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/benchmark.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) && !defined(_M_ARM64)
40 extern "C" uint64_t __rdtsc();
41 #pragma intrinsic(__rdtsc)
42 #endif
43
44 #if !defined(BENCHMARK_OS_WINDOWS) || defined(BENCHMARK_OS_MINGW)
45 #include <sys/time.h>
46 #include <time.h>
47 #endif
48
49 #ifdef BENCHMARK_OS_EMSCRIPTEN
50 #include <emscripten.h>
51 #endif
52
53 namespace benchmark {
54 // NOTE: only i386 and x86_64 have been well tested.
55 // PPC, sparc, alpha, and ia64 are based on
56 // http://peter.kuscsik.com/wordpress/?p=14
57 // with modifications by m3b. See also
58 // https://setisvn.ssl.berkeley.edu/svn/lib/fftw-3.0.1/kernel/cycle.h
59 namespace cycleclock {
60 // This should return the number of cycles since power-on. Thread-safe.
Now()61 inline BENCHMARK_ALWAYS_INLINE int64_t Now() {
62 #if defined(BENCHMARK_OS_MACOSX)
63 // this goes at the top because we need ALL Macs, regardless of
64 // architecture, to return the number of "mach time units" that
65 // have passed since startup. See sysinfo.cc where
66 // InitializeSystemInfo() sets the supposed cpu clock frequency of
67 // macs to the number of mach time units per second, not actual
68 // CPU clock frequency (which can change in the face of CPU
69 // frequency scaling). Also note that when the Mac sleeps, this
70 // counter pauses; it does not continue counting, nor does it
71 // reset to zero.
72 return mach_absolute_time();
73 #elif defined(BENCHMARK_OS_EMSCRIPTEN)
74 // this goes above x86-specific code because old versions of Emscripten
75 // define __x86_64__, although they have nothing to do with it.
76 return static_cast<int64_t>(emscripten_get_now() * 1e+6);
77 #elif defined(__i386__)
78 int64_t ret;
79 __asm__ volatile("rdtsc" : "=A"(ret));
80 return ret;
81 #elif defined(__x86_64__) || defined(__amd64__)
82 uint64_t low, high;
83 __asm__ volatile("rdtsc" : "=a"(low), "=d"(high));
84 return (high << 32) | low;
85 #elif defined(__powerpc__) || defined(__ppc__)
86 // This returns a time-base, which is not always precisely a cycle-count.
87 #if defined(__powerpc64__) || defined(__ppc64__)
88 int64_t tb;
89 asm volatile("mfspr %0, 268" : "=r"(tb));
90 return tb;
91 #else
92 uint32_t tbl, tbu0, tbu1;
93 asm volatile(
94 "mftbu %0\n"
95 "mftb %1\n"
96 "mftbu %2"
97 : "=r"(tbu0), "=r"(tbl), "=r"(tbu1));
98 tbl &= -static_cast<int32_t>(tbu0 == tbu1);
99 // high 32 bits in tbu1; low 32 bits in tbl (tbu0 is no longer needed)
100 return (static_cast<uint64_t>(tbu1) << 32) | tbl;
101 #endif
102 #elif defined(__sparc__)
103 int64_t tick;
104 asm(".byte 0x83, 0x41, 0x00, 0x00");
105 asm("mov %%g1, %0" : "=r"(tick));
106 return tick;
107 #elif defined(__ia64__)
108 int64_t itc;
109 asm("mov %0 = ar.itc" : "=r"(itc));
110 return itc;
111 #elif defined(COMPILER_MSVC) && defined(_M_IX86)
112 // Older MSVC compilers (like 7.x) don't seem to support the
113 // __rdtsc intrinsic properly, so I prefer to use _asm instead
114 // when I know it will work. Otherwise, I'll use __rdtsc and hope
115 // the code is being compiled with a non-ancient compiler.
116 _asm rdtsc
117 #elif defined(COMPILER_MSVC) && defined(_M_ARM64)
118 // See // https://docs.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics
119 // and https://reviews.llvm.org/D53115
120 int64_t virtual_timer_value;
121 virtual_timer_value = _ReadStatusReg(ARM64_CNTVCT);
122 return virtual_timer_value;
123 #elif defined(COMPILER_MSVC)
124 return __rdtsc();
125 #elif defined(BENCHMARK_OS_NACL)
126 // Native Client validator on x86/x86-64 allows RDTSC instructions,
127 // and this case is handled above. Native Client validator on ARM
128 // rejects MRC instructions (used in the ARM-specific sequence below),
129 // so we handle it here. Portable Native Client compiles to
130 // architecture-agnostic bytecode, which doesn't provide any
131 // cycle counter access mnemonics.
