Fix up engine profiling

dedmen · dedmen · commit 0295ef97a312 · 2019-01-23T21:43:25.000+01:00
diff --git a/src/AdapterBrofiler.cpp b/src/AdapterBrofiler.cpp
@@ -157,6 +157,8 @@ uint32_t getRandColor() {
 
 AdapterBrofiler::AdapterBrofiler() {
     type = AdapterType::Brofiler;
+
+    static Brofiler::ThreadScope mainThreadScope("Frame");
 }
 
 
diff --git a/src/AdapterTracy.cpp b/src/AdapterTracy.cpp
@@ -16,6 +16,7 @@ class ScopeInfoTracy final: public ScopeInfo {
 class ScopeTempStorageTracy final : public ScopeTempStorage {
 public:
 	 std::unique_ptr<tracy::ScopedZone> zone;
+     std::thread::id origin;
 };
 
 AdapterTracy::AdapterTracy() {
@@ -50,15 +51,21 @@ std::shared_ptr<ScopeInfo> AdapterTracy::createScope(intercept::types::r_string
 std::shared_ptr<ScopeTempStorage> AdapterTracy::enterScope(std::shared_ptr<ScopeInfo> scope) {
     auto info = std::dynamic_pointer_cast<ScopeInfoTracy>(scope);
     if (!info) return nullptr; //#TODO debugbreak? log error?
+    ensureReady();
 
     auto ret = std::make_shared<ScopeTempStorageTracy>();
+    //ret->origin = std::this_thread::get_id();
     ret->zone = std::make_unique<tracy::ScopedZone>(&info->info, true);
     return ret;
 }
 void AdapterTracy::leaveScope(std::shared_ptr<ScopeTempStorage> tempStorage) {
     auto tmpStorage = std::dynamic_pointer_cast<ScopeTempStorageTracy>(tempStorage);
     if (!tmpStorage) return; //#TODO debugbreak? log error?
 
+
+    //if (tmpStorage->origin != std::this_thread::get_id())
+    //    __debugbreak();
+
     tmpStorage->zone.reset(); //zone destructor ends zone
 }
 
@@ -77,3 +84,17 @@ std::shared_ptr<ScopeInfo> AdapterTracy::createScopeStatic(const char* name, con
     info->info = tracy::SourceLocationData{nullptr, name, filename,fileline, 0};
     return info;
 }
+
+bool AdapterTracy::isConnected() {
+    return tracy::s_profiler.IsConnected();
+}
+
+void AdapterTracy::ensureReady() {
+    if (tracy::s_token.ptr) return;
+
+    tracy::rpmalloc_thread_initialize();
+    tracy::s_token_detail = tracy::moodycamel::ProducerToken(tracy::s_queue);
+    tracy::s_token = tracy::ProducerWrapper{tracy::s_queue.get_explicit_producer(tracy::s_token_detail) };
+
+
+}
diff --git a/src/AdapterTracy.hpp b/src/AdapterTracy.hpp
@@ -18,10 +18,10 @@ class AdapterTracy final : public ProfilerAdapter
 	void setCounter(intercept::types::r_string name, float val) override;
 
 	std::shared_ptr<ScopeInfo> createScopeStatic(const char* name, const char* filename, uint32_t fileline);
-
+    bool isConnected();
 
 private:
-
+    void ensureReady();
 	using scopeCacheKey = std::tuple<intercept::types::r_string, intercept::types::r_string,uint32_t>;
 
 	struct ScopeCacheFastEqual {
diff --git a/src/EngineProfiling.cpp b/src/EngineProfiling.cpp
@@ -1,10 +1,26 @@
 #include "EngineProfiling.h"
 #include "ProfilerAdapter.hpp"
 #include "AdapterTracy.hpp"
+#include <client.hpp>
+#include <shared_mutex>
 
 extern std::shared_ptr<ProfilerAdapter> GProfilerAdapter;
-std::unordered_map<PCounter*, std::shared_ptr<ScopeTempStorage>> openScopes;
+
+
+struct CounterHasher {
+public:
+    size_t operator()(const std::pair<PCounter*, int>& key) const {
+        //intercept pairhash
+        size_t _hash = std::hash<PCounter*>()(key.first);
+        _hash ^= std::hash<uint64_t>()(key.second) + 0x9e3779b9 + (_hash << 6) + (_hash >> 2);
+        return _hash;
+    }
+};
+
+thread_local std::unique_ptr<std::unordered_map<std::pair<PCounter*, int>, std::shared_ptr<ScopeTempStorage>, CounterHasher>> openScopes;
+thread_local bool openScopesInit;
 std::unordered_map<PCounter*, std::shared_ptr<ScopeInfo>> scopeCache;
+std::shared_mutex scopeCacheMtx;
 
