summaryrefslogtreecommitdiff
path: root/mysys/my_rdtsc.c
blob: c624a5ca288730303a5ec0bc1541b959dbf5523b (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
/* Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.

  This program is free software; you can redistribute it and/or modify
  it under the terms of the GNU General Public License as published by
  the Free Software Foundation; version 2 of the License.

  This program is distributed in the hope that it will be useful,
  but WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  GNU General Public License for more details.

  You should have received a copy of the GNU General Public License
  along with this program; if not, write to the Free Software
  Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301  USA */

/*
  rdtsc3 -- multi-platform timer code
  pgulutzan@mysql.com, 2005-08-29
  modified 2008-11-02

  Functions:

  my_timer_cycles           ulonglong cycles
  my_timer_nanoseconds      ulonglong nanoseconds
  my_timer_microseconds     ulonglong "microseconds"
  my_timer_milliseconds     ulonglong milliseconds
  my_timer_ticks            ulonglong ticks
  my_timer_init             initialization / test

  We'll call the first 5 functions (the ones that return
  a ulonglong) "my_timer_xxx" functions.
  Each my_timer_xxx function returns a 64-bit timing value
  since an arbitrary 'epoch' start. Since the only purpose
  is to determine elapsed times, wall-clock time-of-day
  is not known and not relevant.

  The my_timer_init function is necessary for initializing.
  It returns information (underlying routine name,
  frequency, resolution, overhead) about all my_timer_xxx
  functions. A program should call my_timer_init once,
  use the information to decide what my_timer_xxx function
  to use, and subsequently call that function by function
  pointer.

  A typical use would be:
  my_timer_init()        ... once, at program start
  ...
  time1= my_timer_xxx()  ... time before start
  [code that's timed]
  time2= my_timer_xxx()  ... time after end
  elapsed_time= (time2 - time1) - overhead
*/

#include "my_global.h"
#include "my_rdtsc.h"

#if defined(_WIN32)
#include <stdio.h>
#include "windows.h"
#else
#include <stdio.h>
#endif

#if !defined(_WIN32)
#if TIME_WITH_SYS_TIME
#include <sys/time.h>
#include <time.h>           /* for clock_gettime */
#else
#if HAVE_SYS_TIME_H
#include <sys/time.h>
#elif defined(HAVE_TIME_H)
#include <time.h>
#endif
#endif
#endif

#if defined(HAVE_ASM_MSR_H) && defined(HAVE_RDTSCLL)
#include <asm/msr.h>         /* for rdtscll */
#endif

#if defined(HAVE_SYS_TIMEB_H) && defined(HAVE_FTIME)
#include <sys/timeb.h>       /* for ftime */
#endif

#if defined(HAVE_SYS_TIMES_H) && defined(HAVE_TIMES)
#include <sys/times.h>       /* for times */
#endif

#if defined(__INTEL_COMPILER) && defined(__ia64__) && defined(HAVE_IA64INTRIN_H)
#include <ia64intrin.h>    /* for __GetReg */
#endif

#if defined(__APPLE__) && defined(__MACH__)
#include <mach/mach_time.h>
#endif

#if defined(__SUNPRO_CC) && defined(__sparcv9) && defined(_LP64) && !defined(__SunOS_5_7)
extern "C" ulonglong my_timer_cycles_il_sparc64();
#elif defined(__SUNPRO_CC) && defined(_ILP32) && !defined(__SunOS_5_7)
extern "C" ulonglong my_timer_cycles_il_sparc32();
#elif defined(__SUNPRO_CC) && defined(__i386) && defined(_ILP32)
extern "C" ulonglong my_timer_cycles_il_i386();
#elif defined(__SUNPRO_CC) && defined(__x86_64) && defined(_LP64)
extern "C" ulonglong my_timer_cycles_il_x86_64();
#elif defined(__SUNPRO_C) && defined(__sparcv9) && defined(_LP64) && !defined(__SunOS_5_7)
ulonglong my_timer_cycles_il_sparc64();
#elif defined(__SUNPRO_C) && defined(_ILP32) && !defined(__SunOS_5_7)
ulonglong my_timer_cycles_il_sparc32();
#elif defined(__SUNPRO_C) && defined(__i386) && defined(_ILP32)
ulonglong my_timer_cycles_il_i386();
#elif defined(__SUNPRO_C) && defined(__x86_64) && defined(_LP64)
ulonglong my_timer_cycles_il_x86_64();
#endif

#if defined(__INTEL_COMPILER)
/*
  icc warning #1011 is:
  missing return statement at end of non-void function
*/
#pragma warning (disable:1011)
#endif

