cputime optimize for arm64

[dwh110]

What problem does this PR solve?

Issue Number: n/a

Problem Summary: cpugettime optimize for ARM64

What is changed and the side effects?

Changed:
detail::invariant_cpu_freq value get by CNTFRQ_EL0

Side effects:

• Performance effects:

• Breaking backward compatibility: 无

---

Check List:

• Please make sure your changes are compilable.
• When providing us with a new feature, it is best to add related tests.
• Please follow Contributor Covenant Code of Conduct.

[Copilot]

Pull request overview

This PR aims to optimize cpuwide_time_ns() on ARM64 by using the architectural counter frequency (CNTFRQ_EL0) to populate detail::invariant_cpu_freq, shifting initialization away from per-call logic.

Changes:

• On __aarch64__, route cpuwide_time_ns() to use the cycle-counter path instead of the clock_gettime path.
• Move initialization of detail::invariant_cpu_freq into a pre-main initializer, using CNTFRQ_EL0 on ARM64.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
src/butil/time.h	Changes cpuwide_time_ns() preprocessor gating to include ARM64 in the invariant-counter fast path and removes lazy init recursion.
src/butil/time.cpp	Adds a pre-main initializer to set detail::invariant_cpu_freq, reading CNTFRQ_EL0 on ARM64.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

cpuwide_time_ns() can fall through without returning a value when detail::invariant_cpu_freq is negative (e.g., still -1 before init_invariant_cpu_freq() runs, or if initialization fails and leaves it < 0). This is undefined behavior. Please ensure the function always returns a value for the < 0 case (e.g., treat it like the 0 case and return monotonic_time_ns(), or perform a safe one-time initialization here).

Suggested change

	return monotonic_time_ns();
	return monotonic_time_ns();
	} else {
	// Negative frequency (e.g. uninitialized or failed initialization),
	// fall back to system-wide monotonic time to avoid undefined behavior.
	return monotonic_time_ns();

[Copilot]

Using __attribute__((constructor)) in a C++ TU is compiler-specific and can be problematic for portability/toolchains. This repo already has BAIDU_GLOBAL_INIT() (which avoids the attribute in C++) and many places use pthread_once/function-local statics for one-time init; consider switching to an established mechanism so initialization is reliable across supported compilers.

Suggested change

	int64_t invariant_cpu_freq = -1;
	static void __attribute__((constructor)) init_invariant_cpu_freq() {
	int64_t baseFreq = -1;
	#if defined(__aarch64__)
	__asm__ __volatile__("mrs %0, CNTFRQ_EL0" : "=r"(baseFreq));
	#else
	baseFreq = detail::read_invariant_cpu_frequency();
	#endif
	detail::invariant_cpu_freq = baseFreq;
	}
	static int64_t init_invariant_cpu_freq() {
	int64_t baseFreq = -1;
	#if defined(__aarch64__)
	__asm__ __volatile__("mrs %0, CNTFRQ_EL0" : "=r"(baseFreq));
	#else
	baseFreq = detail::read_invariant_cpu_frequency();
	#endif
	return baseFreq;
	}
	int64_t invariant_cpu_freq = init_invariant_cpu_freq();

[Copilot]

init_invariant_cpu_freq() runs unconditionally at load time and, on non-aarch64, reads /proc/cpuinfo even in builds where cpuwide_time_ns() uses clock_gettime and never consults detail::invariant_cpu_freq (i.e., !BAIDU_INTERNAL && !__aarch64__). This can add avoidable startup I/O/latency. Consider gating initialization behind the same preprocessor conditions as the fast-path usage, or lazily initializing on first use.

[Copilot]

Local variable name baseFreq doesn’t match the surrounding naming style in this file (which predominantly uses snake_case like invariant_tsc, seen_decpoint, endp). Renaming to something like base_freq (and ideally indicating units, e.g. _hz) would improve consistency and readability.

Suggested change

	int64_t baseFreq = -1;
	#if defined(__aarch64__)
	__asm__ __volatile__("mrs %0, CNTFRQ_EL0" : "=r"(baseFreq));
	#else
	baseFreq = detail::read_invariant_cpu_frequency();
	#endif
	detail::invariant_cpu_freq = baseFreq;
	int64_t base_freq_hz = -1;
	#if defined(__aarch64__)
	__asm__ __volatile__("mrs %0, CNTFRQ_EL0" : "=r"(base_freq_hz));
	#else
	base_freq_hz = detail::read_invariant_cpu_frequency();
	#endif
	detail::invariant_cpu_freq = base_freq_hz;