8 files changed, 16 insertions, 259 deletions

diff --git a/src/common/CMakeLists.txt b/src/common/CMakeLists.txt
index c839ce173..aa6eee2a3 100644
--- a/src/common/CMakeLists.txt
+++ b/src/common/CMakeLists.txt
@@ -47,7 +47,6 @@ set(HEADERS
             microprofile.h
             microprofileui.h
             platform.h
-            profiler.h
             profiler_reporting.h
             scm_rev.h
             scope_exit.h
diff --git a/src/common/assert.h b/src/common/assert.h
index 6849778b7..cd9b819a9 100644
--- a/src/common/assert.h
+++ b/src/common/assert.h
@@ -39,6 +39,7 @@ static void assert_noinline_call(const Fn& fn) {
     }); } while (0)
 
 #define UNREACHABLE() ASSERT_MSG(false, "Unreachable code!")
+#define UNREACHABLE_MSG(...) ASSERT_MSG(false, __VA_ARGS__)
 
 #ifdef _DEBUG
 #define DEBUG_ASSERT(_a_) ASSERT(_a_)
@@ -49,3 +50,4 @@ static void assert_noinline_call(const Fn& fn) {
 #endif
 
 #define UNIMPLEMENTED() DEBUG_ASSERT_MSG(false, "Unimplemented code!")
+#define UNIMPLEMENTED_MSG(_a_, ...) ASSERT_MSG(false, _a_, __VA_ARGS__)
\ No newline at end of file
diff --git a/src/common/file_util.h b/src/common/file_util.h
index b54a9fb72..3aac4fa46 100644
--- a/src/common/file_util.h
+++ b/src/common/file_util.h
@@ -192,7 +192,9 @@ public:
     size_t ReadArray(T* data, size_t length)
     {
         static_assert(std::is_standard_layout<T>(), "Given array does not consist of standard layout objects");
+#if (__GNUC__ >= 5) || defined(__clang__) || defined(_MSC_VER)
         static_assert(std::is_trivially_copyable<T>(), "Given array does not consist of trivially copyable objects");
+#endif
 
         if (!IsOpen()) {
             m_good = false;
@@ -210,7 +212,9 @@ public:
     size_t WriteArray(const T* data, size_t length)
     {
         static_assert(std::is_standard_layout<T>(), "Given array does not consist of standard layout objects");
+#if (__GNUC__ >= 5) || defined(__clang__) || defined(_MSC_VER)
         static_assert(std::is_trivially_copyable<T>(), "Given array does not consist of trivially copyable objects");
+#endif
 
         if (!IsOpen()) {
             m_good = false;
diff --git a/src/common/microprofile.h b/src/common/microprofile.h
index d3b6cb97c..ef312c6e1 100644
--- a/src/common/microprofile.h
+++ b/src/common/microprofile.h
@@ -4,6 +4,10 @@
 
 #pragma once
 
+// Uncomment this to disable microprofile. This will get you cleaner profiles when using
+// external sampling profilers like "Very Sleepy", and will improve performance somewhat.
+// #define MICROPROFILE_ENABLED 0
+
 // Customized Citra settings.
 // This file wraps the MicroProfile header so that these are consistent everywhere.
 #define MICROPROFILE_WEBSERVER 0
diff --git a/src/common/microprofileui.h b/src/common/microprofileui.h
index 97c369bd9..41abe6b75 100644
--- a/src/common/microprofileui.h
+++ b/src/common/microprofileui.h
@@ -13,4 +13,7 @@
 #define MICROPROFILE_HELP_ALT "Right-Click"
 #define MICROPROFILE_HELP_MOD "Ctrl"
 
+// This isn't included by microprofileui.h :(
+#include <cstdlib> // For std::abs
+
 #include <microprofileui.h>
diff --git a/src/common/profiler.cpp b/src/common/profiler.cpp
index 7792edd2f..49eb3f40c 100644
--- a/src/common/profiler.cpp
+++ b/src/common/profiler.cpp
@@ -7,71 +7,16 @@
 #include <vector>
 
