From d4f871cb6a616f6b1a8a76049e042118571f3dd3 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Fri, 1 Jan 2021 23:28:55 +0100
Subject: X86/NativeClock: Improve performance of clock calculations on hot
 path.

---
 src/common/x64/native_clock.cpp | 69 ++++++++++++++++++++++++++++++++++++++---
 src/common/x64/native_clock.h   |  7 +++++
 2 files changed, 71 insertions(+), 5 deletions(-)

(limited to 'src')

diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index eb8a7782f..e246432d0 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -2,12 +2,17 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <array>
 #include <chrono>
+#include <limits>
 #include <mutex>
 #include <thread>
 
 #ifdef _MSC_VER
 #include <intrin.h>
+
+#pragma intrinsic(__umulh)
+#pragma intrinsic(_udiv128)
 #else
 #include <x86intrin.h>
 #endif
@@ -15,6 +20,55 @@
 #include "common/uint128.h"
 #include "common/x64/native_clock.h"
 
+namespace {
+
+[[nodiscard]] u64 GetFixedPoint64Factor(u64 numerator, u64 divisor) {
+#ifdef __SIZEOF_INT128__
+    const auto base = static_cast<unsigned __int128>(numerator) << 64ULL;
+    return static_cast<u64>(base / divisor);
+#elif defined(_M_X64) || defined(_M_ARM64)
+    std::array<u64, 2> r = {0, numerator};
+    u64 remainder;
+#if _MSC_VER < 1923
+    return udiv128(r[1], r[0], divisor, &remainder);
+#else
+    return _udiv128(r[1], r[0], divisor, &remainder);
+#endif
+#else
+    // This one is bit more inaccurate.
+    return MultiplyAndDivide64(std::numeric_limits<u64>::max(), numerator, divisor);
+#endif
+}
+
+[[nodiscard]] u64 MultiplyHigh(u64 a, u64 b) {
+#ifdef __SIZEOF_INT128__
+    return (static_cast<unsigned __int128>(a) * static_cast<unsigned __int128>(b)) >> 64;
+#elif defined(_M_X64) || defined(_M_ARM64)
+    return __umulh(a, b); // MSVC
+#else
+    // Generic fallback
+    const u64 a_lo = u32(a);
+    const u64 a_hi = a >> 32;
+    const u64 b_lo = u32(b);
+    const u64 b_hi = b >> 32;
+
+    const u64 a_x_b_hi = a_hi * b_hi;
+    const u64 a_x_b_mid = a_hi * b_lo;
+    const u64 b_x_a_mid = b_hi * a_lo;
+    const u64 a_x_b_lo = a_lo * b_lo;
+
+    const u64 carry_bit = (static_cast<u64>(static_cast<u32>(a_x_b_mid)) +
+                           static_cast<u64>(static_cast<u32>(b_x_a_mid)) + (a_x_b_lo >> 32)) >>
+                          32;
+
+    const u64 multhi = a_x_b_hi + (a_x_b_mid >> 32) + (b_x_a_mid >> 32) + carry_bit;
+
+    return multhi;
+#endif
+}
+
+} // namespace
+
 namespace Common {
 
 u64 EstimateRDTSCFrequency() {
@@ -50,6 +104,11 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen
     _mm_mfence();
     last_measure = __rdtsc();
     accumulated_ticks = 0U;
+    ns_rtsc_factor = GetFixedPoint64Factor(1000000000, rtsc_frequency);
+    us_rtsc_factor = GetFixedPoint64Factor(1000000, rtsc_frequency);
+    ms_rtsc_factor = GetFixedPoint64Factor(1000, rtsc_frequency);
+    clock_rtsc_factor = GetFixedPoint64Factor(emulated_clock_frequency, rtsc_frequency);
+    cpu_rtsc_factor = GetFixedPoint64Factor(emulated_cpu_frequency, rtsc_frequency);
 }
 
 u64 NativeClock::GetRTSC() {
@@ -75,27 +134,27 @@ void NativeClock::Pause(bool is_paused) {
 
 std::chrono::nanoseconds NativeClock::GetTimeNS() {
     const u64 rtsc_value = GetRTSC();
-    return std::chrono::nanoseconds{MultiplyAndDivide64(rtsc_value, 1000000000, rtsc_frequency)};
+    return std::chrono::nanoseconds{MultiplyHigh(rtsc_value, ns_rtsc_factor)};
 }
 
 std::chrono::microseconds NativeClock::GetTimeUS() {
     const u64 rtsc_value = GetRTSC();
-    return std::chrono::microseconds{MultiplyAndDivide64(rtsc_value, 1000000, rtsc_frequency)};
+    return std::chrono::microseconds{MultiplyHigh(rtsc_value, us_rtsc_factor)};
 }
 
 std::chrono::milliseconds NativeClock::GetTimeMS() {
     const u64 rtsc_value = GetRTSC();
-    return std::chrono::milliseconds{MultiplyAndDivide64(rtsc_value, 1000, rtsc_frequency)};
+    return std::chrono::milliseconds{MultiplyHigh(rtsc_value, ms_rtsc_factor)};
 }
 
 u64 NativeClock::GetClockCycles() {
     const u64 rtsc_value = GetRTSC();
-    return MultiplyAndDivide64(rtsc_value, emulated_clock_frequency, rtsc_frequency);
+    return MultiplyHigh(rtsc_value, clock_rtsc_factor);
 }
 
 u64 NativeClock::GetCPUCycles() {
     const u64 rtsc_value = GetRTSC();
-    return MultiplyAndDivide64(rtsc_value, emulated_cpu_frequency, rtsc_frequency);
+    return MultiplyHigh(rtsc_value, cpu_rtsc_factor);
 }
 
 } // namespace X64
diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h
index 6d1e32ac8..a7b1ee9e0 100644
--- a/src/common/x64/native_clock.h
+++ b/src/common/x64/native_clock.h
@@ -41,6 +41,13 @@ private:
     u64 last_measure{};
     u64 accumulated_ticks{};
     u64 rtsc_frequency;
+
+    // factors
+    u64 ns_rtsc_factor{};
+    u64 us_rtsc_factor{};
+    u64 ms_rtsc_factor{};
+    u64 clock_rtsc_factor{};
+    u64 cpu_rtsc_factor{};
 };
 } // namespace X64
 
-- 
cgit v1.2.3