132
133 // Native Client does not provide any API to access cycle counter.
134 // Use clock_gettime(CLOCK_MONOTONIC, ...) instead of gettimeofday
135 // because is provides nanosecond resolution (which is noticable at
136 // least for PNaCl modules running on x86 Mac & Linux).
137 // Initialize to always return 0 if clock_gettime fails.
138 struct timespec ts = {0, 0};
139 clock_gettime(CLOCK_MONOTONIC, &ts);
140 return static_cast<int64_t>(ts.tv_sec) * 1000000000 + ts.tv_nsec;
141 #elif defined(__aarch64__)
142 // System timer of ARMv8 runs at a different frequency than the CPU's.
143 // The frequency is fixed, typically in the range 1-50MHz. It can be
144 // read at CNTFRQ special register. We assume the OS has set up
145 // the virtual timer properly.
146 int64_t virtual_timer_value;
147 asm volatile("mrs %0, cntvct_el0" : "=r"(virtual_timer_value));
148 return virtual_timer_value;
149 #elif defined(__ARM_ARCH)
150 // V6 is the earliest arch that has a standard cyclecount
151 // Native Client validator doesn't allow MRC instructions.
152 #if (__ARM_ARCH >= 6)
153 uint32_t pmccntr;
154 uint32_t pmuseren;
155 uint32_t pmcntenset;
156 // Read the user mode perf monitor counter access permissions.
157 asm volatile("mrc p15, 0, %0, c9, c14, 0" : "=r"(pmuseren));
158 if (pmuseren & 1) { // Allows reading perfmon counters for user mode code.
159 asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r"(pmcntenset));
160 if (pmcntenset & 0x80000000ul) { // Is it counting?
161 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(pmccntr));
162 // The counter is set up to count every 64th cycle
163 return static_cast<int64_t>(pmccntr) * 64; // Should optimize to << 6
164 }
165 }
166 #endif
167 struct timeval tv;
168 gettimeofday(&tv, nullptr);
169 return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
170 #elif defined(__mips__) || defined(__m68k__)
171 // mips apparently only allows rdtsc for superusers, so we fall
172 // back to gettimeofday. It's possible clock_gettime would be better.
173 struct timeval tv;
174 gettimeofday(&tv, nullptr);
175 return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
176 #elif defined(__loongarch__)
177 struct timeval tv;
178 gettimeofday(&tv, nullptr);
179 return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
180 #elif defined(__s390__) // Covers both s390 and s390x.
181 // Return the CPU clock.
182 uint64_t tsc;
183 #if defined(BENCHMARK_OS_ZOS) && defined(COMPILER_IBMXL)
184 // z/OS XL compiler HLASM syntax.
185 asm(" stck %0" : "=m"(tsc) : : "cc");
186 #else
187 asm("stck %0" : "=Q"(tsc) : : "cc");
188 #endif
189 return tsc;
190 #elif defined(__riscv) // RISC-V
191 // Use RDCYCLE (and RDCYCLEH on riscv32)
192 #if __riscv_xlen == 32
193 uint32_t cycles_lo, cycles_hi0, cycles_hi1;
194 // This asm also includes the PowerPC overflow handling strategy, as above.
195 // Implemented in assembly because Clang insisted on branching.
196 asm volatile(
197 "rdcycleh %0\n"
198 "rdcycle %1\n"
199 "rdcycleh %2\n"
200 "sub %0, %0, %2\n"
201 "seqz %0, %0\n"
202 "sub %0, zero, %0\n"
203 "and %1, %1, %0\n"
204 : "=r"(cycles_hi0), "=r"(cycles_lo), "=r"(cycles_hi1));
205 return (static_cast<uint64_t>(cycles_hi1) << 32) | cycles_lo;
206 #else
207 uint64_t cycles;
208 asm volatile("rdcycle %0" : "=r"(cycles));
209 return cycles;
210 #endif
211 #elif defined(__e2k__) || defined(__elbrus__)
212 struct timeval tv;
213 gettimeofday(&tv, nullptr);
214 return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
215 #else
216 // The soft failover to a generic implementation is automatic only for ARM.
217 // For other platforms the developer is expected to make an attempt to create
218 // a fast implementation and use generic version if nothing better is available.
219 #error You need to define CycleTimer for your OS and CPU
220 #endif
221 }
222 } // end namespace cycleclock
223 } // end namespace benchmark
224
225 #endif // BENCHMARK_CYCLECLOCK_H_
226