Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * sort_template.h
4 : *
5 : * A template for a sort algorithm that supports varying degrees of
6 : * specialization.
7 : *
8 : * Copyright (c) 2021-2024, PostgreSQL Global Development Group
9 : * Portions Copyright (c) 1992-1994, Regents of the University of California
10 : *
11 : * Usage notes:
12 : *
13 : * To generate functions specialized for a type, the following parameter
14 : * macros should be #define'd before this file is included.
15 : *
16 : * - ST_SORT - the name of a sort function to be generated
17 : * - ST_ELEMENT_TYPE - type of the referenced elements
18 : * - ST_DECLARE - if defined the functions and types are declared
19 : * - ST_DEFINE - if defined the functions and types are defined
20 : * - ST_SCOPE - scope (e.g. extern, static inline) for functions
21 : * - ST_CHECK_FOR_INTERRUPTS - if defined the sort is interruptible
22 : *
23 : * Instead of ST_ELEMENT_TYPE, ST_ELEMENT_TYPE_VOID can be defined. Then
24 : * the generated functions will automatically gain an "element_size"
25 : * parameter. This allows us to generate a traditional qsort function.
26 : *
27 : * One of the following macros must be defined, to show how to compare
28 : * elements. The first two options are arbitrary expressions depending
29 : * on whether an extra pass-through argument is desired, and the third
30 : * option should be defined if the sort function should receive a
31 : * function pointer at runtime.
32 : *
33 : * - ST_COMPARE(a, b) - a simple comparison expression
34 : * - ST_COMPARE(a, b, arg) - variant that takes an extra argument
35 : * - ST_COMPARE_RUNTIME_POINTER - sort function takes a function pointer
36 : *
37 : * NB: If the comparator function is inlined, some compilers may produce
38 : * worse code with the optimized comparison routines in common/int.h than
39 : * with code with the following form:
40 : *
41 : * if (a < b)
42 : * return -1;
43 : * if (a > b)
44 : * return 1;
45 : * return 0;
46 : *
47 : * To say that the comparator and therefore also sort function should
48 : * receive an extra pass-through argument, specify the type of the
49 : * argument.
50 : *
51 : * - ST_COMPARE_ARG_TYPE - type of extra argument
52 : *
53 : * The prototype of the generated sort function is:
54 : *
55 : * void ST_SORT(ST_ELEMENT_TYPE *data, size_t n,
56 : * [size_t element_size,]
57 : * [ST_SORT_compare_function compare,]
58 : * [ST_COMPARE_ARG_TYPE *arg]);
59 : *
60 : * ST_SORT_compare_function is a function pointer of the following type:
61 : *
62 : * int (*)(const ST_ELEMENT_TYPE *a, const ST_ELEMENT_TYPE *b,
63 : * [ST_COMPARE_ARG_TYPE *arg])
64 : *
65 : * HISTORY
66 : *
67 : * Modifications from vanilla NetBSD source:
68 : * - Add do ... while() macro fix
69 : * - Remove __inline, _DIAGASSERTs, __P
70 : * - Remove ill-considered "swap_cnt" switch to insertion sort, in favor
71 : * of a simple check for presorted input.
72 : * - Take care to recurse on the smaller partition, to bound stack usage
73 : * - Convert into a header that can generate specialized functions
74 : *
75 : * IDENTIFICATION
76 : * src/include/lib/sort_template.h
77 : *
78 : *-------------------------------------------------------------------------
79 : */
80 :
81 : /* $NetBSD: qsort.c,v 1.13 2003/08/07 16:43:42 agc Exp $ */
82 :
83 : /*-
84 : * Copyright (c) 1992, 1993
85 : * The Regents of the University of California. All rights reserved.
86 : *
87 : * Redistribution and use in source and binary forms, with or without
88 : * modification, are permitted provided that the following conditions
89 : * are met:
90 : * 1. Redistributions of source code must retain the above copyright
91 : * notice, this list of conditions and the following disclaimer.
92 : * 2. Redistributions in binary form must reproduce the above copyright
93 : * notice, this list of conditions and the following disclaimer in the
94 : * documentation and/or other materials provided with the distribution.
