Preparations for adding a performance profiler for the IR Interpreter

CMakeLists.txt

@@ -1554,6 +1554,8 @@ list(APPEND NativeAppSource
 	UI/CwCheatScreen.cpp
 	UI/InstallZipScreen.h
 	UI/InstallZipScreen.cpp
+	UI/JitCompareScreen.h
+	UI/JitCompareScreen.cpp
 	UI/MemStickScreen.h
 	UI/MemStickScreen.cpp
 	UI/ProfilerDraw.h

Common/LogManager.cpp

@@ -234,7 +234,7 @@ void LogManager::Log(LogLevel level, LogType type, const char *file, int line, c
 			file = fileshort + 1;
 	}
 
-	GetTimeFormatted(message.timestamp);
+	GetCurrentTimeFormatted(message.timestamp);
 
 	if (hleCurrentThreadName) {
 		snprintf(message.header, sizeof(message.header), "%-12.12s %c[%s]: %s:%d",

Common/Render/Text/draw_text_uwp.cpp

@@ -33,7 +33,9 @@ class TextDrawerFontContext {
 			Destroy();
 		}
 
-		if (factory == nullptr) return;
+		if (!factory) {
+			return;
+		}
 
 		HRESULT hr = factory->CreateTextFormat(
 			fname.c_str(),
@@ -45,24 +47,29 @@ class TextDrawerFontContext {
 			L"",
 			&textFmt
 		);
+		if (FAILED(hr)) {
+			ERROR_LOG(G3D, "Failed creating font %s", fname.c_str());
+		}
 	}
 	void Destroy() {
-		textFmt->Release();
+		if (textFmt) {
+			textFmt->Release();
+		}
 		textFmt = nullptr;
 	}
 
-	IDWriteFactory4* factory = nullptr;
-	IDWriteFontCollection1* fontCollection = nullptr;
-	IDWriteTextFormat* textFmt = nullptr;
+	IDWriteFactory4 *factory = nullptr;
+	IDWriteFontCollection1 *fontCollection = nullptr;
+	IDWriteTextFormat *textFmt = nullptr;
 	std::wstring fname;
 	int height;
 	DWRITE_FONT_WEIGHT weight;
 	float dpiScale;
 };
 
 struct TextDrawerContext {
-	ID2D1Bitmap1* bitmap;
-	ID2D1Bitmap1* mirror_bmp;
+	ID2D1Bitmap1 *bitmap;
+	ID2D1Bitmap1 *mirror_bmp;
 };
 
 TextDrawerUWP::TextDrawerUWP(Draw::DrawContext *draw) : TextDrawer(draw), ctx_(nullptr) {

Common/TimeUtil.cpp

@@ -4,6 +4,7 @@
 #include <cstdint>
 
 #include "Common/TimeUtil.h"
+#include "Common/Log.h"
 
 #ifdef HAVE_LIBNX
 #include <switch.h>
@@ -31,21 +32,25 @@
 
 // TODO: https://github.com/floooh/sokol/blob/9a6237fcdf213e6da48e4f9201f144bcb2dcb46f/sokol_time.h#L229-L248
 
-static const double micros = 1000000.0;
-static const double nanos = 1000000000.0;
+constexpr double micros = 1000000.0;
+constexpr double nanos = 1000000000.0;
 
 #if PPSSPP_PLATFORM(WINDOWS)
 
 static LARGE_INTEGER frequency;
 static double frequencyMult;
 static LARGE_INTEGER startTime;
 
-double time_now_d() {
+static inline void InitTime() {
 	if (frequency.QuadPart == 0) {
 		QueryPerformanceFrequency(&frequency);
 		QueryPerformanceCounter(&startTime);
 		frequencyMult = 1.0 / static_cast<double>(frequency.QuadPart);
 	}
+}
+
+double time_now_d() {
+	InitTime();
 	LARGE_INTEGER time;
 	QueryPerformanceCounter(&time);
 	double elapsed = static_cast<double>(time.QuadPart - startTime.QuadPart);
@@ -86,6 +91,22 @@ void yield() {
 	YieldProcessor();
 }
 
+TimeSpan::TimeSpan() {
+	_dbg_assert_(frequencyMult != 0.0);
+	QueryPerformanceCounter(reinterpret_cast<LARGE_INTEGER *>(&nativeStart_));
+}
+
+double TimeSpan::ElapsedSeconds() const {
+	LARGE_INTEGER time;
+	QueryPerformanceCounter(&time);
+	double elapsed = static_cast<double>(time.QuadPart - nativeStart_);
+	return elapsed * frequencyMult;
+}
+
+int64_t TimeSpan::ElapsedNanos() const {
+	return (int64_t)(ElapsedSeconds() * 1000000000.0);
+}
+
 #elif PPSSPP_PLATFORM(ANDROID) || PPSSPP_PLATFORM(LINUX) || PPSSPP_PLATFORM(MAC) || PPSSPP_PLATFORM(IOS)
 
 // The only intended use is to match the timings in VK_GOOGLE_display_timing
@@ -128,10 +149,34 @@ void yield() {
 #endif
 }
 