 #include "common/assert.h"
-#include "common/profiler.h"
 #include "common/profiler_reporting.h"
 #include "common/synchronized_wrapper.h"
 
-#if defined(_MSC_VER) && _MSC_VER <= 1800 // MSVC 2013.
-    #define WIN32_LEAN_AND_MEAN
-    #include <Windows.h> // For QueryPerformanceCounter/Frequency
-#endif
-
 namespace Common {
 namespace Profiling {
 
-#if ENABLE_PROFILING
-thread_local Timer* Timer::current_timer = nullptr;
-#endif
-
-#if defined(_MSC_VER) && _MSC_VER <= 1800 // MSVC 2013
-QPCClock::time_point QPCClock::now() {
-    static LARGE_INTEGER freq;
-    // Use this dummy local static to ensure this gets initialized once.
-    static BOOL dummy = QueryPerformanceFrequency(&freq);
-
-    LARGE_INTEGER ticks;
-    QueryPerformanceCounter(&ticks);
-
-    // This is prone to overflow when multiplying, which is why I'm using micro instead of nano. The
-    // correct way to approach this would be to just return ticks as a time_point and then subtract
-    // and do this conversion when creating a duration from two time_points, however, as far as I
-    // could tell the C++ requirements for these types are incompatible with this approach.
-    return time_point(duration(ticks.QuadPart * std::micro::den / freq.QuadPart));
-}
-#endif
-
-TimingCategory::TimingCategory(const char* name, TimingCategory* parent)
-        : accumulated_duration(0) {
-
-    ProfilingManager& manager = GetProfilingManager();
-    category_id = manager.RegisterTimingCategory(this, name);
-    if (parent != nullptr)
-        manager.SetTimingCategoryParent(category_id, parent->category_id);
-}
-
 ProfilingManager::ProfilingManager()
         : last_frame_end(Clock::now()), this_frame_start(Clock::now()) {
 }
 
-unsigned int ProfilingManager::RegisterTimingCategory(TimingCategory* category, const char* name) {
-    TimingCategoryInfo info;
-    info.category = category;
-    info.name = name;
-    info.parent = TimingCategoryInfo::NO_PARENT;
-
-    unsigned int id = (unsigned int)timing_categories.size();
-    timing_categories.push_back(std::move(info));
-
-    return id;
-}
-
-void ProfilingManager::SetTimingCategoryParent(unsigned int category, unsigned int parent) {
-    ASSERT(category < timing_categories.size());
-    ASSERT(parent < timing_categories.size());
-
-    timing_categories[category].parent = parent;
-}
-
 void ProfilingManager::BeginFrame() {
     this_frame_start = Clock::now();
 }
@@ -82,11 +27,6 @@ void ProfilingManager::FinishFrame() {
     results.interframe_time = now - last_frame_end;
     results.frame_time = now - this_frame_start;
 
-    results.time_per_category.resize(timing_categories.size());
-    for (size_t i = 0; i < timing_categories.size(); ++i) {
-        results.time_per_category[i] = timing_categories[i].category->GetAccumulatedTime();
-    }
-
     last_frame_end = now;
 }
 
@@ -100,26 +40,9 @@ void TimingResultsAggregator::Clear() {
     window_size = cursor = 0;
 }
 
-void TimingResultsAggregator::SetNumberOfCategories(size_t n) {
-    size_t old_size = times_per_category.size();
-    if (n == old_size)
-        return;
-
-    times_per_category.resize(n);
-
-    for (size_t i = old_size; i < n; ++i) {
-        times_per_category[i].resize(max_window_size, Duration::zero());
-    }
-}
-
 void TimingResultsAggregator::AddFrame(const ProfilingFrameResult& frame_result) {
-    SetNumberOfCategories(frame_result.time_per_category.size());
-
     interframe_times[cursor] = frame_result.interframe_time;
     frame_times[cursor] = frame_result.frame_time;
-    for (size_t i = 0; i < frame_result.time_per_category.size(); ++i) {
-        times_per_category[i][cursor] = frame_result.time_per_category[i];
-    }
 
