1Intro
2-----
3This describes an adaptive, stable, natural mergesort, modestly called
4timsort (hey, I earned it <wink>).  It has supernatural performance on many
5kinds of partially ordered arrays (less than lg(N!) comparisons needed, and
6as few as N-1), yet as fast as Python's previous highly tuned samplesort
7hybrid on random arrays.
8
9In a nutshell, the main routine marches over the array once, left to right,
10alternately identifying the next run, then merging it into the previous
11runs "intelligently".  Everything else is complication for speed, and some
12hard-won measure of memory efficiency.
13
14
15Comparison with Python's Samplesort Hybrid
16------------------------------------------
17+ timsort can require a temp array containing as many as N//2 pointers,
18  which means as many as 2*N extra bytes on 32-bit boxes.  It can be
19  expected to require a temp array this large when sorting random data; on
20  data with significant structure, it may get away without using any extra
21  heap memory.  This appears to be the strongest argument against it, but
22  compared to the size of an object, 2 temp bytes worst-case (also expected-
23  case for random data) doesn't scare me much.
24
25  It turns out that Perl is moving to a stable mergesort, and the code for
26  that appears always to require a temp array with room for at least N
27  pointers. (Note that I wouldn't want to do that even if space weren't an
28  issue; I believe its efforts at memory frugality also save timsort
29  significant pointer-copying costs, and allow it to have a smaller working
30  set.)
31
32+ Across about four hours of generating random arrays, and sorting them
33  under both methods, samplesort required about 1.5% more comparisons
34  (the program is at the end of this file).
35
36+ In real life, this may be faster or slower on random arrays than
37  samplesort was, depending on platform quirks.  Since it does fewer
38  comparisons on average, it can be expected to do better the more
39  expensive a comparison function is.  OTOH, it does more data movement
40  (pointer copying) than samplesort, and that may negate its small
41  comparison advantage (depending on platform quirks) unless comparison
42  is very expensive.
43
44+ On arrays with many kinds of pre-existing order, this blows samplesort out
45  of the water.  It's significantly faster than samplesort even on some
46  cases samplesort was special-casing the snot out of.  I believe that lists
47  very often do have exploitable partial order in real life, and this is the
48  strongest argument in favor of timsort (indeed, samplesort's special cases
49  for extreme partial order are appreciated by real users, and timsort goes
50  much deeper than those, in particular naturally covering every case where
51  someone has suggested "and it would be cool if list.sort() had a special
52  case for this too ... and for that ...").
53
54+ Here are exact comparison counts across all the tests in sortperf.py,
55  when run with arguments "15 20 1".
56
57  Column Key:
58      *sort: random data
59      \sort: descending data
60      /sort: ascending data
61      3sort: ascending, then 3 random exchanges
62      +sort: ascending, then 10 random at the end
63      %sort: ascending, then randomly replace 1% of elements w/ random values
64      ~sort: many duplicates
65      =sort: all equal
66      !sort: worst case scenario
67
68  First the trivial cases, trivial for samplesort because it special-cased
69  them, and trivial for timsort because it naturally works on runs.  Within
70  an "n" block, the first line gives the # of compares done by samplesort,
71  the second line by timsort, and the third line is the percentage by
72  which the samplesort count exceeds the timsort count:
73
74      n   \sort   /sort   =sort
75-------  ------  ------  ------
76  32768   32768   32767   32767  samplesort
77          32767   32767   32767  timsort
78          0.00%   0.00%   0.00%  (samplesort - timsort) / timsort
79
80  65536   65536   65535   65535
81          65535   65535   65535
82          0.00%   0.00%   0.00%
83
84 131072  131072  131071  131071
85         131071  131071  131071
86          0.00%   0.00%   0.00%
87
88 262144  262144  262143  262143
89         262143  262143  262143
90          0.00%   0.00%   0.00%
91
92 524288  524288  524287  524287
93         524287  524287  524287
94          0.00%   0.00%   0.00%
95
961048576 1048576 1048575 1048575
97        1048575 1048575 1048575
98          0.00%   0.00%   0.00%
99
100  The algorithms are effectively identical in these cases, except that
101  timsort does one less compare in \sort.
102
103  Now for the more interesting cases.  Where lg(x) is the logarithm of x to
104  the base 2 (e.g., lg(8)=3), lg(n!) is the information-theoretic limit for
105  the best any comparison-based sorting algorithm can do on average (across
106  all permutations).  When a method gets significantly below that, it's
107  either astronomically lucky, or is finding exploitable structure in the
108  data.