 extern "C" {
     uintptr_t profEndJmpback;
@@ -13,42 +29,65 @@ extern "C" {
 
     void shouldTime();
     void doEnd();
+    void scopeCompleted();
 }
 
 bool PCounter::shouldTime() {
+    if (slot < 0) return false;
+    //if (!boss) return false;
+    //if (!boss->da) return false;
+    //if (!boss->stuffzi) return false;
+    //if (!boss->stuffzi[slot].enabled) return false;
 
-    if (GProfilerAdapter->getType() == AdapterType::Tracy) {
-        auto tracyProf = std::dynamic_pointer_cast<AdapterTracy>(GProfilerAdapter);
-
-        auto found = scopeCache.find(this);
-        if (found == scopeCache.end()) {
-            auto res = scopeCache.insert({this, tracyProf->createScopeStatic(name, cat, 0)});
-            found = res.first;
-        }
-
-        //make sure the add is not inside the scope
-        openScopes[this] = nullptr;
-        auto ins = openScopes.find(this);
-
-        ins->second = tracyProf->enterScope(found->second);   
+    if (GProfilerAdapter->getType() != AdapterType::Tracy) return false;
+    
+    auto tracyProf = std::dynamic_pointer_cast<AdapterTracy>(GProfilerAdapter);
+    if (!tracyProf->isConnected()) return false;
+
+
+    std::shared_lock lock(scopeCacheMtx);
+    auto found = scopeCache.find(this);
+    if (found == scopeCache.end()) {
+        lock.unlock();
+        std::unique_lock lockInternal(scopeCacheMtx);
+        auto res = scopeCache.insert({ this, tracyProf->createScopeStatic(name, cat, 0) });
+        lockInternal.unlock();//#TODO this is unsafe
+        lock.lock();
+        found = res.first;
     }
 
+    if (!openScopes)
+        openScopes = std::make_unique<std::unordered_map<std::pair<PCounter*, int>, std::shared_ptr<ScopeTempStorage>, CounterHasher>>();
+
+    //make sure the add is not inside the scope
+    auto p = std::make_pair(this, slot);
+    auto ins = openScopes->insert_or_assign(p,nullptr);
+    auto tmp = tracyProf->enterScope(found->second);
+    //if (tmp)
+        ins.first->second = tmp;
+    //else
+    //    openScopes.erase(p);
+    
 
-
-
-    if (slot < 0) return false;
-    if (!boss) return false;
-    if (!boss->da) return false;
-    if (!boss->stuffzi) return false;
-    if (!boss->stuffzi[slot].enabled) return false;
     return true;
 }
 
 void ScopeProf::doEnd() {
-    auto found = openScopes.find(counter);
-    if (found == openScopes.end()) return;
+    if (!openScopes || openScopes->empty() || !counter) return;
+    auto found = openScopes->find({ counter, counter->slot });
+    if (found == openScopes->end()) return;
+    GProfilerAdapter->leaveScope(found->second);
+    openScopes->erase(found);
+}
+
+void ArmaProf::scopeCompleted(int64_t start, int64_t end, intercept::types::r_string* stuff, PCounter* counter) {
+    
+    if (!openScopes || openScopes->empty() || !counter) return;
+    auto found = openScopes->find({ counter, counter->slot });
+    if (found == openScopes->end()) return;
     GProfilerAdapter->leaveScope(found->second);
-    openScopes.erase(found);
+    openScopes->erase(found);
+
 }
 
 
@@ -57,28 +96,48 @@ HookManager::Pattern pat_doEnd{
     "\x40\x53\x48\x83\xEC\x30\x80\x79\x11\x00\x48\x8B\xD9\x75\x09\x80\x3D\x00\x00\x00\x00\x00\x74\x38\x80\x3D\x00\x00\x00\x00\x00\x74\x0B\x0F\x31\x48\xC1\xE2\x20\x48\x0B\xC2\xEB\x05\xE8\x00\x00\x00\x00\x48\x8B\x13\x4C\x8B\xC0\x48\x8B\x43\x08\x4C\x8D\x4B\x18\x48\x8D\x0D\x00\x00\x00\x00\x48\x89\x44\x24\x00\xE8\x00\x00\x00\x00\x48\x8B\x53\x18\x48\x85\xD2\x74\x1A\xF0\xFF\x0A\x75\x0D\x48\x8B\x0D\x00\x00\x00\x00\x48\x8B\x01\xFF\x50\x18\x48\xC7\x43\x00\x00\x00\x00\x00\x48\x83\xC4\x30\x5B\xC3"
 };
 