     ++cursor;
     if (cursor == max_window_size)
@@ -162,11 +85,6 @@ AggregatedFrameResult TimingResultsAggregator::GetAggregatedResults() const {
         result.fps = 0.0f;
     }
 
-    result.time_per_category.resize(times_per_category.size());
-    for (size_t i = 0; i < times_per_category.size(); ++i) {
-        result.time_per_category[i] = AggregateField(times_per_category[i], window_size);
-    }
-
     return result;
 }
 
diff --git a/src/common/profiler.h b/src/common/profiler.h
deleted file mode 100644
index 3e967b4bc..000000000
--- a/src/common/profiler.h
+++ /dev/null
@@ -1,152 +0,0 @@
-// Copyright 2015 Citra Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#pragma once
-
-#include <atomic>
-#include <chrono>
-
-#include "common/assert.h"
-#include "common/thread.h"
-
-namespace Common {
-namespace Profiling {
-
-// If this is defined to 0, it turns all Timers into no-ops.
-#ifndef ENABLE_PROFILING
-#define ENABLE_PROFILING 1
-#endif
-
-#if defined(_MSC_VER) && _MSC_VER <= 1800 // MSVC 2013
-// MSVC up to 2013 doesn't use QueryPerformanceCounter for high_resolution_clock, so it has bad
-// precision. We manually implement a clock based on QPC to get good results.
-
-struct QPCClock {
-    using duration = std::chrono::microseconds;
-    using time_point = std::chrono::time_point<QPCClock>;
-    using rep = duration::rep;
-    using period = duration::period;
-    static const bool is_steady = false;
-
-    static time_point now();
-};
-
-using Clock = QPCClock;
-#else
-using Clock = std::chrono::high_resolution_clock;
-#endif
-
-using Duration = Clock::duration;
-
-/**
- * Represents a timing category that measured time can be accounted towards. Should be declared as a
- * global variable and passed to Timers.
- */
-class TimingCategory final {
-public:
-    TimingCategory(const char* name, TimingCategory* parent = nullptr);
-
-    unsigned int GetCategoryId() const {
-        return category_id;
-    }
-
-    /// Adds some time to this category. Can safely be called from multiple threads at the same time.
-    void AddTime(Duration amount) {
-        std::atomic_fetch_add_explicit(
-                &accumulated_duration, amount.count(),
-                std::memory_order_relaxed);
-    }
-
-    /**
-     * Atomically retrieves the accumulated measured time for this category and resets the counter
-     * to zero. Can be safely called concurrently with AddTime.
-     */
-    Duration GetAccumulatedTime() {
-        return Duration(std::atomic_exchange_explicit(
-                &accumulated_duration, (Duration::rep)0,
-                std::memory_order_relaxed));
-    }
-
-private:
-    unsigned int category_id;
-    std::atomic<Duration::rep> accumulated_duration;
-};
-
-/**
- * Measures time elapsed between a call to Start and a call to Stop and attributes it to the given
- * TimingCategory. Start/Stop can be called multiple times on the same timer, but each call must be
- * appropriately paired.
- *
- * When a Timer is started, it automatically pauses a previously running timer on the same thread,
- * which is resumed when it is stopped. As such, no special action needs to be taken to avoid
- * double-accounting of time on two categories.
- */
-class Timer {
-public:
-    Timer(TimingCategory& category) : category(category) {
-    }
-
-    void Start() {
-#if ENABLE_PROFILING
-        ASSERT(!running);
-        previous_timer = current_timer;
-        current_timer = this;
-        if (previous_timer != nullptr)
-            previous_timer->StopTiming();
-
-        StartTiming();
-#endif
-    }
-
-    void Stop() {
-#if ENABLE_PROFILING
-        ASSERT(running);
-        StopTiming();
-
-        if (previous_timer != nullptr)
-            previous_timer->StartTiming();
-        current_timer = previous_timer;
-#endif
-    }
-
-private:
-#if ENABLE_PROFILING
-    void StartTiming() {
-        start = Clock::now();
-        running = true;
-    }
-
-    void StopTiming() {
-        auto duration = Clock::now() - start;
-        running = false;
-        category.AddTime(std::chrono::duration_cast<Duration>(duration));
-    }
-
-    Clock::time_point start;
-    bool running = false;
-
-    Timer* previous_timer;
-    static thread_local Timer* current_timer;
-#endif
-
-    TimingCategory& category;
-};
-
-/**
- * A Timer that automatically starts timing when created and stops at the end of the scope. Should
- * be used in the majority of cases.
- */
-class ScopeTimer : public Timer {
-public:
-    ScopeTimer(TimingCategory& category) : Timer(category) {
-        Start();
-    }
-
-    ~ScopeTimer() {
-        Stop();
-    }
-};
-
-} // namespace Profiling
-} // namespace Common
diff --git a/src/common/profiler_reporting.h b/src/common/profiler_reporting.h
index df98e05b7..fa1ac883f 100644
--- a/src/common/profiler_reporting.h
+++ b/src/common/profiler_reporting.h
@@ -4,22 +4,17 @@
 