109
110
111      n   lg(n!)    *sort    3sort     +sort   %sort    ~sort     !sort
112-------  -------   ------   -------  -------  ------  -------  --------
113  32768   444255   453096   453614    32908   452871   130491    469141 old
114                   448885    33016    33007    50426   182083     65534 new
115                    0.94% 1273.92%   -0.30%  798.09%  -28.33%   615.87% %ch from new
116
117  65536   954037   972699   981940    65686   973104   260029   1004607
118                   962991    65821    65808   101667   364341    131070
119                    1.01% 1391.83%   -0.19%  857.15%  -28.63%   666.47%
120
121 131072  2039137  2101881  2091491   131232  2092894   554790   2161379
122                  2057533   131410   131361   206193   728871    262142
123                    2.16% 1491.58%   -0.10%  915.02%  -23.88%   724.51%
124
125 262144  4340409  4464460  4403233   262314  4445884  1107842   4584560
126                  4377402   262437   262459   416347  1457945    524286
127                    1.99% 1577.82%   -0.06%  967.83%  -24.01%   774.44%
128
129 524288  9205096  9453356  9408463   524468  9441930  2218577   9692015
130                  9278734   524580   524633   837947  2916107   1048574
131                   1.88%  1693.52%   -0.03% 1026.79%  -23.92%   824.30%
132
1331048576 19458756 19950272 19838588  1048766 19912134  4430649  20434212
134                 19606028  1048958  1048941  1694896  5832445   2097150
135                    1.76% 1791.27%   -0.02% 1074.83%  -24.03%   874.38%
136
137  Discussion of cases:
138
139  *sort:  There's no structure in random data to exploit, so the theoretical
140  limit is lg(n!).  Both methods get close to that, and timsort is hugging
141  it (indeed, in a *marginal* sense, it's a spectacular improvement --
142  there's only about 1% left before hitting the wall, and timsort knows
143  darned well it's doing compares that won't pay on random data -- but so
144  does the samplesort hybrid).  For contrast, Hoare's original random-pivot
145  quicksort does about 39% more compares than the limit, and the median-of-3
146  variant about 19% more.
147
148  3sort, %sort, and !sort:  No contest; there's structure in this data, but
149  not of the specific kinds samplesort special-cases.  Note that structure
150  in !sort wasn't put there on purpose -- it was crafted as a worst case for
151  a previous quicksort implementation.  That timsort nails it came as a
152  surprise to me (although it's obvious in retrospect).
153
154  +sort:  samplesort special-cases this data, and does a few less compares
155  than timsort.  However, timsort runs this case significantly faster on all
156  boxes we have timings for, because timsort is in the business of merging
157  runs efficiently, while samplesort does much more data movement in this
158  (for it) special case.
159
160  ~sort:  samplesort's special cases for large masses of equal elements are
161  extremely effective on ~sort's specific data pattern, and timsort just
162  isn't going to get close to that, despite that it's clearly getting a
163  great deal of benefit out of the duplicates (the # of compares is much less
164  than lg(n!)).  ~sort has a perfectly uniform distribution of just 4
165  distinct values, and as the distribution gets more skewed, samplesort's
166  equal-element gimmicks become less effective, while timsort's adaptive
167  strategies find more to exploit; in a database supplied by Kevin Altis, a
168  sort on its highly skewed "on which stock exchange does this company's
169  stock trade?" field ran over twice as fast under timsort.
170
171  However, despite that timsort does many more comparisons on ~sort, and
172  that on several platforms ~sort runs highly significantly slower under
173  timsort, on other platforms ~sort runs highly significantly faster under
174  timsort.  No other kind of data has shown this wild x-platform behavior,
175  and we don't have an explanation for it.  The only thing I can think of
176  that could transform what "should be" highly significant slowdowns into
177  highly significant speedups on some boxes are catastrophic cache effects
178  in samplesort.
179
180  But timsort "should be" slower than samplesort on ~sort, so it's hard
181  to count that it isn't on some boxes as a strike against it <wink>.
182
183+ Here's the highwater mark for the number of heap-based temp slots (4
184  bytes each on this box) needed by each test, again with arguments
185  "15 20 1":
186
187   2**i  *sort \sort /sort  3sort  +sort  %sort  ~sort  =sort  !sort
188  32768  16384     0     0   6256      0  10821  12288      0  16383
189  65536  32766     0     0  21652      0  31276  24576      0  32767
190 131072  65534     0     0  17258      0  58112  49152      0  65535
191 262144 131072     0     0  35660      0 123561  98304      0 131071
192 524288 262142     0     0  31302      0 212057 196608      0 262143
1931048576 524286     0     0 312438      0 484942 393216      0 524287
194
195  Discussion:  The tests that end up doing (close to) perfectly balanced
196  merges (*sort, !sort) need all N//2 temp slots (or almost all).  ~sort
197  also ends up doing balanced merges, but systematically benefits a lot from
198  the preliminary pre-merge searches described under "Merge Memory" later.
