1// ----------------------------------------------------------------------
2// CycleClock
3// A CycleClock tells you the current time in Cycles. The "time"
4// is actually time since power-on. This is like time() but doesn't
5// involve a system call and is much more precise.
6//
7// NOTE: Not all cpu/platform/kernel combinations guarantee that this
8// clock increments at a constant rate or is synchronized across all logical
9// cpus in a system.
10//
11// If you need the above guarantees, please consider using a different
12// API. There are efforts to provide an interface which provides a millisecond
13// granularity and implemented as a memory read. A memory read is generally
14// cheaper than the CycleClock for many architectures.
15//
16// Also, in some out of order CPU implementations, the CycleClock is not
17// serializing. So if you're trying to count at cycles granularity, your
18// data might be inaccurate due to out of order instruction execution.
19// ----------------------------------------------------------------------
20
21#ifndef BENCHMARK_CYCLECLOCK_H_
22#define BENCHMARK_CYCLECLOCK_H_
23
24#include <cstdint>
25
26#include "benchmark/macros.h"
27#include "internal_macros.h"
28
29#if defined(BENCHMARK_OS_MACOSX)
30#include <mach/mach_time.h>
31#endif
32// For MSVC, we want to use '_asm rdtsc' when possible (since it works
33// with even ancient MSVC compilers), and when not possible the
34// __rdtsc intrinsic, declared in <intrin.h>. Unfortunately, in some
35// environments, <windows.h> and <intrin.h> have conflicting
36// declarations of some other intrinsics, breaking compilation.
37// Therefore, we simply declare __rdtsc ourselves. See also
38// http://connect.microsoft.com/VisualStudio/feedback/details/262047
39#if defined(COMPILER_MSVC) && !defined(_M_IX86)
40extern "C" uint64_t __rdtsc();
41#pragma intrinsic(__rdtsc)
42#endif
43
44#ifndef BENCHMARK_OS_WINDOWS
45#include <sys/time.h>
46#endif
47
48namespace benchmark {
49// NOTE: only i386 and x86_64 have been well tested.
50// PPC, sparc, alpha, and ia64 are based on
51// http://peter.kuscsik.com/wordpress/?p=14
52// with modifications by m3b. See also
53// https://setisvn.ssl.berkeley.edu/svn/lib/fftw-3.0.1/kernel/cycle.h
54namespace cycleclock {
55// This should return the number of cycles since power-on. Thread-safe.
56inline BENCHMARK_ALWAYS_INLINE int64_t Now() {
57#if defined(BENCHMARK_OS_MACOSX)
58 // this goes at the top because we need ALL Macs, regardless of
59 // architecture, to return the number of "mach time units" that
60 // have passed since startup. See sysinfo.cc where
61 // InitializeSystemInfo() sets the supposed cpu clock frequency of
62 // macs to the number of mach time units per second, not actual
63 // CPU clock frequency (which can change in the face of CPU
64 // frequency scaling). Also note that when the Mac sleeps, this
65 // counter pauses; it does not continue counting, nor does it
66 // reset to zero.
67 return mach_absolute_time();
68#elif defined(__i386__)
69 int64_t ret;
70 __asm__ volatile("rdtsc" : "=A"(ret));
71 return ret;
72#elif defined(__x86_64__) || defined(__amd64__)
73 uint64_t low, high;
74 __asm__ volatile("rdtsc" : "=a"(low), "=d"(high));
75 return (high << 32) | low;
76#elif defined(__powerpc__) || defined(__ppc__)
77 // This returns a time-base, which is not always precisely a cycle-count.
78 int64_t tbl, tbu0, tbu1;
79 asm("mftbu %0" : "=r"(tbu0));
80 asm("mftb %0" : "=r"(tbl));
81 asm("mftbu %0" : "=r"(tbu1));
82 tbl &= -static_cast<int64>(tbu0 == tbu1);
83 // high 32 bits in tbu1; low 32 bits in tbl (tbu0 is garbage)
84 return (tbu1 << 32) | tbl;
85#elif defined(__sparc__)
86 int64_t tick;
87 asm(".byte 0x83, 0x41, 0x00, 0x00");
88 asm("mov %%g1, %0" : "=r"(tick));
89 return tick;
90#elif defined(__ia64__)
91 int64_t itc;
92 asm("mov %0 = ar.itc" : "=r"(itc));
93 return itc;
94#elif defined(COMPILER_MSVC) && defined(_M_IX86)
95 // Older MSVC compilers (like 7.x) don't seem to support the
96 // __rdtsc intrinsic properly, so I prefer to use _asm instead
97 // when I know it will work. Otherwise, I'll use __rdtsc and hope
98 // the code is being compiled with a non-ancient compiler.
99 _asm rdtsc
100#elif defined(COMPILER_MSVC)
101 return __rdtsc();
102#elif defined(__aarch64__)
103 // System timer of ARMv8 runs at a different frequency than the CPU's.
104 // The frequency is fixed, typically in the range 1-50MHz. It can be
105 // read at CNTFRQ special register. We assume the OS has set up
106 // the virtual timer properly.
107 int64_t virtual_timer_value;
108 asm volatile("mrs %0, cntvct_el0" : "=r"(virtual_timer_value));
109 return virtual_timer_value;
110#elif defined(__ARM_ARCH)
111#if (__ARM_ARCH >= 6) // V6 is the earliest arch that has a standard cyclecount
112 uint32_t pmccntr;
113 uint32_t pmuseren;
114 uint32_t pmcntenset;
115 // Read the user mode perf monitor counter access permissions.
116 asm volatile("mrc p15, 0, %0, c9, c14, 0" : "=r"(pmuseren));
117 if (pmuseren & 1) { // Allows reading perfmon counters for user mode code.
118 asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r"(pmcntenset));
119 if (pmcntenset & 0x80000000ul) { // Is it counting?
120 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(pmccntr));
121 // The counter is set up to count every 64th cycle
122 return static_cast<int64_t>(pmccntr) * 64; // Should optimize to << 6
123 }
124 }
125#endif
126 struct timeval tv;
127 gettimeofday(&tv, nullptr);
128 return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
129#elif defined(__mips__)
130 // mips apparently only allows rdtsc for superusers, so we fall
131 // back to gettimeofday. It's possible clock_gettime would be better.
132 struct timeval tv;
133 gettimeofday(&tv, nullptr);
134 return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
135#else
136// The soft failover to a generic implementation is automatic only for ARM.
137// For other platforms the developer is expected to make an attempt to create
138// a fast implementation and use generic version if nothing better is available.
139#error You need to define CycleTimer for your OS and CPU
140#endif
141}
142} // end namespace cycleclock
143} // end namespace benchmark
144
145#endif // BENCHMARK_CYCLECLOCK_H_
146