From 9578c23aa189a02716e7e154feece4ea61787570 Mon Sep 17 00:00:00 2001
From: Marisa Kirisame <marisa@sayachan.org>
Date: Tue, 8 Mar 2022 17:54:19 +0100
Subject: [PATCH] Use RDTSC on Linux if possible for stats.

---
 src/common/engine/stats.h                  | 52 ++++++++++++++++++++++
 src/common/platform/posix/sdl/i_system.cpp | 46 +++++++++++++++++++
 2 files changed, 98 insertions(+)

diff --git a/src/common/engine/stats.h b/src/common/engine/stats.h
index 64b6e9d75c..3530d18834 100644
--- a/src/common/engine/stats.h
+++ b/src/common/engine/stats.h
@@ -54,6 +54,42 @@ public:
 
 #include <time.h>
 
+// [MK] try to use RDTSC on linux if possible
+// avoids excess latency of clock_gettime() on some platforms
+#ifdef __linux__
+extern bool PerfAvailable;
+extern double PerfToSec, PerfToMillisec;
+
+inline uint64_t rdtsc()
+{
+#ifdef __amd64__
+	uint64_t tsc;
+	asm volatile("rdtsc; shlq $32, %%rdx; orq %%rdx, %%rax":"=a"(tsc)::"%rdx");
+	return tsc;
+#elif defined __ppc__
+	unsigned int lower, upper, temp;
+	do
+	{
+		asm volatile ("mftbu %0 \n mftb %1 \n mftbu %2 \n"
+			: "=r"(upper), "=r"(lower), "=r"(temp));
+	}
+	while (upper != temp);
+	return (static_cast<unsigned long long>(upper) << 32) | lower;
+#elif defined __aarch64__
+	// TODO: Implement and test on ARM64
+	return 0;
+#else // i386
+	if (CPU.bRDTSC)
+	{
+		uint64_t tsc;
+		asm volatile ("\trdtsc\n" : "=A" (tsc));
+		return tsc;
+	}
+	return 0;
+#endif // __amd64__
+}
+#endif
+
 class cycle_t
 {
 public:
@@ -64,6 +100,14 @@ public:
 
 	void Clock()
 	{
+#ifdef __linux__
+		if ( PerfAvailable )
+		{
+			int64_t time = rdtsc();
+			Sec -= time * PerfToSec;
+			return;
+		}
+#endif
 		timespec ts;
 
 		clock_gettime(CLOCK_MONOTONIC, &ts);
@@ -72,6 +116,14 @@ public:
 
 	void Unclock()
 	{
+#ifdef __linux__
+		if ( PerfAvailable )
+		{
+			int64_t time = rdtsc();
+			Sec += time * PerfToSec;
+			return;
+		}
+#endif
 		timespec ts;
 
 		clock_gettime(CLOCK_MONOTONIC, &ts);
diff --git a/src/common/platform/posix/sdl/i_system.cpp b/src/common/platform/posix/sdl/i_system.cpp
index 0bca723049..243c64edc0 100644
--- a/src/common/platform/posix/sdl/i_system.cpp
+++ b/src/common/platform/posix/sdl/i_system.cpp
@@ -47,6 +47,13 @@
 #include <unistd.h>
 #include <sys/ioctl.h>
 
+#ifdef __linux__
+#include <asm/unistd.h>
+#include <linux/perf_event.h>
+#include <sys/mman.h>
+#include "printf.h"
+#endif
+
 #include <SDL.h>
 
 #include "version.h"
@@ -131,8 +138,47 @@ void I_ShowFatalError(const char *message)
 #endif
 }
 
+bool PerfAvailable;
+
 void CalculateCPUSpeed()
 {
+	PerfAvailable = false;
+	PerfToMillisec = PerfToSec = 0.;
+#ifdef __linux__
+	// [MK] read from perf values if we can
+	struct perf_event_attr pe =
+	{
+		.type = PERF_TYPE_HARDWARE,
+		.size = sizeof(struct perf_event_attr),
+		.config = PERF_COUNT_HW_INSTRUCTIONS,
+		.disabled = 1,
+		.exclude_kernel = 1,
+		.exclude_hv = 1
+	};
+	int fd = syscall(__NR_perf_event_open, &pe, 0, -1, -1, 0);
+	if (fd == -1)
+	{
+		return;
+	}
+	void *addr = mmap(nullptr, 4096, PROT_READ, MAP_SHARED, fd, 0);
+	if (addr == nullptr)
+	{
+		close(fd);
+		return;
+	}
+	struct perf_event_mmap_page *pc = (struct perf_event_mmap_page *)addr;
+	if (pc->cap_user_time != 1)
+	{
+		close(fd);
+		return;
+	}
+	double mhz = (1000LU << pc->time_shift) / (double)pc->time_mult;
+	PerfAvailable = true;
+	PerfToSec = .000001/mhz;
+	PerfToMillisec = PerfToSec*1000.;
+	if (!batchrun) Printf("CPU speed: %.0f MHz\n", mhz);
+	close(fd);
+#endif
 }
 
 void CleanProgressBar()