199  %sort approaches having a balanced merge at the end because the random
200  selection of elements to replace is expected to produce an out-of-order
201  element near the midpoint.  \sort, /sort, =sort are the trivial one-run
202  cases, needing no merging at all.  +sort ends up having one very long run
203  and one very short, and so gets all the temp space it needs from the small
204  temparray member of the MergeState struct (note that the same would be
205  true if the new random elements were prefixed to the sorted list instead,
206  but not if they appeared "in the middle").  3sort approaches N//3 temp
207  slots twice, but the run lengths that remain after 3 random exchanges
208  clearly has very high variance.
209
210
211A detailed description of timsort follows.
212
213Runs
214----
215count_run() returns the # of elements in the next run.  A run is either
216"ascending", which means non-decreasing:
217
218    a0 <= a1 <= a2 <= ...
219
220or "descending", which means strictly decreasing:
221
222    a0 > a1 > a2 > ...
223
224Note that a run is always at least 2 long, unless we start at the array's
225last element.
226
227The definition of descending is strict, because the main routine reverses
228a descending run in-place, transforming a descending run into an ascending
229run.  Reversal is done via the obvious fast "swap elements starting at each
230end, and converge at the middle" method, and that can violate stability if
231the slice contains any equal elements.  Using a strict definition of
232descending ensures that a descending run contains distinct elements.
233
234If an array is random, it's very unlikely we'll see long runs.  If a natural
235run contains less than minrun elements (see next section), the main loop
236artificially boosts it to minrun elements, via a stable binary insertion sort
237applied to the right number of array elements following the short natural
238run.  In a random array, *all* runs are likely to be minrun long as a
239result.  This has two primary good effects:
240
2411. Random data strongly tends then toward perfectly balanced (both runs have
242   the same length) merges, which is the most efficient way to proceed when
243   data is random.
244
2452. Because runs are never very short, the rest of the code doesn't make
246   heroic efforts to shave a few cycles off per-merge overheads.  For
247   example, reasonable use of function calls is made, rather than trying to
248   inline everything.  Since there are no more than N/minrun runs to begin
249   with, a few "extra" function calls per merge is barely measurable.
250
251
252Computing minrun
253----------------
254If N < 64, minrun is N.  IOW, binary insertion sort is used for the whole
255array then; it's hard to beat that given the overheads of trying something
256fancier (see note BINSORT).
257
258When N is a power of 2, testing on random data showed that minrun values of
25916, 32, 64 and 128 worked about equally well.  At 256 the data-movement cost
260in binary insertion sort clearly hurt, and at 8 the increase in the number
261of function calls clearly hurt.  Picking *some* power of 2 is important
262here, so that the merges end up perfectly balanced (see next section).  We
263pick 32 as a good value in the sweet range; picking a value at the low end
264allows the adaptive gimmicks more opportunity to exploit shorter natural
265runs.
266
267Because sortperf.py only tries powers of 2, it took a long time to notice
268that 32 isn't a good choice for the general case!  Consider N=2112:
269
270>>> divmod(2112, 32)
271(66, 0)
272>>>
273
274If the data is randomly ordered, we're very likely to end up with 66 runs
275each of length 32.  The first 64 of these trigger a sequence of perfectly
276balanced merges (see next section), leaving runs of lengths 2048 and 64 to
277merge at the end.  The adaptive gimmicks can do that with fewer than 2048+64
278compares, but it's still more compares than necessary, and-- mergesort's
279bugaboo relative to samplesort --a lot more data movement (O(N) copies just
280to get 64 elements into place).
281
282If we take minrun=33 in this case, then we're very likely to end up with 64
283runs each of length 33, and then all merges are perfectly balanced.  Better!
284
285What we want to avoid is picking minrun such that in
286
287    q, r = divmod(N, minrun)
288
289q is a power of 2 and r>0 (then the last merge only gets r elements into
290place, and r < minrun is small compared to N), or q a little larger than a
291power of 2 regardless of r (then we've got a case similar to "2112", again
292leaving too little work for the last merge to do).
293
294Instead we pick a minrun in range(32, 65) such that N/minrun is exactly a
295power of 2, or if that isn't possible, is close to, but strictly less than,
296a power of 2.  This is easier to do than it may sound:  take the first 6
297bits of N, and add 1 if any of the remaining bits are set.  In fact, that
298rule covers every case in this section, including small N and exact powers
299of 2; merge_compute_minrun() is a deceptively simple function.