+HookManager::Pattern pat_scopeCompleted{
+    "xxxx?xxxx?xxxxxxxxxxxx????xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx?xxxxxxxx????xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx????xxx????xxxxxx????xxxx?xx????xxx",
+    "\x48\x89\x5C\x24\x00\x48\x89\x6C\x24\x00\x57\x41\x54\x41\x57\x48\x83\xEC\x20\x48\x8B\x81\x00\x00\x00\x00\x49\x8B\xF8\x4D\x8B\xE1\x48\x3B\xD0\x48\x8B\xD9\x48\x0F\x4C\xD0\x48\xC1\xF8\x04\x48\xC1\xFF\x04\x48\xC1\xFA\x04\x44\x8B\xFA\x2B\xFA\x44\x2B\xF8\x48\x8B\x44\x24\x00\x48\x63\x68\x18\x85\xED\x0F\x88\x00\x00\x00\x00\x8B\x41\x68\x66\x0F\x6E\xC7\x8B\xC8\xD1\xF9\x66\x0F\x6E\xD0\x8B\x43\x6C\x0F\x5B\xC0\x2B\xC1\x66\x0F\x6E\xC8\x0F\x5B\xD2\x42\x8D\x04\x3F\xF3\x0F\x59\xD0\x66\x0F\x6E\xC0\x0F\x5B\xC9\x0F\x5B\xC0\xF3\x0F\x59\xC8\x0F\x2F\xD1\x73\x48\x80\x3B\x00\x0F\x84\x00\x00\x00\x00\x4C\x69\xC5\x00\x00\x00\x00\x4C\x03\x43\x20\x0F\x84\x00\x00\x00\x00\x41\x0F\x0D\x48\x00\x41\xB9\x00\x00\x00\x00\x0F\x1F\x00"
+};
 
 HookManager::Pattern pat_shouldTime{
     "xxxxxxxxxxxxxxxxxxxxxxxxxx????xxxxxxxxxxxxxxxxxxxx????xxxxxxx",
     "\x48\x63\x41\x18\x85\xC0\x78\x32\x4C\x8B\x01\x33\xD2\x4D\x85\xC0\x74\x12\x41\x38\x10\x74\x0D\x48\x69\xC8\x00\x00\x00\x00\x49\x03\x48\x20\xEB\x03\x48\x8B\xCA\x48\x85\xC9\x74\x0A\x38\x51\x4A\x74\x05\xBA\x00\x00\x00\x00\x0F\xB6\xC2\xC3\x32\xC0\xC3"
 };
 
 
-EngineProfiling::EngineProfiling()
-{
+EngineProfiling::EngineProfiling() {
+    //order is important
+    //hooks.placeHook(hookTypes::doEnd, pat_doEnd, reinterpret_cast<uintptr_t>(doEnd), profEndJmpback, 1, true);
+    hooks.placeHook(hookTypes::scopeCompleted, pat_scopeCompleted, reinterpret_cast<uintptr_t>(scopeCompleted), profEndJmpback, 0);
+    hooks.placeHook(hookTypes::shouldTime, pat_shouldTime, reinterpret_cast<uintptr_t>(shouldTime), shouldTimeJmpback, 0);
 
-     hooks.placeHook(hookTypes::shouldTime, pat_shouldTime, reinterpret_cast<uintptr_t>(shouldTime), shouldTimeJmpback, 0);
-     hooks.placeHook(hookTypes::doEnd, pat_doEnd, reinterpret_cast<uintptr_t>(doEnd), profEndJmpback, 0);
+    auto found = hooks.findPattern(pat_doEnd, 0xD);
 
 
-}
 
+    auto stuffByte = found + 0x2 + 2;
+    uint32_t offs = *((uint32_t*)stuffByte);
+    uint64_t addr = stuffByte + 4+1 + offs;
+    uint64_t base = addr - 0x121;
 