 #pragma once
 
+#include <chrono>
 #include <cstddef>
 #include <vector>
 
-#include "common/profiler.h"
 #include "common/synchronized_wrapper.h"
 
 namespace Common {
 namespace Profiling {
 
-struct TimingCategoryInfo {
-    static const unsigned int NO_PARENT = -1;
-
-    TimingCategory* category;
-    const char* name;
-    unsigned int parent;
-};
+using Clock = std::chrono::high_resolution_clock;
+using Duration = Clock::duration;
 
 struct ProfilingFrameResult {
     /// Time since the last delivered frame
@@ -27,22 +22,12 @@ struct ProfilingFrameResult {
 
     /// Time spent processing a frame, excluding VSync
     Duration frame_time;
-
-    /// Total amount of time spent inside each category in this frame. Indexed by the category id
-    std::vector<Duration> time_per_category;
 };
 
 class ProfilingManager final {
 public:
     ProfilingManager();
 
-    unsigned int RegisterTimingCategory(TimingCategory* category, const char* name);
-    void SetTimingCategoryParent(unsigned int category, unsigned int parent);
-
-    const std::vector<TimingCategoryInfo>& GetTimingCategoriesInfo() const {
-        return timing_categories;
-    }
-
     /// This should be called after swapping screen buffers.
     void BeginFrame();
     /// This should be called before swapping screen buffers.
@@ -54,7 +39,6 @@ public:
     }
 
 private:
-    std::vector<TimingCategoryInfo> timing_categories;
     Clock::time_point last_frame_end;
     Clock::time_point this_frame_start;
 
@@ -73,9 +57,6 @@ struct AggregatedFrameResult {
     AggregatedDuration frame_time;
 
     float fps;
-
-    /// Total amount of time spent inside each category in this frame. Indexed by the category id
-    std::vector<AggregatedDuration> time_per_category;
 };
 
 class TimingResultsAggregator final {
@@ -83,7 +64,6 @@ public:
     TimingResultsAggregator(size_t window_size);
 
     void Clear();
-    void SetNumberOfCategories(size_t n);
 
     void AddFrame(const ProfilingFrameResult& frame_result);
 
@@ -95,7 +75,6 @@ public:
 
     std::vector<Duration> interframe_times;
     std::vector<Duration> frame_times;
-    std::vector<std::vector<Duration>> times_per_category;
 };
 
 ProfilingManager& GetProfilingManager();