/*
  For cycles, we depend on RDTSC for x86 platforms,
  or on time buffer (which is not really a cycle count
  but a separate counter with less than nanosecond
  resolution) for most PowerPC platforms, or on
  gethrtime which is okay for hpux and solaris, or on
  clock_gettime(CLOCK_SGI_CYCLE) for Irix platforms,
  or on read_real_time for aix platforms. There is
  nothing for Alpha platforms, they would be tricky.
*/

ulonglong my_timer_cycles(void)
{
#if defined(__GNUC__) && defined(__i386__)
  /* This works much better if compiled with "gcc -O3". */
  ulonglong result;
  __asm__ __volatile__ ("rdtsc" : "=A" (result));
  return result;
#elif defined(__SUNPRO_C) && defined(__i386)
  __asm("rdtsc");
#elif defined(__GNUC__) && defined(__x86_64__)
  ulonglong result;
  __asm__ __volatile__ ("rdtsc\n\t" \
                        "shlq $32,%%rdx\n\t" \
                        "orq %%rdx,%%rax"
                        : "=a" (result) :: "%edx");
  return result;
#elif defined(HAVE_ASM_MSR_H) && defined(HAVE_RDTSCLL)
  {
    ulonglong result;
    rdtscll(result);
    return result;
  }
#elif defined(_WIN32) && defined(_M_IX86)
  __asm {rdtsc};
#elif defined(_WIN64) && defined(_M_X64)
  /* For 64-bit Windows: unsigned __int64 __rdtsc(); */
  return __rdtsc();
#elif defined(__INTEL_COMPILER) && defined(__ia64__) && defined(HAVE_IA64INTRIN_H)
  return (ulonglong) __getReg(_IA64_REG_AR_ITC); /* (3116) */
#elif defined(__GNUC__) && defined(__ia64__)
  {
    ulonglong result;
    __asm __volatile__ ("mov %0=ar.itc" : "=r" (result));
    return result;
  }
#elif defined(__GNUC__) && (defined(__powerpc__) || defined(__POWERPC__) || (defined(_POWER) && defined(_AIX52))) && (defined(__64BIT__) || defined(_ARCH_PPC64))
  {
    ulonglong result;
    __asm __volatile__ ("mftb %0" : "=r" (result));
    return result;
  }
#elif defined(__GNUC__) && (defined(__powerpc__) || defined(__POWERPC__) || (defined(_POWER) && defined(_AIX52))) && (!defined(__64BIT__) && !defined(_ARCH_PPC64))
  {
    /*
      mftbu means "move from time-buffer-upper to result".
      The loop is saying: x1=upper, x2=lower, x3=upper,
      if x1!=x3 there was an overflow so repeat.
    */
    unsigned int x1, x2, x3;
    ulonglong result;
    for (;;)
    {
       __asm __volatile__ ( "mftbu %0" : "=r"(x1) );
       __asm __volatile__ ( "mftb %0" : "=r"(x2) );
       __asm __volatile__ ( "mftbu %0" : "=r"(x3) );
       if (x1 == x3) break;
    }
    result = x1;
    return ( result << 32 ) | x2;
  }
#elif (defined(__SUNPRO_CC) || defined(__SUNPRO_C)) && defined(__sparcv9) && defined(_LP64) && !defined(__SunOS_5_7)
  return (my_timer_cycles_il_sparc64());
#elif (defined(__SUNPRO_CC) || defined(__SUNPRO_C)) && defined(_ILP32) && !defined(__SunOS_5_7)
  return (my_timer_cycles_il_sparc32());
#elif (defined(__SUNPRO_CC) || defined(__SUNPRO_C)) && defined(__i386) && defined(_ILP32)
  /* This is probably redundant for __SUNPRO_C. */
  return (my_timer_cycles_il_i386());
#elif (defined(__SUNPRO_CC) || defined(__SUNPRO_C)) && defined(__x86_64) && defined(_LP64)
  return (my_timer_cycles_il_x86_64());
#elif defined(__GNUC__) && defined(__sparcv9) && defined(_LP64)  && (__GNUC__>2)
  {
    ulonglong result;
    __asm __volatile__ ("rd %%tick,%0" : "=r" (result));
    return result;
  }
#elif defined(__GNUC__) && defined(__sparc__) && !defined(_LP64) && (__GNUC__>2)
  {
      union {
              ulonglong wholeresult;
              struct {
                      ulong high;
                      ulong low;
              }       splitresult;
      } result;
    __asm __volatile__ ("rd %%tick,%1; srlx %1,32,%0" : "=r" (result.splitresult.high), "=r" (result.splitresult.low));
    return result.wholeresult;
  }
#elif defined(__sgi) && defined(HAVE_CLOCK_GETTIME) && defined(CLOCK_SGI_CYCLE)
  {
    struct timespec tp;
    clock_gettime(CLOCK_SGI_CYCLE, &tp);
    return (ulonglong) tp.tv_sec * 1000000000 + (ulonglong) tp.tv_nsec;
  }
#elif defined(HAVE_SYS_TIMES_H) && defined(HAVE_GETHRTIME)
  /* gethrtime may appear as either cycle or nanosecond counter */
  return (ulonglong) gethrtime();
#else
  return 0;
#endif
}