+TimeSpan::TimeSpan() {
+	struct timespec ts;
+	clock_gettime(CLOCK_MONOTONIC, &ts);
+	nativeStart_ = ts.tv_sec;
+	nsecs_ = ts.tv_nsec;
+}
+
+int64_t TimeSpan::ElapsedNanos() const {
+	struct timespec ts;
+	clock_gettime(CLOCK_MONOTONIC, &ts);
+	int64_t secs = ts.tv_sec - nativeStart_;
+	int64_t nsecs = ts.tv_nsec - nsecs_;
+	if (nsecs < 0) {
+		secs--;
+		nsecs += 1000000000;
+	}
+	return secs * 1000000000ULL + nsecs;
+}
+
+double TimeSpan::ElapsedSeconds() const {
+	return (double)ElapsedNanos() * (1.0 / nanos);
+}
+
 #else
 
+static time_t start;
+
 double time_now_d() {
-	static time_t start;
 	struct timeval tv;
 	gettimeofday(&tv, nullptr);
 	if (start == 0) {
@@ -141,7 +186,6 @@ double time_now_d() {
 }
 
 uint64_t time_now_raw() {
-	static time_t start;
 	struct timeval tv;
 	gettimeofday(&tv, nullptr);
 	if (start == 0) {
@@ -164,6 +208,30 @@ double time_now_unix_utc() {
 	return time_now_raw();
 }
 
+TimeSpan::TimeSpan() {
+	struct timeval tv;
+	gettimeofday(&tv, nullptr);
+	nativeStart_ = tv.tv_sec;
+	nsecs_ = tv.tv_usec;
+}
+
+int64_t TimeSpan::ElapsedNanos() const {
+	struct timespec ts;
+	clock_gettime(CLOCK_MONOTONIC, &ts);
+
+	int64_t secs = ts.tv_sec - nativeStart_;
+	int64_t usecs = ts.tv_nsec - nsecs_;
+	if (usecs < 0) {
+		secs--;
+		usecs += 1000000;
+	}
+	return secs * 1000000000 + usecs * 1000;
+}
+
+double TimeSpan::ElapsedSeconds() const {
+	return (double)ElapsedNanos() * (1.0 / 1000000000);
+}
+
 #endif
 
 void sleep_ms(int ms) {
@@ -180,7 +248,7 @@ void sleep_ms(int ms) {
 
 // Return the current time formatted as Minutes:Seconds:Milliseconds
 // in the form 00:00:000.
-void GetTimeFormatted(char formattedTime[13]) {
+void GetCurrentTimeFormatted(char formattedTime[13]) {
 	time_t sysTime;
 	time(&sysTime);
 

Common/TimeUtil.h

@@ -1,5 +1,4 @@
 #pragma once
-
 #include <cstdint>
 
 // Seconds.
@@ -22,18 +21,31 @@ void sleep_ms(int ms);
 // Yield. Signals that this thread is busy-waiting but wants to allow other hyperthreads to run.
 void yield();
 
-void GetTimeFormatted(char formattedTime[13]);
+void GetCurrentTimeFormatted(char formattedTime[13]);
 
 // Rust-style Instant for clear and easy timing.
 class Instant {
 public:
 	static Instant Now() {
 		return Instant(time_now_d());
 	}
-	double Elapsed() const {
+	double ElapsedSeconds() const {
 		return time_now_d() - instantTime_;
 	}
 private:
 	explicit Instant(double initTime) : instantTime_(initTime) {}
 	double instantTime_;
 };
+
+// Most accurate timer possible - no extra double conversions. Only for spans.
+class TimeSpan {
+public:
+	TimeSpan();
+	double ElapsedSeconds() const;
+	int64_t ElapsedNanos() const;
+private:
+	uint64_t nativeStart_;
+#ifndef _WIN32
+	int64_t nsecs_;
+#endif
+};

Core/HW/Atrac3Standalone.cpp

@@ -23,14 +23,14 @@ class Atrac3Audio : public AudioDecoder {
 			if (at3pCtx_) {
 				codecOpen_ = true;
 			} else {
-				ERROR_LOG(ME, "Failed to open atrac3+ context! (channels=%d blockAlign=%d ed=%d)", channels, blockAlign, extraDataSize);
+				ERROR_LOG(ME, "Failed to open atrac3+ context! (channels=%d blockAlign=%d ed=%d)", channels, (int)blockAlign, (int)extraDataSize);
 			}
 		} else if (audioType_ == PSP_CODEC_AT3) {
 			at3Ctx_ = atrac3_alloc(channels, &blockAlign_, extraData, (int)extraDataSize);
 			if (at3Ctx_) {
 				codecOpen_ = true;
 			} else {
-				ERROR_LOG(ME, "Failed to open atrac3 context! !channels=%d blockAlign=%d ed=%d)", channels, blockAlign, extraDataSize);
+				ERROR_LOG(ME, "Failed to open atrac3 context! !channels=%d blockAlign=%d ed=%d)", channels, (int)blockAlign, (int)extraDataSize);
 			}
 		}
 		for (int i = 0; i < 2; i++) {