95 : * 3. Neither the name of the University nor the names of its contributors
96 : * may be used to endorse or promote products derived from this software
97 : * without specific prior written permission.
98 : *
99 : * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
100 : * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
101 : * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
102 : * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
103 : * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
104 : * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
105 : * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
106 : * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
107 : * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
108 : * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
109 : * SUCH DAMAGE.
110 : */
111 :
112 : /*
113 : * Qsort routine based on J. L. Bentley and M. D. McIlroy,
114 : * "Engineering a sort function",
115 : * Software--Practice and Experience 23 (1993) 1249-1265.
116 : *
117 : * We have modified their original by adding a check for already-sorted
118 : * input, which seems to be a win per discussions on pgsql-hackers around
119 : * 2006-03-21.
120 : *
121 : * Also, we recurse on the smaller partition and iterate on the larger one,
122 : * which ensures we cannot recurse more than log(N) levels (since the
123 : * partition recursed to is surely no more than half of the input). Bentley
124 : * and McIlroy explicitly rejected doing this on the grounds that it's "not
125 : * worth the effort", but we have seen crashes in the field due to stack
126 : * overrun, so that judgment seems wrong.
127 : */
128 :
129 : #define ST_MAKE_PREFIX(a) CppConcat(a,_)
130 : #define ST_MAKE_NAME(a,b) ST_MAKE_NAME_(ST_MAKE_PREFIX(a),b)
131 : #define ST_MAKE_NAME_(a,b) CppConcat(a,b)
132 :
133 : /*
134 : * If the element type is void, we'll also need an element_size argument
135 : * because we don't know the size.
136 : */
137 : #ifdef ST_ELEMENT_TYPE_VOID
138 : #define ST_ELEMENT_TYPE void
139 : #define ST_SORT_PROTO_ELEMENT_SIZE , size_t element_size
140 : #define ST_SORT_INVOKE_ELEMENT_SIZE , element_size
141 : #else
142 : #define ST_SORT_PROTO_ELEMENT_SIZE
143 : #define ST_SORT_INVOKE_ELEMENT_SIZE
144 : #endif
145 :
146 : /*
147 : * If the user wants to be able to pass in compare functions at runtime,
148 : * we'll need to make that an argument of the sort and med3 functions.
149 : */
150 : #ifdef ST_COMPARE_RUNTIME_POINTER
151 : /*
152 : * The type of the comparator function pointer that ST_SORT will take, unless
153 : * you've already declared a type name manually and want to use that instead of
154 : * having a new one defined.
155 : */
156 : #ifndef ST_COMPARATOR_TYPE_NAME
157 : #define ST_COMPARATOR_TYPE_NAME ST_MAKE_NAME(ST_SORT, compare_function)
158 : #endif
159 : #define ST_COMPARE compare
160 : #ifndef ST_COMPARE_ARG_TYPE
161 : #define ST_SORT_PROTO_COMPARE , ST_COMPARATOR_TYPE_NAME compare
162 : #define ST_SORT_INVOKE_COMPARE , compare
163 : #else
164 : #define ST_SORT_PROTO_COMPARE , ST_COMPARATOR_TYPE_NAME compare
165 : #define ST_SORT_INVOKE_COMPARE , compare
166 : #endif
167 : #else
168 : #define ST_SORT_PROTO_COMPARE
169 : #define ST_SORT_INVOKE_COMPARE
170 : #endif
171 :
172 : /*
173 : * If the user wants to use a compare function or expression that takes an
174 : * extra argument, we'll need to make that an argument of the sort, compare and
175 : * med3 functions.