#if defined(__INTEL_COMPILER)
/* re-enable warning#1011 which was only for my_timer_cycles() */
/* There may be an icc bug which means we must leave disabled. */
#pragma warning (default:1011)
#endif

/*
  For nanoseconds, most platforms have nothing available that
  (a) doesn't require bringing in a 40-kb librt.so library
  (b) really has nanosecond resolution.
*/

ulonglong my_timer_nanoseconds(void)
{
#if defined(HAVE_READ_REAL_TIME)
  {
    timebasestruct_t tr;
    read_real_time(&tr, TIMEBASE_SZ);
    return (ulonglong) tr.tb_high * 1000000000 + (ulonglong) tr.tb_low;
  }
#elif defined(HAVE_SYS_TIMES_H) && defined(HAVE_GETHRTIME)
  /* SunOS 5.10+, Solaris, HP-UX: hrtime_t gethrtime(void) */
  return (ulonglong) gethrtime();
#elif defined(HAVE_CLOCK_GETTIME) && defined(CLOCK_REALTIME)
  {
    struct timespec tp;
    clock_gettime(CLOCK_REALTIME, &tp);
    return (ulonglong) tp.tv_sec * 1000000000 + (ulonglong) tp.tv_nsec;
  }
#elif defined(__APPLE__) && defined(__MACH__)
  {
    ulonglong tm;
    static mach_timebase_info_data_t timebase_info= {0,0};
    if (timebase_info.denom == 0)
      (void) mach_timebase_info(&timebase_info);
    tm= mach_absolute_time();
    return (tm * timebase_info.numer) / timebase_info.denom;
  }
#else
  return 0;
#endif
}

/*
  For microseconds, gettimeofday() is available on
  almost all platforms. On Windows we use
  QueryPerformanceCounter which will usually tick over
  3.5 million times per second, and we don't throw
  away the extra precision. (On Windows Server 2003
  the frequency is same as the cycle frequency.)
*/

ulonglong my_timer_microseconds(void)
{
#if defined(HAVE_GETTIMEOFDAY)
  {
    static ulonglong last_value= 0;
    struct timeval tv;
    if (gettimeofday(&tv, NULL) == 0)
      last_value= (ulonglong) tv.tv_sec * 1000000 + (ulonglong) tv.tv_usec;
    else
    {
      /*
        There are reports that gettimeofday(2) can have intermittent failures
        on some platform, see for example Bug#36819.
        We are not trying again or looping, just returning the best value possible
        under the circumstances ...
      */
      last_value++;
    }
    return last_value;
  }
#elif defined(_WIN32)
  {
    /* QueryPerformanceCounter usually works with about 1/3 microsecond. */
    LARGE_INTEGER t_cnt;

    QueryPerformanceCounter(&t_cnt);
    return (ulonglong) t_cnt.QuadPart;
  }
#else
  return 0;
#endif
}

/*
  For milliseconds, we use ftime() if it's supported
  or time()*1000 if it's not. With modern versions of
  Windows and with HP Itanium, resolution is 10-15
  milliseconds.
*/

ulonglong my_timer_milliseconds(void)
{
#if defined(HAVE_SYS_TIMEB_H) && defined(HAVE_FTIME)
  /* ftime() is obsolete but maybe the platform is old */
  struct timeb ft;
  ftime(&ft);
  return (ulonglong)ft.time * 1000 + (ulonglong)ft.millitm;
#elif defined(HAVE_TIME)
  return (ulonglong) time(NULL) * 1000;
#elif defined(_WIN32)
   FILETIME ft;
   GetSystemTimeAsFileTime( &ft );
   return ((ulonglong)ft.dwLowDateTime +
                  (((ulonglong)ft.dwHighDateTime) << 32))/10000;
#else
  return 0;
#endif
}

/*
  For ticks, which we handle with times(), the frequency
  is usually 100/second and the overhead is surprisingly
  bad, sometimes even worse than gettimeofday's overhead.
*/

ulonglong my_timer_ticks(void)
{
#if defined(HAVE_SYS_TIMES_H) && defined(HAVE_TIMES)
  {
    struct tms times_buf;
    return (ulonglong) times(&times_buf);
  }
#elif defined(_WIN32)
  return (ulonglong) GetTickCount();
#else
  return 0;
#endif
}

/*
  The my_timer_init() function and its sub-functions
  have several loops which call timers. If there's
  something wrong with a timer -- which has never
  happened in tests -- we want the loop to end after
  an arbitrary number of iterations, and my_timer_info
  will show a discouraging result. The arbitrary
  number is 1,000,000.
*/
#define MY_TIMER_ITERATIONS 1000000

