sysdeps/ia64/hp-timing.h

   1 /* High precision, low overhead timing functions.  IA-64 version.
   2    Copyright (C) 2001 Free Software Foundation, Inc.
   3    This file is part of the GNU C Library.
   4    Contributed by Ulrich Drepper <drepper@cygnus.com>, 2001.
   5
   6    The GNU C Library is free software; you can redistribute it and/or
   7    modify it under the terms of the GNU Lesser General Public
   8    License as published by the Free Software Foundation; either
   9    version 2.1 of the License, or (at your option) any later version.
  10
  11    The GNU C Library is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14    Lesser General Public License for more details.
  15
  16    You should have received a copy of the GNU Lesser General Public
  17    License along with the GNU C Library; if not, write to the Free
  18    Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
  19    02111-1307 USA.  */
  20
  21 #ifndef _HP_TIMING_H
  22 #define _HP_TIMING_H    1
  23
  24 #include <string.h>
  25 #include <sys/param.h>
  26 #include <stdio-common/_itoa.h>
  27 #include <ia64intrin.h>
  28
  29 /* The macros defined here use the timestamp counter in IA-64.  They
  30    provide a very accurate way to measure the time with very little
  31    overhead.  The time values themself have no real meaning, only
  32    differences are interesting.
  33
  34    The list of macros we need includes the following:
  35
  36    - HP_TIMING_AVAIL: test for availability.
  37
  38    - HP_TIMING_INLINE: this macro is non-zero if the functionality is not
  39      implemented using function calls but instead uses some inlined code
  40      which might simply consist of a few assembler instructions.  We have to
  41      know this since we might want to use the macros here in places where we
  42      cannot make function calls.
  43
  44    - hp_timing_t: This is the type for variables used to store the time
  45      values.
  46
  47    - HP_TIMING_ZERO: clear `hp_timing_t' object.
  48
  49    - HP_TIMING_NOW: place timestamp for current time in variable given as
  50      parameter.
  51
  52    - HP_TIMING_DIFF_INIT: do whatever is necessary to be able to use the
  53      HP_TIMING_DIFF macro.
  54
  55    - HP_TIMING_DIFF: compute difference between two times and store it
  56      in a third.  Source and destination might overlap.
  57
  58    - HP_TIMING_ACCUM: add time difference to another variable.  This might
  59      be a bit more complicated to implement for some platforms as the
  60      operation should be thread-safe and 64bit arithmetic on 32bit platforms
  61      is not.
  62
  63    - HP_TIMING_ACCUM_NT: this is the variant for situations where we know
  64      there are no threads involved.
  65
  66    - HP_TIMING_PRINT: write decimal representation of the timing value into
  67      the given string.  This operation need not be inline even though
  68      HP_TIMING_INLINE is specified.
  69
  70 */
  71
  72 /* We always assume having the timestamp register.  */
  73 #define HP_TIMING_AVAIL         (1)
  74
  75 /* We indeed have inlined functions.  */
  76 #define HP_TIMING_INLINE        (1)
  77
  78 /* We use 64bit values for the times.  */
  79 typedef unsigned long int hp_timing_t;
  80
  81 /* Internal variable used to store the overhead of the measurement
  82    opcodes.  */
  83 extern hp_timing_t __libc_hp_timing_overhead;
  84
  85 /* Set timestamp value to zero.  */
  86 #define HP_TIMING_ZERO(Var)     (Var) = (0)
  87
  88
  89 /* The Itanium/Merced has a bug where the ar.itc register value read
  90    is not correct in some situations.  The solution is to read again.
  91    For now we always do this until we know how to recognize a fixed
  92    processor implementation.  */
  93 #define REPEAT_READ(val) __builtin_expect ((int) val == -1, 0)
  94
  95 /* That's quite simple.  Use the `rdtsc' instruction.  Note that the value
  96    might not be 100% accurate since there might be some more instructions
  97    running in this moment.  This could be changed by using a barrier like
  98    'cpuid' right before the `rdtsc' instruciton.  But we are not interested
  99    in accurate clock cycles here so we don't do this.  */
 100 #define HP_TIMING_NOW(Var) \
 101   ({ unsigned long int __itc;                                                 \
 102      do                                                                       \
 103        asm volatile ("mov %0=ar.itc" : "=r" (__itc) : : "memory");            \
 104      while (REPEAT_READ (__itc));                                             \
 105      Var = __itc; })
 106
 107 /* Use two 'ar.itc' instructions in a row to find out how long it takes.  */
 108 #define HP_TIMING_DIFF_INIT() \
 109   do {                                                                        \
 110     int __cnt = 5;                                                            \
 111     __libc_hp_timing_overhead = ~0ul;                                         \
 112     do                                                                        \
 113       {                                                                       \
 114         hp_timing_t __t1, __t2;                                               \
 115         HP_TIMING_NOW (__t1);                                                 \
 116         HP_TIMING_NOW (__t2);                                                 \
 117         if (__t2 - __t1 < __libc_hp_timing_overhead)                          \
 118           __libc_hp_timing_overhead = __t2 - __t1;                            \
 119       }                                                                       \
 120     while (--__cnt > 0);                                                      \
 121   } while (0)
 122
 123 /* It's simple arithmetic for us.  */
 124 #define HP_TIMING_DIFF(Diff, Start, End)        (Diff) = ((End) - (Start))
 125
 126 /* We have to jump through hoops to get this correctly implemented.  */
 127 #define HP_TIMING_ACCUM(Sum, Diff) \
 128   do {                                                                        \
 129     hp_timing_t __oldval;                                                     \
 130     hp_timing_t __diff = (Diff) - __libc_hp_timing_overhead;                  \
 131     hp_timing_t __newval;                                                     \
 132     do                                                                        \
 133       {                                                                       \
 134         __oldval = (Sum);                                                     \
 135         __newval = __oldval + __diff;                                         \
 136       }                                                                       \
 137     while (! __sync_bool_compare_and_swap (&Sum, __oldvar, __newval));        \
 138   } while (0)
 139
 140 /* No threads, no extra work.  */
 141 #define HP_TIMING_ACCUM_NT(Sum, Diff)   (Sum) += (Diff)
 142
 143 /* Print the time value.  */
 144 #define HP_TIMING_PRINT(Buf, Len, Val) \
 145   do {                                                                        \
 146     char __buf[20];                                                           \
 147     char *__cp = _itoa_word (Val, __buf + sizeof (__buf), 10, 0);             \
 148     int __len = (Len);                                                        \
 149     char *__dest = (Buf);                                                     \
 150     while (__len-- > 0 && __cp < __buf + sizeof (__buf))                      \
 151       *__dest++ = *__cp++;                                                    \
 152     memcpy (__dest, " clock cycles", MIN (__len, sizeof (" clock cycles")));  \
 153   } while (0)
 154
 155 #endif  /* hp-timing.h */