From fdd4d019ef6f37421358ae70f3d5e036857dd758 Mon Sep 17 00:00:00 2001 From: merry Date: Sat, 2 Apr 2022 21:54:39 +0100 Subject: native_clock: Use lfence with rdtsc --- src/common/x64/native_clock.cpp | 47 +++++++++++++++++++++++++++++------------ 1 file changed, 33 insertions(+), 14 deletions(-) (limited to 'src/common') diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp index 7a3f21dcf..63364f839 100644 --- a/src/common/x64/native_clock.cpp +++ b/src/common/x64/native_clock.cpp @@ -10,25 +10,47 @@ #include "common/uint128.h" #include "common/x64/native_clock.h" +#ifdef _MSC_VER +#include +#endif + namespace Common { +inline u64 FencedRDTSC() { +#ifdef _MSC_VER + _mm_lfence(); + _ReadWriteBarrier(); + const u64 result = __rdtsc(); + _mm_lfence(); + _ReadWriteBarrier(); + return result; +#else + u64 result; + asm volatile("lfence\n\t" + "rdtsc\n\t" + "shl $32, %%rdx\n\t" + "or %%rdx, %0\n\t" + "lfence" + : "=a"(result) + : + : "rdx", "memory", "cc"); + return result; +#endif +} + u64 EstimateRDTSCFrequency() { // Discard the first result measuring the rdtsc. - _mm_mfence(); - __rdtsc(); + FencedRDTSC(); std::this_thread::sleep_for(std::chrono::milliseconds{1}); - _mm_mfence(); - __rdtsc(); + FencedRDTSC(); // Get the current time. const auto start_time = std::chrono::steady_clock::now(); - _mm_mfence(); - const u64 tsc_start = __rdtsc(); + const u64 tsc_start = FencedRDTSC(); // Wait for 200 milliseconds. std::this_thread::sleep_for(std::chrono::milliseconds{200}); const auto end_time = std::chrono::steady_clock::now(); - _mm_mfence(); - const u64 tsc_end = __rdtsc(); + const u64 tsc_end = FencedRDTSC(); // Calculate differences. const u64 timer_diff = static_cast( std::chrono::duration_cast(end_time - start_time).count()); @@ -42,8 +64,7 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen u64 rtsc_frequency_) : WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, true), rtsc_frequency{ rtsc_frequency_} { - _mm_mfence(); - time_point.inner.last_measure = __rdtsc(); + time_point.inner.last_measure = FencedRDTSC(); time_point.inner.accumulated_ticks = 0U; ns_rtsc_factor = GetFixedPoint64Factor(NS_RATIO, rtsc_frequency); us_rtsc_factor = GetFixedPoint64Factor(US_RATIO, rtsc_frequency); @@ -58,8 +79,7 @@ u64 NativeClock::GetRTSC() { current_time_point.pack = Common::AtomicLoad128(time_point.pack.data()); do { - _mm_mfence(); - const u64 current_measure = __rdtsc(); + const u64 current_measure = FencedRDTSC(); u64 diff = current_measure - current_time_point.inner.last_measure; diff = diff & ~static_cast(static_cast(diff) >> 63); // max(diff, 0) new_time_point.inner.last_measure = current_measure > current_time_point.inner.last_measure @@ -80,8 +100,7 @@ void NativeClock::Pause(bool is_paused) { current_time_point.pack = Common::AtomicLoad128(time_point.pack.data()); do { new_time_point.pack = current_time_point.pack; - _mm_mfence(); - new_time_point.inner.last_measure = __rdtsc(); + new_time_point.inner.last_measure = FencedRDTSC(); } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack, current_time_point.pack, current_time_point.pack)); } -- cgit v1.2.3 From 4052bfb4ada22bc922461dab09bef886e3f114ae Mon Sep 17 00:00:00 2001 From: Merry Date: Sun, 3 Apr 2022 00:13:27 +0100 Subject: native_clock: Internal linkage for FencedRDTSC __forceinline required on MSVC for function to be inlined --- src/common/x64/native_clock.cpp | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) (limited to 'src/common') diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp index 63364f839..7fd9d22f8 100644 --- a/src/common/x64/native_clock.cpp +++ b/src/common/x64/native_clock.cpp @@ -16,15 +16,17 @@ namespace Common { -inline u64 FencedRDTSC() { #ifdef _MSC_VER +__forceinline static u64 FencedRDTSC() { _mm_lfence(); _ReadWriteBarrier(); const u64 result = __rdtsc(); _mm_lfence(); _ReadWriteBarrier(); return result; +} #else +static u64 FencedRDTSC() { u64 result; asm volatile("lfence\n\t" "rdtsc\n\t" @@ -35,8 +37,8 @@ inline u64 FencedRDTSC() { : : "rdx", "memory", "cc"); return result; -#endif } +#endif u64 EstimateRDTSCFrequency() { // Discard the first result measuring the rdtsc. -- cgit v1.2.3