/*
  Calculate overhead. Called from my_timer_init().
  Usually best_timer_overhead = cycles.overhead or
  nanoseconds.overhead, so returned amount is in
  cycles or nanoseconds. We repeat the calculation
  ten times, so that we can disregard effects of
  caching or interrupts. Result is quite consistent
  for cycles, at least. But remember it's a minimum.
*/

static void my_timer_init_overhead(ulonglong *overhead,
                                   ulonglong (*cycle_timer)(void),
                                   ulonglong (*this_timer)(void),
                                   ulonglong best_timer_overhead)
{
  ulonglong time1, time2;
  int i;

  /* *overhead, least of 20 calculations - cycles.overhead */
  for (i= 0, *overhead= 1000000000; i < 20; ++i)
  {
    time1= cycle_timer();
    this_timer(); /* rather than 'time_tmp= timer();' */
    time2= cycle_timer() - time1;
    if (*overhead > time2)
      *overhead= time2;
  }
  *overhead-= best_timer_overhead;
}

/*
  Calculate Resolution. Called from my_timer_init().
  If a timer goes up by jumps, e.g. 1050, 1075, 1100, ...
  then the best resolution is the minimum jump, e.g. 25.
  If it's always divisible by 1000 then it's just a
  result of multiplication of a lower-precision timer
  result, e.g. nanoseconds are often microseconds * 1000.
  If the minimum jump is less than an arbitrary passed
  figure (a guess based on maximum overhead * 2), ignore.
  Usually we end up with nanoseconds = 1 because it's too
  hard to detect anything <= 100 nanoseconds.
  Often GetTickCount() has resolution = 15.
  We don't check with ticks because they take too long.
*/
static ulonglong my_timer_init_resolution(ulonglong (*this_timer)(void),
                                          ulonglong overhead_times_2)
{
  ulonglong time1, time2;
  ulonglong best_jump;
  int i, jumps, divisible_by_1000, divisible_by_1000000;

  divisible_by_1000= divisible_by_1000000= 0;
  best_jump= 1000000;
  for (i= jumps= 0; jumps < 3 && i < MY_TIMER_ITERATIONS * 10; ++i)
  {
    time1= this_timer();
    time2= this_timer();
    time2-= time1;
    if (time2)
    {
      ++jumps;
      if (!(time2 % 1000))
      {
        ++divisible_by_1000;
        if (!(time2 % 1000000))
          ++divisible_by_1000000;
      }
      if (best_jump > time2)
        best_jump= time2;
      /* For milliseconds, one jump is enough. */
      if (overhead_times_2 == 0)
        break;
    }
  }
  if (jumps == 3)
  {
    if (jumps == divisible_by_1000000)
      return 1000000;
    if (jumps == divisible_by_1000)
      return 1000;
  }
  if (best_jump > overhead_times_2)
    return best_jump;
  return 1;
}

/*
  Calculate cycle frequency by seeing how many cycles pass
  in a 200-microsecond period. I tried with 10-microsecond
  periods originally, and the result was often very wrong.
*/

static ulonglong my_timer_init_frequency(MY_TIMER_INFO *mti)
{
  int i;
  ulonglong time1, time2, time3, time4;
  time1= my_timer_cycles();
  time2= my_timer_microseconds();
  time3= time2; /* Avoids a Microsoft/IBM compiler warning */
  for (i= 0; i < MY_TIMER_ITERATIONS; ++i)
  {
    time3= my_timer_microseconds();
    if (time3 - time2 > 200) break;
  }
  time4= my_timer_cycles() - mti->cycles.overhead;
  time4-= mti->microseconds.overhead;
  return (mti->microseconds.frequency * (time4 - time1)) / (time3 - time2);
}

/*
  Call my_timer_init before the first call to my_timer_xxx().
  If something must be initialized, it happens here.
  Set: what routine is being used e.g. "asm_x86"
  Set: function, overhead, actual frequency, resolution.
*/

void my_timer_init(MY_TIMER_INFO *mti)
{
  ulonglong (*best_timer)(void);
  ulonglong best_timer_overhead;
  ulonglong time1, time2;
  int i;