300
301
302The Merge Pattern
303-----------------
304In order to exploit regularities in the data, we're merging on natural
305run lengths, and they can become wildly unbalanced.  That's a Good Thing
306for this sort!  It means we have to find a way to manage an assortment of
307potentially very different run lengths, though.
308
309Stability constrains permissible merging patterns.  For example, if we have
3103 consecutive runs of lengths
311
312    A:10000  B:20000  C:10000
313
314we dare not merge A with C first, because if A, B and C happen to contain
315a common element, it would get out of order wrt its occurrence(s) in B.  The
316merging must be done as (A+B)+C or A+(B+C) instead.
317
318So merging is always done on two consecutive runs at a time, and in-place,
319although this may require some temp memory (more on that later).
320
321When a run is identified, its length is passed to found_new_run() to
322potentially merge runs on a stack of pending runs.  We would like to delay
323merging as long as possible in order to exploit patterns that may come up
324later, but we like even more to do merging as soon as possible to exploit
325that the run just found is still high in the memory hierarchy.  We also can't
326delay merging "too long" because it consumes memory to remember the runs that
327are still unmerged, and the stack has a fixed size.
328
329The original version of this code used the first thing I made up that didn't
330obviously suck ;-) It was loosely based on invariants involving the Fibonacci
331sequence.
332
333It worked OK, but it was hard to reason about, and was subtle enough that the
334intended invariants weren't actually preserved.  Researchers discovered that
335when trying to complete a computer-generated correctness proof.  That was
336easily-enough repaired, but the discovery spurred quite a bit of academic
337interest in truly good ways to manage incremental merging on the fly.
338
339At least a dozen different approaches were developed, some provably having
340near-optimal worst case behavior with respect to the entropy of the
341distribution of run lengths.  Some details can be found in bpo-34561.
342
343The code now uses the "powersort" merge strategy from:
344
345    "Nearly-Optimal Mergesorts: Fast, Practical Sorting Methods
346     That Optimally Adapt to Existing Runs"
347    J. Ian Munro and Sebastian Wild
348
349The code is pretty simple, but the justification is quite involved, as it's
350based on fast approximations to optimal binary search trees, which are
351substantial topics on their own.
352
353Here we'll just cover some pragmatic details:
354
355The `powerloop()` function computes a run's "power". Say two adjacent runs
356begin at index s1. The first run has length n1, and the second run (starting
357at index s1+n1, called "s2" below) has length n2. The list has total length n.
358The "power" of the first run is a small integer, the depth of the node
359connecting the two runs in an ideal binary merge tree, where power 1 is the
360root node, and the power increases by 1 for each level deeper in the tree.
361
362The power is the least integer L such that the "midpoint interval" contains
363a rational number of the form J/2**L. The midpoint interval is the semi-
364closed interval:
365
366    ((s1 + n1/2)/n, (s2 + n2/2)/n]
367
368Yes, that's brain-busting at first ;-) Concretely, if (s1 + n1/2)/n and
369(s2 + n2/2)/n are computed to infinite precision in binary, the power L is
370the first position at which the 2**-L bit differs between the expansions.
371Since the left end of the interval is less than the right end, the first
372differing bit must be a 0 bit in the left quotient and a 1 bit in the right
373quotient.
374
375`powerloop()` emulates these divisions, 1 bit at a time, using comparisons,
376subtractions, and shifts in a loop.
377
378You'll notice the paper uses an O(1) method instead, but that relies on two
379things we don't have:
380
381- An O(1) "count leading zeroes" primitive. We can find such a thing as a C
382  extension on most platforms, but not all, and there's no uniform spelling
383  on the platforms that support it.
384
385- Integer division on an integer type twice as wide as needed to hold the
386  list length. But the latter is Py_ssize_t for us, and is typically the
387  widest native signed integer type the platform supports.
388
389But since runs in our algorithm are almost never very short, the once-per-run
390overhead of `powerloop()` seems lost in the noise.
391
392Detail: why is Py_ssize_t "wide enough" in `powerloop()`?  We do, after all,
393shift integers of that width left by 1.  How do we know that won't spill into
394the sign bit?  The trick is that we have some slop. `n` (the total list
395length) is the number of list elements, which is at most 4 times (on a 32-box,
396with 4-byte pointers) smaller than than the largest size_t. So at least the
397leading two bits of the integers we're using are clear.
398
399Since we can't compute a run's power before seeing the run that follows it,
400the most-recently identified run is never merged by `found_new_run()`.