-EngineProfiling::~EngineProfiling()
-{
+    armaP = reinterpret_cast<ArmaProf*>(base);
+    armaP->blip.clear();
+    armaP->forceCapture = true;
+    armaP->capture = true;
+
+    //disable captureSlowFrame because it can set forceCapture to false
+    static auto stuff = intercept::client::host::register_sqf_command("diag_captureSlowFrame", "", [](uintptr_t, game_value_parameter) -> game_value
+        {
+            return {};
+        }, game_data_type::NOTHING, game_data_type::ARRAY);
 }
 
 
+EngineProfiling::~EngineProfiling() {}
+
+
 extern "C" {
 
     
diff --git a/src/EngineProfiling.h b/src/EngineProfiling.h
@@ -40,10 +40,10 @@ class ArmaProf {
 	friend class PCounter;
 public:
 
-	void scopeCompleted(int64_t start, int64_t end, intercept::types::r_string* stuff, void* stuff2);
+    __declspec(noinline) void scopeCompleted(int64_t start, int64_t end, intercept::types::r_string* stuff, PCounter* stuff2);
 
 
-private:
+public:
 	//This is engine stuff.
 	bool da,db,dc;
 	int dd,df;
@@ -78,7 +78,11 @@ class ArmaProf {
 	int blios;
 	intercept::types::r_string blip;
 	float blop;
-	bool asd,adasd;
+
+    float dummy, dummy2, dummy3;//no idea what dis is.. Stuff above is probably wrong somewhere
+
+    bool forceCapture;
+    bool capture;
 	int stuffiz;
 
 	int64_t framestart;
@@ -99,6 +103,6 @@ class EngineProfiling {
 	EngineProfiling();
 	~EngineProfiling();
 
-
+    ArmaProf* armaP;
 	HookManager hooks;
 };
diff --git a/src/HookManager.cpp b/src/HookManager.cpp
@@ -12,7 +12,7 @@ HookManager::HookManager() {
     engineSize = static_cast<uintptr_t>(modInfo.SizeOfImage);
 }
 
-bool HookManager::placeHook(hookTypes type, const Pattern& pat, uintptr_t jmpTo, uintptr_t & jmpBackRef, uint8_t jmpBackOffset) {
+bool HookManager::placeHook(hookTypes type, const Pattern& pat, uintptr_t jmpTo, uintptr_t & jmpBackRef, uint8_t jmpBackOffset, bool taintRax) {
 
     auto found = findPattern(pat);
     if (found == 0) {
@@ -21,7 +21,7 @@ bool HookManager::placeHook(hookTypes type, const Pattern& pat, uintptr_t jmpTo,
 //#endif
         return false;
     }
-    jmpBackRef = placeHookTotalOffs(found, jmpTo) + jmpBackOffset;
+    jmpBackRef = placeHookTotalOffs(found, jmpTo, taintRax) + jmpBackOffset;
     return true;
 }
 
@@ -42,7 +42,7 @@ uintptr_t HookManager::placeHook(uintptr_t offset, uintptr_t jmpTo, uint8_t jmpB
     return placeHookTotalOffs(totalOffset, jmpTo) + jmpBackOffset;
 }
 
-uintptr_t HookManager::placeHookTotalOffs(uintptr_t totalOffset, uintptr_t jmpTo) {
+uintptr_t HookManager::placeHookTotalOffs(uintptr_t totalOffset, uintptr_t jmpTo, bool taintRax) {
     DWORD dwVirtualProtectBackup;
 