176 : */
177 : #ifdef ST_COMPARE_ARG_TYPE
178 : #define ST_SORT_PROTO_ARG , ST_COMPARE_ARG_TYPE *arg
179 : #define ST_SORT_INVOKE_ARG , arg
180 : #else
181 : #define ST_SORT_PROTO_ARG
182 : #define ST_SORT_INVOKE_ARG
183 : #endif
184 :
185 : #ifdef ST_DECLARE
186 :
187 : #ifdef ST_COMPARE_RUNTIME_POINTER
188 : typedef int (*ST_COMPARATOR_TYPE_NAME) (const ST_ELEMENT_TYPE *,
189 : const ST_ELEMENT_TYPE * ST_SORT_PROTO_ARG);
190 : #endif
191 :
192 : /* Declare the sort function. Note optional arguments at end. */
193 : ST_SCOPE void ST_SORT(ST_ELEMENT_TYPE * first, size_t n
194 : ST_SORT_PROTO_ELEMENT_SIZE
195 : ST_SORT_PROTO_COMPARE
196 : ST_SORT_PROTO_ARG);
197 :
198 : #endif
199 :
200 : #ifdef ST_DEFINE
201 :
202 : /* sort private helper functions */
203 : #define ST_MED3 ST_MAKE_NAME(ST_SORT, med3)
204 : #define ST_SWAP ST_MAKE_NAME(ST_SORT, swap)
205 : #define ST_SWAPN ST_MAKE_NAME(ST_SORT, swapn)
206 :
207 : /* Users expecting to run very large sorts may need them to be interruptible. */
208 : #ifdef ST_CHECK_FOR_INTERRUPTS
209 : #define DO_CHECK_FOR_INTERRUPTS() CHECK_FOR_INTERRUPTS()
210 : #else
211 : #define DO_CHECK_FOR_INTERRUPTS()
212 : #endif
213 :
214 : /*
215 : * Create wrapper macros that know how to invoke compare, med3 and sort with
216 : * the right arguments.
217 : */
218 : #ifdef ST_COMPARE_RUNTIME_POINTER
219 : #define DO_COMPARE(a_, b_) ST_COMPARE((a_), (b_) ST_SORT_INVOKE_ARG)
220 : #elif defined(ST_COMPARE_ARG_TYPE)
221 : #define DO_COMPARE(a_, b_) ST_COMPARE((a_), (b_), arg)
222 : #else
223 : #define DO_COMPARE(a_, b_) ST_COMPARE((a_), (b_))
224 : #endif
225 : #define DO_MED3(a_, b_, c_) \
226 : ST_MED3((a_), (b_), (c_) \
227 : ST_SORT_INVOKE_COMPARE \
228 : ST_SORT_INVOKE_ARG)
229 : #define DO_SORT(a_, n_) \
230 : ST_SORT((a_), (n_) \
231 : ST_SORT_INVOKE_ELEMENT_SIZE \
232 : ST_SORT_INVOKE_COMPARE \
233 : ST_SORT_INVOKE_ARG)
234 :
235 : /*
236 : * If we're working with void pointers, we'll use pointer arithmetic based on
237 : * uint8, and use the runtime element_size to step through the array and swap
238 : * elements. Otherwise we'll work with ST_ELEMENT_TYPE.
239 : */
240 : #ifndef ST_ELEMENT_TYPE_VOID
241 : #define ST_POINTER_TYPE ST_ELEMENT_TYPE
242 : #define ST_POINTER_STEP 1
243 : #define DO_SWAPN(a_, b_, n_) ST_SWAPN((a_), (b_), (n_))
244 : #define DO_SWAP(a_, b_) ST_SWAP((a_), (b_))
245 : #else
246 : #define ST_POINTER_TYPE uint8
247 : #define ST_POINTER_STEP element_size
248 : #define DO_SWAPN(a_, b_, n_) ST_SWAPN((a_), (b_), (n_))
249 : #define DO_SWAP(a_, b_) DO_SWAPN((a_), (b_), element_size)
250 : #endif
251 :
252 : /*
253 : * Find the median of three values. Currently, performance seems to be best
254 : * if the comparator is inlined here, but the med3 function is not inlined
255 : * in the qsort function.
256 : *
257 : * Refer to the comment at the top of this file for known caveats to consider
258 : * when writing inlined comparator functions.
259 : */
260 : static pg_noinline ST_ELEMENT_TYPE *
261 35189338 : ST_MED3(ST_ELEMENT_TYPE * a,
262 : ST_ELEMENT_TYPE * b,
263 : ST_ELEMENT_TYPE * c
264 : ST_SORT_PROTO_COMPARE
265 : ST_SORT_PROTO_ARG)
266 : {
267 35189338 : return DO_COMPARE(a, b) < 0 ?