  /* cycles */
  mti->cycles.frequency= 1000000000;
#if defined(__GNUC__) && defined(__i386__)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_X86;
#elif defined(__SUNPRO_C) && defined(__i386)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_X86;
#elif defined(__GNUC__) && defined(__x86_64__)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_X86_64;
#elif defined(HAVE_ASM_MSR_H) && defined(HAVE_RDTSCLL)
  mti->cycles.routine= MY_TIMER_ROUTINE_RDTSCLL;
#elif defined(_WIN32) && defined(_M_IX86)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_X86_WIN;
#elif defined(_WIN64) && defined(_M_X64)
  mti->cycles.routine= MY_TIMER_ROUTINE_RDTSC;
#elif defined(__INTEL_COMPILER) && defined(__ia64__) && defined(HAVE_IA64INTRIN_H)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_IA64;
#elif defined(__GNUC__) && defined(__ia64__)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_IA64;
#elif defined(__GNUC__) && (defined(__powerpc__) || defined(__POWERPC__) || (defined(_POWER) && defined(_AIX52))) && (defined(__64BIT__) || defined(_ARCH_PPC64))
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_PPC64;
#elif defined(__GNUC__) && (defined(__powerpc__) || defined(__POWERPC__) || (defined(_POWER) && defined(_AIX52))) && (!defined(__64BIT__) && !defined(_ARCH_PPC64))
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_PPC;
#elif (defined(__SUNPRO_CC) || defined(__SUNPRO_C)) && defined(__sparcv9) && defined(_LP64) && !defined(__SunOS_5_7)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_SUNPRO_SPARC64;
#elif (defined(__SUNPRO_CC) || defined(__SUNPRO_C)) && defined(_ILP32) && !defined(__SunOS_5_7)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_SUNPRO_SPARC32;
#elif (defined(__SUNPRO_CC) || defined(__SUNPRO_C)) && defined(__i386) && defined(_ILP32)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_SUNPRO_I386;
#elif (defined(__SUNPRO_CC) || defined(__SUNPRO_C)) && defined(__x86_64) && defined(_LP64)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_SUNPRO_X86_64;
#elif defined(__GNUC__) && defined(__sparcv9) && defined(_LP64) && (__GNUC__>2)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_GCC_SPARC64;
#elif defined(__GNUC__) && defined(__sparc__) && !defined(_LP64) && (__GNUC__>2)
  mti->cycles.routine= MY_TIMER_ROUTINE_ASM_GCC_SPARC32;
#elif defined(__sgi) && defined(HAVE_CLOCK_GETTIME) && defined(CLOCK_SGI_CYCLE)
  mti->cycles.routine= MY_TIMER_ROUTINE_SGI_CYCLE;
#elif defined(HAVE_SYS_TIMES_H) && defined(HAVE_GETHRTIME)
  mti->cycles.routine= MY_TIMER_ROUTINE_GETHRTIME;
#else
  mti->cycles.routine= 0;
#endif

  if (!mti->cycles.routine || !my_timer_cycles())
  {
    mti->cycles.routine= 0;
    mti->cycles.resolution= 0;
    mti->cycles.frequency= 0;
    mti->cycles.overhead= 0;
  }

  /* nanoseconds */
  mti->nanoseconds.frequency=  1000000000; /* initial assumption */
#if defined(HAVE_READ_REAL_TIME)
  mti->nanoseconds.routine= MY_TIMER_ROUTINE_READ_REAL_TIME;
#elif defined(HAVE_SYS_TIMES_H) && defined(HAVE_GETHRTIME)
  mti->nanoseconds.routine= MY_TIMER_ROUTINE_GETHRTIME;
#elif defined(HAVE_CLOCK_GETTIME)
  mti->nanoseconds.routine= MY_TIMER_ROUTINE_CLOCK_GETTIME;
#elif defined(__APPLE__) && defined(__MACH__)
  mti->nanoseconds.routine= MY_TIMER_ROUTINE_MACH_ABSOLUTE_TIME;
#else
  mti->nanoseconds.routine= 0;
#endif
  if (!mti->nanoseconds.routine || !my_timer_nanoseconds())
  {
    mti->nanoseconds.routine= 0;
    mti->nanoseconds.resolution= 0;
    mti->nanoseconds.frequency= 0;
    mti->nanoseconds.overhead= 0;
  }

  /* microseconds */
  mti->microseconds.frequency= 1000000; /* initial assumption */
#if defined(HAVE_GETTIMEOFDAY)
   mti->microseconds.routine= MY_TIMER_ROUTINE_GETTIMEOFDAY;
#elif defined(_WIN32)
  {
    LARGE_INTEGER li;
    /* Windows: typical frequency = 3579545, actually 1/3 microsecond. */
    if (!QueryPerformanceFrequency(&li))
      mti->microseconds.routine= 0;
    else
    {
      mti->microseconds.frequency= li.QuadPart;
      mti->microseconds.routine= MY_TIMER_ROUTINE_QUERYPERFORMANCECOUNTER;
    }
  }
#else
  mti->microseconds.routine= 0;
#endif
  if (!mti->microseconds.routine || !my_timer_microseconds())
  {
    mti->microseconds.routine= 0;
    mti->microseconds.resolution= 0;
    mti->microseconds.frequency= 0;
    mti->microseconds.overhead= 0;
  }

