved. Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. === Sample implementation === Short summary: no restrictions. While I wrote all of this code myself, I read a lot of code before I began writing. However, all such code is, itself, either open source or public domain, rendering irrelevant any issue of "copyright taint." The natural merge sorts (pure list, destructive list, and vector) are not only my own code, but are implementations of an algorithm of my own devising. They run in O(n lg n) worst case, O(n) best case, and require only a logarithmic number of stack frames. And they are stable. And the destructive-list variant allocates zero cons cells; it simply rearranges the cells of the input list. Hence the sample implementation is Copyright © 1998 by Olin Shivers and made available under the same copyright as the SRFI text (see above).