268 20019422 : (DO_COMPARE(b, c) < 0 ? b : (DO_COMPARE(a, c) < 0 ? c : a))
269 55208760 : : (DO_COMPARE(b, c) > 0 ? b : (DO_COMPARE(a, c) < 0 ? a : c));
270 : }
271 :
272 : static inline void
273 2473241696 : ST_SWAP(ST_POINTER_TYPE * a, ST_POINTER_TYPE * b)
274 : {
275 2473241696 : ST_POINTER_TYPE tmp = *a;
276 :
277 2473241696 : *a = *b;
278 2473241696 : *b = tmp;
279 2473241696 : }
280 :
281 : static inline void
282 248602630 : ST_SWAPN(ST_POINTER_TYPE * a, ST_POINTER_TYPE * b, size_t n)
283 : {
284 2682125294 : for (size_t i = 0; i < n; ++i)
285 2433522664 : ST_SWAP(&a[i], &b[i]);
286 248602630 : }
287 :
288 : /*
289 : * Sort an array.
290 : */
291 : ST_SCOPE void
292 22026290 : ST_SORT(ST_ELEMENT_TYPE * data, size_t n
293 : ST_SORT_PROTO_ELEMENT_SIZE
294 : ST_SORT_PROTO_COMPARE
295 : ST_SORT_PROTO_ARG)
296 : {
297 22026290 : ST_POINTER_TYPE *a = (ST_POINTER_TYPE *) data,
298 : *pa,
299 : *pb,
300 : *pc,
301 : *pd,
302 : *pl,
303 : *pm,
304 : *pn;
305 : size_t d1,
306 : d2;
307 : int r,
308 : presorted;
309 :
310 48419826 : loop:
311 20212136 : DO_CHECK_FOR_INTERRUPTS();
312 48419826 : if (n < 7)
313 : {
314 78923254 : for (pm = a + ST_POINTER_STEP; pm < a + n * ST_POINTER_STEP;
315 57493752 : pm += ST_POINTER_STEP)
316 120502248 : for (pl = pm; pl > a && DO_COMPARE(pl - ST_POINTER_STEP, pl) > 0;
317 63008436 : pl -= ST_POINTER_STEP)
318 63008436 : DO_SWAP(pl, pl - ST_POINTER_STEP);
319 21429442 : return;
320 : }
321 26990324 : presorted = 1;
322 295831966 : for (pm = a + ST_POINTER_STEP; pm < a + n * ST_POINTER_STEP;
323 268841642 : pm += ST_POINTER_STEP)
324 : {
325 69494518 : DO_CHECK_FOR_INTERRUPTS();
326 295239494 : if (DO_COMPARE(pm - ST_POINTER_STEP, pm) > 0)
327 : {
328 26397828 : presorted = 0;
329 26397828 : break;
330 : }
331 : }
332 26990300 : if (presorted)
333 592472 : return;
334 26397828 : pm = a + (n / 2) * ST_POINTER_STEP;
335 26397828 : if (n > 7)
336 : {
337 23784034 : pl = a;
338 23784034 : pn = a + (n - 1) * ST_POINTER_STEP;
339 23784034 : if (n > 40)
340 : {
341 3801768 : size_t d = (n / 8) * ST_POINTER_STEP;
342 :
343 3801768 : pl = DO_MED3(pl, pl + d, pl + 2 * d);
344 3801768 : pm = DO_MED3(pm - d, pm, pm + d);
345 3801768 : pn = DO_MED3(pn - 2 * d, pn - d, pn);
346 : }
347 23784034 : pm = DO_MED3(pl, pm, pn);
348 : }
349 26397828 : DO_SWAP(a, pm);
350 26397828 : pa = pb = a + ST_POINTER_STEP;
351 26397828 : pc = pd = a + (n - 1) * ST_POINTER_STEP;
352 : for (;;)
353 : {
354 527081172 : while (pb <= pc && (r = DO_COMPARE(pb, a)) <= 0)
355 : {
356 360594424 : if (r == 0)
357 : {
358 3225096 : DO_SWAP(pa, pb);
359 3225096 : pa += ST_POINTER_STEP;