  /* milliseconds */
  mti->milliseconds.frequency= 1000; /* initial assumption */
#if defined(HAVE_SYS_TIMEB_H) && defined(HAVE_FTIME)
  mti->milliseconds.routine= MY_TIMER_ROUTINE_FTIME;
#elif defined(_WIN32)
  mti->milliseconds.routine= MY_TIMER_ROUTINE_GETSYSTEMTIMEASFILETIME;
#elif defined(HAVE_TIME)
  mti->milliseconds.routine= MY_TIMER_ROUTINE_TIME;
#else
  mti->milliseconds.routine= 0;
#endif
  if (!mti->milliseconds.routine || !my_timer_milliseconds())
  {
    mti->milliseconds.routine= 0;
    mti->milliseconds.resolution= 0;
    mti->milliseconds.frequency= 0;
    mti->milliseconds.overhead= 0;
  }

  /* ticks */
  mti->ticks.frequency= 100; /* permanent assumption */
#if defined(HAVE_SYS_TIMES_H) && defined(HAVE_TIMES)
  mti->ticks.routine= MY_TIMER_ROUTINE_TIMES;
#elif defined(_WIN32)
  mti->ticks.routine= MY_TIMER_ROUTINE_GETTICKCOUNT;
#else
  mti->ticks.routine= 0;
#endif
  if (!mti->ticks.routine || !my_timer_ticks())
  {
    mti->ticks.routine= 0;
    mti->ticks.resolution= 0;
    mti->ticks.frequency= 0;
    mti->ticks.overhead= 0;
  }

  /*
    Calculate overhead in terms of the timer that
    gives the best resolution: cycles or nanoseconds.
    I doubt it ever will be as bad as microseconds.
  */
  if (mti->cycles.routine)
    best_timer= &my_timer_cycles;
  else
  {
    if (mti->nanoseconds.routine)
    {
      best_timer= &my_timer_nanoseconds;
    }
    else
      best_timer= &my_timer_microseconds;
  }

  /* best_timer_overhead = least of 20 calculations */
  for (i= 0, best_timer_overhead= 1000000000; i < 20; ++i)
  {
    time1= best_timer();
    time2= best_timer() - time1;
    if (best_timer_overhead > time2)
      best_timer_overhead= time2;
  }
  if (mti->cycles.routine)
    my_timer_init_overhead(&mti->cycles.overhead,
                           best_timer,
                           &my_timer_cycles,
                           best_timer_overhead);
  if (mti->nanoseconds.routine)
    my_timer_init_overhead(&mti->nanoseconds.overhead,
                           best_timer,
                           &my_timer_nanoseconds,
                           best_timer_overhead);
  if (mti->microseconds.routine)
    my_timer_init_overhead(&mti->microseconds.overhead,
                           best_timer,
                           &my_timer_microseconds,
                           best_timer_overhead);
  if (mti->milliseconds.routine)
    my_timer_init_overhead(&mti->milliseconds.overhead,
                           best_timer,
                           &my_timer_milliseconds,
                           best_timer_overhead);
  if (mti->ticks.routine)
    my_timer_init_overhead(&mti->ticks.overhead,
                           best_timer,
                           &my_timer_ticks,
                           best_timer_overhead);

/*
  Calculate resolution for nanoseconds or microseconds
  or milliseconds, by seeing if it's always divisible
  by 1000, and by noticing how much jumping occurs.
  For ticks, just assume the resolution is 1.
*/
  if (mti->cycles.routine)
    mti->cycles.resolution= 1;
  if (mti->nanoseconds.routine)
    mti->nanoseconds.resolution=
    my_timer_init_resolution(&my_timer_nanoseconds, 20000);
  if (mti->microseconds.routine)
    mti->microseconds.resolution=
    my_timer_init_resolution(&my_timer_microseconds, 20);
  if (mti->milliseconds.routine)
  {
    if (mti->milliseconds.routine == MY_TIMER_ROUTINE_TIME)
      mti->milliseconds.resolution= 1000;
    else
      mti->milliseconds.resolution=
      my_timer_init_resolution(&my_timer_milliseconds, 0);
  }
  if (mti->ticks.routine)
    mti->ticks.resolution= 1;

/*
  Calculate cycles frequency,
  if we have both a cycles routine and a microseconds routine.
  In tests, this usually results in a figure within 2% of
  what "cat /proc/cpuinfo" says.
  If the microseconds routine is QueryPerformanceCounter
  (i.e. it's Windows), and the microseconds frequency is >
  500,000,000 (i.e. it's Windows Server so it uses RDTSC)
  and the microseconds resolution is > 100 (i.e. dreadful),
  then calculate cycles frequency = microseconds frequency.
*/
  if (mti->cycles.routine
  &&  mti->microseconds.routine)
  {
    if (mti->microseconds.routine ==
    MY_TIMER_ROUTINE_QUERYPERFORMANCECOUNTER
    &&  mti->microseconds.frequency > 500000000
    &&  mti->microseconds.resolution > 100)
      mti->cycles.frequency= mti->microseconds.frequency;
    else
    {
      time1= my_timer_init_frequency(mti);
      /* Repeat once in case there was an interruption. */
      time2= my_timer_init_frequency(mti);
      if (time1 < time2) mti->cycles.frequency= time1;
      else mti->cycles.frequency= time2;
    }
  }