401Instead a new run is only used to compute the 2nd-most-recent run's power.
402Then adjacent runs are merged so long as their saved power (tree depth) is
403greater than that newly computed power. When found_new_run() returns, only
404then is a new run pushed on to the stack of pending runs.
405
406A key invariant is that powers on the run stack are strictly decreasing
407(starting from the run at the top of the stack).
408
409Note that even powersort's strategy isn't always truly optimal. It can't be.
410Computing an optimal merge sequence can be done in time quadratic in the
411number of runs, which is very much slower, and also requires finding &
412remembering _all_ the runs' lengths (of which there may be billions) in
413advance.  It's remarkable, though, how close to optimal this strategy gets.
414
415Curious factoid: of all the alternatives I've seen in the literature,
416powersort's is the only one that's always truly optimal for a collection of 3
417run lengths (for three lengths A B C, it's always optimal to first merge the
418shorter of A and C with B).
419
420
421Merge Memory
422------------
423Merging adjacent runs of lengths A and B in-place, and in linear time, is
424difficult.  Theoretical constructions are known that can do it, but they're
425too difficult and slow for practical use.  But if we have temp memory equal
426to min(A, B), it's easy.
427
428If A is smaller (function merge_lo), copy A to a temp array, leave B alone,
429and then we can do the obvious merge algorithm left to right, from the temp
430area and B, starting the stores into where A used to live.  There's always a
431free area in the original area comprising a number of elements equal to the
432number not yet merged from the temp array (trivially true at the start;
433proceed by induction).  The only tricky bit is that if a comparison raises an
434exception, we have to remember to copy the remaining elements back in from
435the temp area, lest the array end up with duplicate entries from B.  But
436that's exactly the same thing we need to do if we reach the end of B first,
437so the exit code is pleasantly common to both the normal and error cases.
438
439If B is smaller (function merge_hi, which is merge_lo's "mirror image"),
440much the same, except that we need to merge right to left, copying B into a
441temp array and starting the stores at the right end of where B used to live.
442
443A refinement:  When we're about to merge adjacent runs A and B, we first do
444a form of binary search (more on that later) to see where B[0] should end up
445in A.  Elements in A preceding that point are already in their final
446positions, effectively shrinking the size of A.  Likewise we also search to
447see where A[-1] should end up in B, and elements of B after that point can
448also be ignored.  This cuts the amount of temp memory needed by the same
449amount.
450
451These preliminary searches may not pay off, and can be expected *not* to
452repay their cost if the data is random.  But they can win huge in all of
453time, copying, and memory savings when they do pay, so this is one of the
454"per-merge overheads" mentioned above that we're happy to endure because
455there is at most one very short run.  It's generally true in this algorithm
456that we're willing to gamble a little to win a lot, even though the net
457expectation is negative for random data.
458
459
460Merge Algorithms
461----------------
462merge_lo() and merge_hi() are where the bulk of the time is spent.  merge_lo
463deals with runs where A <= B, and merge_hi where A > B.  They don't know
464whether the data is clustered or uniform, but a lovely thing about merging
465is that many kinds of clustering "reveal themselves" by how many times in a
466row the winning merge element comes from the same run.  We'll only discuss
467merge_lo here; merge_hi is exactly analogous.
468
469Merging begins in the usual, obvious way, comparing the first element of A
470to the first of B, and moving B[0] to the merge area if it's less than A[0],
471else moving A[0] to the merge area.  Call that the "one pair at a time"
472mode.  The only twist here is keeping track of how many times in a row "the
473winner" comes from the same run.
474
475If that count reaches MIN_GALLOP, we switch to "galloping mode".  Here
476we *search* B for where A[0] belongs, and move over all the B's before
477that point in one chunk to the merge area, then move A[0] to the merge
478area.  Then we search A for where B[0] belongs, and similarly move a
479slice of A in one chunk.  Then back to searching B for where A[0] belongs,
480etc.  We stay in galloping mode until both searches find slices to copy
481less than MIN_GALLOP elements long, at which point we go back to one-pair-
482at-a-time mode.
483
484A refinement:  The MergeState struct contains the value of min_gallop that
485controls when we enter galloping mode, initialized to MIN_GALLOP.
486merge_lo() and merge_hi() adjust this higher when galloping isn't paying
487off, and lower when it is.
488
489
490Galloping
491---------
492Still without loss of generality, assume A is the shorter run.  In galloping
493mode, we first look for A[0] in B.  We do this via "galloping", comparing
494A[0] in turn to B[0], B[1], B[3], B[7], ..., B[2**j - 1], ..., until finding
495the k such that B[2**(k-1) - 1] < A[0] <= B[2**k - 1].  This takes at most
496roughly lg(B) comparisons, and, unlike a straight binary search, favors
497finding the right spot early in B (more on that later).