360 : }
361 360594424 : pb += ST_POINTER_STEP;
362 194621082 : DO_CHECK_FOR_INTERRUPTS();
363 : }
364 498888382 : while (pb <= pc && (r = DO_COMPARE(pc, a)) >= 0)
365 : {
366 332401634 : if (r == 0)
367 : {
368 2805726 : DO_SWAP(pc, pd);
369 2805726 : pd -= ST_POINTER_STEP;
370 : }
371 332401634 : pc -= ST_POINTER_STEP;
372 162216906 : DO_CHECK_FOR_INTERRUPTS();
373 : }
374 166486748 : if (pb > pc)
375 26397828 : break;
376 140088920 : DO_SWAP(pb, pc);
377 140088920 : pb += ST_POINTER_STEP;
378 140088920 : pc -= ST_POINTER_STEP;
379 : }
380 26397828 : pn = a + n * ST_POINTER_STEP;
381 26397828 : d1 = Min(pa - a, pb - pa);
382 26397828 : DO_SWAPN(a, pb - d1, d1);
383 26397828 : d1 = Min(pd - pc, pn - pd - ST_POINTER_STEP);
384 26397828 : DO_SWAPN(pb, pn - d1, d1);
385 26397828 : d1 = pb - pa;
386 26397828 : d2 = pd - pc;
387 26397828 : if (d1 <= d2)
388 : {
389 : /* Recurse on left partition, then iterate on right partition */
390 13440904 : if (d1 > ST_POINTER_STEP)
391 10920086 : DO_SORT(a, d1 / ST_POINTER_STEP);
392 13440904 : if (d2 > ST_POINTER_STEP)
393 : {
394 : /* Iterate rather than recurse to save stack space */
395 : /* DO_SORT(pn - d2, d2 / ST_POINTER_STEP) */
396 13438434 : a = pn - d2;
397 13438434 : n = d2 / ST_POINTER_STEP;
398 13438434 : goto loop;
399 : }
400 : }
401 : else
402 : {
403 : /* Recurse on right partition, then iterate on left partition */
404 12956924 : if (d2 > ST_POINTER_STEP)
405 7010100 : DO_SORT(pn - d2, d2 / ST_POINTER_STEP);
406 12956924 : if (d1 > ST_POINTER_STEP)
407 : {
408 : /* Iterate rather than recurse to save stack space */
409 : /* DO_SORT(a, d1 / ST_POINTER_STEP) */
410 12955102 : n = d1 / ST_POINTER_STEP;
411 12955102 : goto loop;
412 : }
413 : }
414 : }
415 : #endif
416 :
417 : #undef DO_CHECK_FOR_INTERRUPTS
418 : #undef DO_COMPARE
419 : #undef DO_MED3
420 : #undef DO_SORT
421 : #undef DO_SWAP
422 : #undef DO_SWAPN
423 : #undef ST_CHECK_FOR_INTERRUPTS
424 : #undef ST_COMPARATOR_TYPE_NAME
425 : #undef ST_COMPARE
426 : #undef ST_COMPARE_ARG_TYPE
427 : #undef ST_COMPARE_RUNTIME_POINTER
428 : #undef ST_ELEMENT_TYPE
429 : #undef ST_ELEMENT_TYPE_VOID
430 : #undef ST_MAKE_NAME
431 : #undef ST_MAKE_NAME_
432 : #undef ST_MAKE_PREFIX
433 : #undef ST_MED3
434 : #undef ST_POINTER_STEP
435 : #undef ST_POINTER_TYPE
436 : #undef ST_SCOPE
437 : #undef ST_SORT
438 : #undef ST_SORT_INVOKE_ARG
439 : #undef ST_SORT_INVOKE_COMPARE
440 : #undef ST_SORT_INVOKE_ELEMENT_SIZE
441 : #undef ST_SORT_PROTO_ARG
442 : #undef ST_SORT_PROTO_COMPARE
443 : #undef ST_SORT_PROTO_ELEMENT_SIZE
444 : #undef ST_SWAP
445 : #undef ST_SWAPN
|