/*
  Calculate milliseconds frequency =
  (cycles-frequency/#-of-cycles) * #-of-milliseconds,
  if we have both a milliseconds routine and a cycles
  routine.
  This will be inaccurate if milliseconds resolution > 1.
  This is probably only useful when testing new platforms.
*/
  if (mti->milliseconds.routine
  &&  mti->milliseconds.resolution < 1000
  &&  mti->microseconds.routine
  &&  mti->cycles.routine)
  {
    ulonglong time3, time4;
    time1= my_timer_cycles();
    time2= my_timer_milliseconds();
    time3= time2; /* Avoids a Microsoft/IBM compiler warning */
    for (i= 0; i < MY_TIMER_ITERATIONS * 1000; ++i)
    {
      time3= my_timer_milliseconds();
      if (time3 - time2 > 10) break;
    }
    time4= my_timer_cycles();
    mti->milliseconds.frequency=
    (mti->cycles.frequency * (time3 - time2)) / (time4 - time1);
  }

/*
  Calculate ticks.frequency =
  (cycles-frequency/#-of-cycles * #-of-ticks,
  if we have both a ticks routine and a cycles
  routine,
  This is probably only useful when testing new platforms.
*/
  if (mti->ticks.routine
  &&  mti->microseconds.routine
  &&  mti->cycles.routine)
  {
    ulonglong time3, time4;
    time1= my_timer_cycles();
    time2= my_timer_ticks();
    time3= time2; /* Avoids a Microsoft/IBM compiler warning */
    for (i= 0; i < MY_TIMER_ITERATIONS * 1000; ++i)
    {
      time3= my_timer_ticks();
      if (time3 - time2 > 10) break;
    }
    time4= my_timer_cycles();
    mti->ticks.frequency=
    (mti->cycles.frequency * (time3 - time2)) / (time4 - time1);
  }
}

/*
   Additional Comments
   -------------------

   This is for timing, i.e. finding out how long a piece of code
   takes. If you want time of day matching a wall clock, the
   my_timer_xxx functions won't help you.

   The best timer is the one with highest frequency, lowest
   overhead, and resolution=1. The my_timer_info() routine will tell
   you at runtime which timer that is. Usually it will be
   my_timer_cycles() but be aware that, although it's best,
   it has possible flaws and dangers. Depending on platform:
   - The frequency might change. We don't test for this. It
     happens on laptops for power saving, and on blade servers
     for avoiding overheating.
   - The overhead that my_timer_init() returns is the minimum.
     In fact it could be slightly greater because of caching or
     because you call the routine by address, as recommended.
     It could be hugely greater if there's an interrupt.
   - The x86 cycle counter, RDTSC doesn't "serialize". That is,
     if there is out-of-order execution, rdtsc might be processed
     after an instruction that logically follows it.
     (We could force serialization, but that would be slower.)
   - It is possible to set a flag which renders RDTSC
     inoperative. Somebody responsible for the kernel
     of the operating system would have to make this
     decision. For the platforms we've tested with, there's
     no such problem.
   - With a multi-processor arrangement, it's possible
     to get the cycle count from one processor in
     thread X, and the cycle count from another processor
     in thread Y. They may not always be in synch.
   - You can't depend on a cycle counter being available for
     all platforms. On Alphas, the
     cycle counter is only 32-bit, so it would overflow quickly,
     so we don't bother with it. On platforms that we haven't
     tested, there might be some if/endif combination that we
     didn't expect, or some assembler routine that we didn't
     supply.

   The recommended way to use the timer routines is:
   1. Somewhere near the beginning of the program, call
      my_timer_init(). This should only be necessary once,
      although you can call it again if you think that the
      frequency has changed.
   2. Determine the best timer based on frequency, resolution,
      overhead -- all things that my_timer_init() returns.
      Preserve the address of the timer and the my_timer_into
      results in an easily-accessible place.
   3. Instrument the code section that you're monitoring, thus:
      time1= my_timer_xxx();
      Instrumented code;
      time2= my_timer_xxx();
      elapsed_time= (time2 - time1) - overhead;
      If the timer is always on, then overhead is always there,
      so don't subtract it.
   4. Save the elapsed time, or add it to a totaller.
   5. When all timing processes are complete, transfer the
      saved / totalled elapsed time to permanent storage.
      Optionally you can convert cycles to microseconds at
      this point. (Don't do so every time you calculate
      elapsed_time! That would waste time and lose precision!)
      For converting cycles to microseconds, use the frequency
      that my_timer_init() returns. You'll also need to convert
      if the my_timer_microseconds() function is the Windows
      function QueryPerformanceCounter(), since that's sometimes
      a counter with precision slightly better than microseconds.