498
499After finding such a k, the region of uncertainty is reduced to 2**(k-1) - 1
500consecutive elements, and a straight binary search requires exactly k-1
501additional comparisons to nail it (see note REGION OF UNCERTAINTY).  Then we
502copy all the B's up to that point in one chunk, and then copy A[0].  Note
503that no matter where A[0] belongs in B, the combination of galloping + binary
504search finds it in no more than about 2*lg(B) comparisons.
505
506If we did a straight binary search, we could find it in no more than
507ceiling(lg(B+1)) comparisons -- but straight binary search takes that many
508comparisons no matter where A[0] belongs.  Straight binary search thus loses
509to galloping unless the run is quite long, and we simply can't guess
510whether it is in advance.
511
512If data is random and runs have the same length, A[0] belongs at B[0] half
513the time, at B[1] a quarter of the time, and so on:  a consecutive winning
514sub-run in B of length k occurs with probability 1/2**(k+1).  So long
515winning sub-runs are extremely unlikely in random data, and guessing that a
516winning sub-run is going to be long is a dangerous game.
517
518OTOH, if data is lopsided or lumpy or contains many duplicates, long
519stretches of winning sub-runs are very likely, and cutting the number of
520comparisons needed to find one from O(B) to O(log B) is a huge win.
521
522Galloping compromises by getting out fast if there isn't a long winning
523sub-run, yet finding such very efficiently when they exist.
524
525I first learned about the galloping strategy in a related context; see:
526
527    "Adaptive Set Intersections, Unions, and Differences" (2000)
528    Erik D. Demaine, Alejandro López-Ortiz, J. Ian Munro
529
530and its followup(s).  An earlier paper called the same strategy
531"exponential search":
532
533   "Optimistic Sorting and Information Theoretic Complexity"
534   Peter McIlroy
535   SODA (Fourth Annual ACM-SIAM Symposium on Discrete Algorithms), pp
536   467-474, Austin, Texas, 25-27 January 1993.
537
538and it probably dates back to an earlier paper by Bentley and Yao.  The
539McIlroy paper in particular has good analysis of a mergesort that's
540probably strongly related to this one in its galloping strategy.
541
542
543Galloping with a Broken Leg
544---------------------------
545So why don't we always gallop?  Because it can lose, on two counts:
546
5471. While we're willing to endure small per-merge overheads, per-comparison
548   overheads are a different story.  Calling Yet Another Function per
549   comparison is expensive, and gallop_left() and gallop_right() are
550   too long-winded for sane inlining.
551
5522. Galloping can-- alas --require more comparisons than linear one-at-time
553   search, depending on the data.
554
555#2 requires details.  If A[0] belongs before B[0], galloping requires 1
556compare to determine that, same as linear search, except it costs more
557to call the gallop function.  If A[0] belongs right before B[1], galloping
558requires 2 compares, again same as linear search.  On the third compare,
559galloping checks A[0] against B[3], and if it's <=, requires one more
560compare to determine whether A[0] belongs at B[2] or B[3].  That's a total
561of 4 compares, but if A[0] does belong at B[2], linear search would have
562discovered that in only 3 compares, and that's a huge loss!  Really.  It's
563an increase of 33% in the number of compares needed, and comparisons are
564expensive in Python.
565
566index in B where    # compares linear  # gallop  # binary  gallop
567A[0] belongs        search needs       compares  compares  total
568----------------    -----------------  --------  --------  ------
569               0                    1         1         0       1
570
571               1                    2         2         0       2
572
573               2                    3         3         1       4
574               3                    4         3         1       4
575
576               4                    5         4         2       6
577               5                    6         4         2       6
578               6                    7         4         2       6
579               7                    8         4         2       6
580
581               8                    9         5         3       8
582               9                   10         5         3       8
583              10                   11         5         3       8
584              11                   12         5         3       8
585                                        ...
586
587In general, if A[0] belongs at B[i], linear search requires i+1 comparisons
588to determine that, and galloping a total of 2*floor(lg(i))+2 comparisons.
589The advantage of galloping is unbounded as i grows, but it doesn't win at
590all until i=6.  Before then, it loses twice (at i=2 and i=4), and ties
591at the other values.  At and after i=6, galloping always wins.
