1 /* High precision, low overhead timing functions. powerpc64 version.
2 Copyright (C) 2005, 2008 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4 Contributed by Ulrich Drepper <drepper@cygnus.com>, 1998.
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.
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.
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
22 #define _HP_TIMING_H 1
25 #include <sys/param.h>
26 #include <stdio-common/_itoa.h>
29 /* The macros defined here use the powerpc 64-bit time base register.
30 The time base is nominally clocked at 1/8th the CPU clock, but this
33 The list of macros we need includes the following:
35 - HP_TIMING_AVAIL: test for availability.
37 - HP_TIMING_INLINE: this macro is non-zero if the functionality is not
38 implemented using function calls but instead uses some inlined code
39 which might simply consist of a few assembler instructions. We have to
40 know this since we might want to use the macros here in places where we
41 cannot make function calls.
43 - hp_timing_t: This is the type for variables used to store the time
46 - HP_TIMING_ZERO: clear `hp_timing_t' object.
48 - HP_TIMING_NOW: place timestamp for current time in variable given as
51 - HP_TIMING_DIFF_INIT: do whatever is necessary to be able to use the
54 - HP_TIMING_DIFF: compute difference between two times and store it
55 in a third. Source and destination might overlap.
57 - HP_TIMING_ACCUM: add time difference to another variable. This might
58 be a bit more complicated to implement for some platforms as the
59 operation should be thread-safe and 64bit arithmetic on 32bit platforms
62 - HP_TIMING_ACCUM_NT: this is the variant for situations where we know
63 there are no threads involved.
65 - HP_TIMING_PRINT: write decimal representation of the timing value into
66 the given string. This operation need not be inline even though
67 HP_TIMING_INLINE is specified.
71 /* We always assume having the timestamp register. */
72 #define HP_TIMING_AVAIL (1)
74 /* We indeed have inlined functions. */
75 #define HP_TIMING_INLINE (1)
77 /* We use 64bit values for the times. */
78 typedef unsigned long long int hp_timing_t
;
80 /* Set timestamp value to zero. */
81 #define HP_TIMING_ZERO(Var) (Var) = (0)
83 /* That's quite simple. Use the `mftb' instruction. Note that the value
84 might not be 100% accurate since there might be some more instructions
85 running in this moment. This could be changed by using a barrier like
86 'lwsync' right before the `mftb' instruciton. But we are not interested
87 in accurate clock cycles here so we don't do this. */
89 #define HP_TIMING_NOW(Var) __asm__ __volatile__ ("mfspr %0,268" : "=r" (Var))
91 #define HP_TIMING_NOW(Var) __asm__ __volatile__ ("mftb %0" : "=r" (Var))
94 /* Use two 'mftb' instructions in a row to find out how long it takes.
95 On current POWER4, POWER5, and 970 processors mftb take ~10 cycles. */
96 #define HP_TIMING_DIFF_INIT() \
98 if (GLRO(dl_hp_timing_overhead) == 0) \
101 GLRO(dl_hp_timing_overhead) = ~0ull; \
104 hp_timing_t __t1, __t2; \
105 HP_TIMING_NOW (__t1); \
106 HP_TIMING_NOW (__t2); \
107 if (__t2 - __t1 < GLRO(dl_hp_timing_overhead)) \
108 GLRO(dl_hp_timing_overhead) = __t2 - __t1; \
110 while (--__cnt > 0); \
114 /* It's simple arithmetic in 64-bit. */
115 #define HP_TIMING_DIFF(Diff, Start, End) (Diff) = ((End) - (Start))
117 /* We need to insure that this add is atomic in threaded environments. We use
118 __arch_atomic_exchange_and_add_64 from atomic.h to get thread safety. */
119 #define HP_TIMING_ACCUM(Sum, Diff) \
121 hp_timing_t __diff = (Diff) - GLRO(dl_hp_timing_overhead); \
122 __arch_atomic_exchange_and_add_64 (&(Sum), __diff); \
125 /* No threads, no extra work. */
126 #define HP_TIMING_ACCUM_NT(Sum, Diff) (Sum) += (Diff)
128 /* Print the time value. */
129 #define HP_TIMING_PRINT(Buf, Len, Val) \
132 char *__cp = _itoa (Val, __buf + sizeof (__buf), 10, 0); \
133 size_t __len = (Len); \
134 char *__dest = (Buf); \
135 while (__len-- > 0 && __cp < __buf + sizeof (__buf)) \
136 *__dest++ = *__cp++; \
137 memcpy (__dest, " ticks", MIN (__len, sizeof (" ticks"))); \
140 #endif /* hp-timing.h */