   Since we recommend calls by function pointer, we supply
   no inline functions.

   Some comments on the many candidate routines for timing ...

   clock() -- We don't use because it would overflow frequently.

   clock_gettime() -- In tests, clock_gettime often had
   resolution = 1000.

   ftime() -- A "man ftime" says: "This function is obsolete.
   Don't use it." On every platform that we tested, if ftime()
   was available, then so was gettimeofday(), and gettimeofday()
   overhead was always at least as good as ftime() overhead.

   gettimeofday() -- available on most platforms, though not
   on Windows. There is a hardware timer (sometimes a Programmable
   Interrupt Timer or "PIT") (sometimes a "HPET") used for
   interrupt generation. When it interrupts (a "tick" or "jiffy",
   typically 1 centisecond) it sets xtime. For gettimeofday, a
   Linux kernel routine usually gets xtime and then gets rdtsc
   to get elapsed nanoseconds since the last tick. On Red Hat
   Enterprise Linux 3, there was once a bug which caused the
   resolution to be 1000, i.e. one centisecond. We never check
   for time-zone change.

   getnstimeofday() -- something to watch for in future Linux

   do_gettimeofday() -- exists on Linux but not for "userland"

   get_cycles() -- a multi-platform function, worth watching
   in future Linux versions. But we found platform-specific
   functions which were better documented in operating-system
   manuals. And get_cycles() can fail or return a useless
   32-bit number. It might be available on some platforms,
   such as arm, which we didn't test.  Using
   "include <linux/timex.h>" or "include <asm/timex.h>"
   can lead to autoconf or compile errors, depending on system.

   rdtsc, __rdtsc, rdtscll: available for x86 with Linux BSD,
   Solaris, Windows. See "possible flaws and dangers" comments.

   times(): what we use for ticks. Should just read the last
   (xtime) tick count, therefore should be fast, but usually
   isn't.

   GetTickCount(): we use this for my_timer_ticks() on
   Windows. Actually it really is a tick counter, so resolution
   >= 10 milliseconds unless you have a very old Windows version.
   With Windows 95 or 98 or ME, timeGetTime() has better resolution than
   GetTickCount (1ms rather than 55ms). But with Windows NT or XP or 2000,
   they're both getting from a variable in the Process Environment Block
   (PEB), and the variable is set by the programmable interrupt timer, so
   the resolution is the same (usually 10-15 milliseconds). Also timeGetTime
   is slower on old machines:
   http://www.doumo.jp/aon-java/jsp/postgretips/tips.jsp?tips=74.
   Also timeGetTime requires linking winmm.lib,
   Therefore we use GetTickCount.
   It will overflow every 49 days because the return is 32-bit.
   There is also a GetTickCount64 but it requires Vista or Windows Server 2008.
   (As for GetSystemTimeAsFileTime, its precision is spurious, it
   just reads the tick variable like the other functions do.
   However, we don't expect it to overflow every 49 days, so we
   will prefer it for my_timer_milliseconds().)

   QueryPerformanceCounter() we use this for my_timer_microseconds()
   on Windows. 1-PIT-tick (often 1/3-microsecond). Usually reads
   the PIT so it's slow. On some Windows variants, uses RDTSC.

   GetLocalTime() this is available on Windows but we don't use it.

   getclock(): documented for Alpha, but not found during tests.

   mach_absolute_time() and UpTime() are recommended for Apple.
   Inititally they weren't tried, because asm_ppc seems to do the job.
   But now we use mach_absolute_time for nanoseconds.

   Any clock-based timer can be affected by NPT (ntpd program),
   which means:
   - full-second correction can occur for leap second
   - tiny corrections can occcur approimately every 11 minutes
     (but I think they only affect the RTC which isn't the PIT).

   We define "precision" as "frequency" and "high precision" is
   "frequency better than 1 microsecond". We define "resolution"
   as a synonym for "granularity". We define "accuracy" as
   "closeness to the truth" as established by some authoritative
   clock, but we can't measure accuracy.

   Do not expect any of our timers to be monotonic; we
   won't guarantee that they return constantly-increasing
   unique numbers.

   We tested with AIX, Solaris (x86 + Sparc), Linux (x86 +
   Itanium), Windows, 64-bit Windows, QNX, FreeBSD, HPUX,
   Irix, Mac. We didn't test with SCO.

*/