592
593We can't guess in advance when it's going to win, though, so we do one pair
594at a time until the evidence seems strong that galloping may pay.  MIN_GALLOP
595is 7, and that's pretty strong evidence.  However, if the data is random, it
596simply will trigger galloping mode purely by luck every now and again, and
597it's quite likely to hit one of the losing cases next.  On the other hand,
598in cases like ~sort, galloping always pays, and MIN_GALLOP is larger than it
599"should be" then.  So the MergeState struct keeps a min_gallop variable
600that merge_lo and merge_hi adjust:  the longer we stay in galloping mode,
601the smaller min_gallop gets, making it easier to transition back to
602galloping mode (if we ever leave it in the current merge, and at the
603start of the next merge).  But whenever the gallop loop doesn't pay,
604min_gallop is increased by one, making it harder to transition back
605to galloping mode (and again both within a merge and across merges).  For
606random data, this all but eliminates the gallop penalty:  min_gallop grows
607large enough that we almost never get into galloping mode.  And for cases
608like ~sort, min_gallop can fall to as low as 1.  This seems to work well,
609but in all it's a minor improvement over using a fixed MIN_GALLOP value.
610
611
612Galloping Complication
613----------------------
614The description above was for merge_lo.  merge_hi has to merge "from the
615other end", and really needs to gallop starting at the last element in a run
616instead of the first.  Galloping from the first still works, but does more
617comparisons than it should (this is significant -- I timed it both ways). For
618this reason, the gallop_left() and gallop_right() (see note LEFT OR RIGHT)
619functions have a "hint" argument, which is the index at which galloping
620should begin.  So galloping can actually start at any index, and proceed at
621offsets of 1, 3, 7, 15, ... or -1, -3, -7, -15, ... from the starting index.
622
623In the code as I type it's always called with either 0 or n-1 (where n is
624the # of elements in a run).  It's tempting to try to do something fancier,
625melding galloping with some form of interpolation search; for example, if
626we're merging a run of length 1 with a run of length 10000, index 5000 is
627probably a better guess at the final result than either 0 or 9999.  But
628it's unclear how to generalize that intuition usefully, and merging of
629wildly unbalanced runs already enjoys excellent performance.
630
631~sort is a good example of when balanced runs could benefit from a better
632hint value:  to the extent possible, this would like to use a starting
633offset equal to the previous value of acount/bcount.  Doing so saves about
63410% of the compares in ~sort.  However, doing so is also a mixed bag,
635hurting other cases.
636
637
638Comparing Average # of Compares on Random Arrays
639------------------------------------------------
640[NOTE:  This was done when the new algorithm used about 0.1% more compares
641 on random data than does its current incarnation.]
642
643Here list.sort() is samplesort, and list.msort() this sort:
644
645"""
646import random
647from time import clock as now
648
649def fill(n):
650    from random import random
651    return [random() for i in range(n)]
652
653def mycmp(x, y):
654    global ncmp
655    ncmp += 1
656    return cmp(x, y)
657
658def timeit(values, method):
659    global ncmp
660    X = values[:]
661    bound = getattr(X, method)
662    ncmp = 0
663    t1 = now()
664    bound(mycmp)
665    t2 = now()
666    return t2-t1, ncmp
667
668format = "%5s  %9.2f  %11d"
669f2     = "%5s  %9.2f  %11.2f"
670
671def drive():
672    count = sst = sscmp = mst = mscmp = nelts = 0
673    while True:
674        n = random.randrange(100000)
675        nelts += n
676        x = fill(n)
677
678        t, c = timeit(x, 'sort')
679        sst += t
680        sscmp += c
681
682        t, c = timeit(x, 'msort')
683        mst += t
684        mscmp += c
685
686        count += 1
687        if count % 10:
688            continue
689
690        print "count", count, "nelts", nelts
691        print format % ("sort",  sst, sscmp)
692        print format % ("msort", mst, mscmp)
693        print f2     % ("", (sst-mst)*1e2/mst, (sscmp-mscmp)*1e2/mscmp)
694
695drive()
696"""
697
698I ran this on Windows and kept using the computer lightly while it was
699running.  time.clock() is wall-clock time on Windows, with better than
700microsecond resolution.  samplesort started with a 1.52% #-of-comparisons
701disadvantage, fell quickly to 1.48%, and then fluctuated within that small
702range.  Here's the last chunk of output before I killed the job:
703
704count 2630 nelts 130906543
705 sort    6110.80   1937887573
706msort    6002.78   1909389381
707            1.80         1.49
708
709We've done nearly 2 billion comparisons apiece at Python speed there, and
710that's enough <wink>.
711
712For random arrays of size 2 (yes, there are only 2 interesting ones),
713samplesort has a 50%(!) comparison disadvantage.  This is a consequence of
714samplesort special-casing at most one ascending run at the start, then
715falling back to the general case if it doesn't find an ascending run
716immediately.  The consequence is that it ends up using two compares to sort
717[2, 1].  Gratifyingly, timsort doesn't do any special-casing, so had to be
718taught how to deal with mixtures of ascending and descending runs
719efficiently in all cases.
720
721
722NOTES
723-----
724
725BINSORT
726A "binary insertion sort" is just like a textbook insertion sort, but instead
727of locating the correct position of the next item via linear (one at a time)
728search, an equivalent to Python's bisect.bisect_right is used to find the
729correct position in logarithmic time.  Most texts don't mention this
730variation, and those that do usually say it's not worth the bother:  insertion
731sort remains quadratic (expected and worst cases) either way.  Speeding the
732search doesn't reduce the quadratic data movement costs.
733
734But in CPython's case, comparisons are extraordinarily expensive compared to
735moving data, and the details matter.  Moving objects is just copying
736pointers.  Comparisons can be arbitrarily expensive (can invoke arbitrary
737user-supplied Python code), but even in simple cases (like 3 < 4) _all_
738decisions are made at runtime:  what's the type of the left comparand?  the
739type of the right?  do they need to be coerced to a common type?  where's the
740code to compare these types?  And so on.  Even the simplest Python comparison
741triggers a large pile of C-level pointer dereferences, conditionals, and
742function calls.
743
744So cutting the number of compares is almost always measurably helpful in
745CPython, and the savings swamp the quadratic-time data movement costs for
746reasonable minrun values.
747
748
749LEFT OR RIGHT
750gallop_left() and gallop_right() are akin to the Python bisect module's
751bisect_left() and bisect_right():  they're the same unless the slice they're
752searching contains a (at least one) value equal to the value being searched
753for.  In that case, gallop_left() returns the position immediately before the
754leftmost equal value, and gallop_right() the position immediately after the
755rightmost equal value.  The distinction is needed to preserve stability.  In
756general, when merging adjacent runs A and B, gallop_left is used to search
757thru B for where an element from A belongs, and gallop_right to search thru A
758for where an element from B belongs.
759
760
761REGION OF UNCERTAINTY
762Two kinds of confusion seem to be common about the claim that after finding
763a k such that
764
765    B[2**(k-1) - 1] < A[0] <= B[2**k - 1]
766
767then a binary search requires exactly k-1 tries to find A[0]'s proper
768location. For concreteness, say k=3, so B[3] < A[0] <= B[7].
769
770The first confusion takes the form "OK, then the region of uncertainty is at
771indices 3, 4, 5, 6 and 7:  that's 5 elements, not the claimed 2**(k-1) - 1 =
7723"; or the region is viewed as a Python slice and the objection is "but that's
773the slice B[3:7], so has 7-3 = 4 elements".  Resolution:  we've already
774compared A[0] against B[3] and against B[7], so A[0]'s correct location is
775already known wrt _both_ endpoints.  What remains is to find A[0]'s correct
776location wrt B[4], B[5] and B[6], which spans 3 elements.  Or in general, the
777slice (leaving off both endpoints) (2**(k-1)-1)+1 through (2**k-1)-1
778inclusive = 2**(k-1) through (2**k-1)-1 inclusive, which has
779    (2**k-1)-1 - 2**(k-1) + 1 =
780    2**k-1 - 2**(k-1) =
781    2*2**(k-1)-1 - 2**(k-1) =
782    (2-1)*2**(k-1) - 1 =
783    2**(k-1) - 1
784elements.
785
786The second confusion:  "k-1 = 2 binary searches can find the correct location
787among 2**(k-1) = 4 elements, but you're only applying it to 3 elements:  we
788could make this more efficient by arranging for the region of uncertainty to
789span 2**(k-1) elements."  Resolution:  that confuses "elements" with
790"locations".  In a slice with N elements, there are N+1 _locations_.  In the
791example, with the region of uncertainty B[4], B[5], B[6], there are 4
792locations:  before B[4], between B[4] and B[5], between B[5] and B[6], and
793after B[6].  In general, across 2**(k-1)-1 elements, there are 2**(k-1)
794locations.  That's why k-1 binary searches are necessary and sufficient.
795
796OPTIMIZATION OF INDIVIDUAL COMPARISONS
797As noted above, even the simplest Python comparison triggers a large pile of
798C-level pointer dereferences, conditionals, and function calls.  This can be
799partially mitigated by pre-scanning the data to determine whether the data is
800homogeneous with respect to type.  If so, it is sometimes possible to
801substitute faster type-specific comparisons for the slower, generic
802PyObject_RichCompareBool.
803