 
@@ -53,19 +53,50 @@ uintptr_t HookManager::placeHookTotalOffs(uintptr_t totalOffset, uintptr_t jmpTo
     64bit
     FF 25 64bit relative
     */
-#ifdef X64
+#ifdef _WIN64
     //auto distance = std::max(totalOffset, jmpTo) - std::min(totalOffset, jmpTo);
     // if distance < 2GB (2147483648) we could use the 32bit relative jmp
-    VirtualProtect(reinterpret_cast<LPVOID>(totalOffset), 14u, 0x40u, &dwVirtualProtectBackup);
-    auto jmpInstr = reinterpret_cast<unsigned char*>(totalOffset);
-    auto addrOffs = reinterpret_cast<uint32_t*>(totalOffset + 1);
-    *jmpInstr = 0x68; //push DWORD
-    *addrOffs = static_cast<uint32_t>(jmpTo) /*- totalOffset - 6*/;//offset
-    *reinterpret_cast<uint32_t*>(totalOffset + 5) = 0x042444C7; //MOV [RSP+4],
-    *reinterpret_cast<uint32_t*>(totalOffset + 9) = static_cast<uint64_t>(jmpTo) >> 32;//DWORD
-    *reinterpret_cast<unsigned char*>(totalOffset + 13) = 0xc3;//ret
-    VirtualProtect(reinterpret_cast<LPVOID>(totalOffset), 14u, dwVirtualProtectBackup, &dwVirtualProtectBackup);
-    return totalOffset + 14;
+
+
+    if (taintRax) {
+        //This is shorter, but messes up RAX
+
+        /*
+        push rax; //to restore it inside out target
+        moveabs rax, address
+        push rax
+        ret
+         */
+
+        VirtualProtect(reinterpret_cast<LPVOID>(totalOffset), 12u, 0x40u, &dwVirtualProtectBackup);
+        auto memberRax = reinterpret_cast<unsigned char*>(totalOffset);
+        auto jmpInstr1 = reinterpret_cast<unsigned char*>(totalOffset+1);
+        auto jmpInstr2 = reinterpret_cast<unsigned char*>(totalOffset+2);
+        auto addrOffs = reinterpret_cast<uint64_t*>(totalOffset + 3);
+        *memberRax = 0x50; //push rax
+        *jmpInstr1 = 0x48; //moveabs
+        *jmpInstr2 = 0xb8; //into rax
+        *addrOffs = static_cast<uint64_t>(jmpTo) /*- totalOffset - 6*/;//offset
+
+        auto pushInstr = reinterpret_cast<unsigned char*>(totalOffset + 11);
+        *pushInstr = 0x50; //push rax
+        *reinterpret_cast<unsigned char*>(totalOffset + 12) = 0xc3;//ret
+        VirtualProtect(reinterpret_cast<LPVOID>(totalOffset), 12u, dwVirtualProtectBackup, &dwVirtualProtectBackup);
+        return totalOffset + 12;
+    } else {
+        VirtualProtect(reinterpret_cast<LPVOID>(totalOffset), 14u, 0x40u, &dwVirtualProtectBackup);
+        auto jmpInstr = reinterpret_cast<unsigned char*>(totalOffset);
+        auto addrOffs = reinterpret_cast<uint32_t*>(totalOffset + 1);
+        *jmpInstr = 0x68; //push DWORD
+        *addrOffs = static_cast<uint32_t>(jmpTo) /*- totalOffset - 6*/;//offset
+        *reinterpret_cast<uint32_t*>(totalOffset + 5) = 0x042444C7; //MOV [RSP+4],
+        *reinterpret_cast<uint32_t*>(totalOffset + 9) = static_cast<uint64_t>(jmpTo) >> 32;//DWORD
+        *reinterpret_cast<unsigned char*>(totalOffset + 13) = 0xc3;//ret
+        VirtualProtect(reinterpret_cast<LPVOID>(totalOffset), 14u, dwVirtualProtectBackup, &dwVirtualProtectBackup);
+        return totalOffset + 14;
+    }
+
+
 #else
     VirtualProtect(reinterpret_cast<LPVOID>(totalOffset), 5u, 0x40u, &dwVirtualProtectBackup);
     auto jmpInstr = reinterpret_cast<unsigned char *>(totalOffset);
diff --git a/src/HookManager.hpp b/src/HookManager.hpp
@@ -10,6 +10,7 @@
 enum class hookTypes {
     shouldTime,   //FrameEnd/FrameStart
     doEnd,
+    scopeCompleted,
     End
 };
 
@@ -30,10 +31,10 @@ class HookManager {//Implementation in dllmain
 
 
     HookManager();
-    bool placeHook(hookTypes, const Pattern& pat, uintptr_t jmpTo, uintptr_t& jmpBackRef, uint8_t jmpBackOffset = 0);
+    bool placeHook(hookTypes, const Pattern& pat, uintptr_t jmpTo, uintptr_t& jmpBackRef, uint8_t jmpBackOffset = 0, bool taintRax = false);
     bool placeHook(hookTypes, const Pattern& pat, uintptr_t jmpTo);
     uintptr_t placeHook(uintptr_t offset, uintptr_t jmpTo, uint8_t jmpBackOffset = 0);
-    uintptr_t placeHookTotalOffs(uintptr_t offset, uintptr_t jmpTo);
+    uintptr_t placeHookTotalOffs(uintptr_t offset, uintptr_t jmpTo, bool taintRax = false);
     bool MatchPattern(uintptr_t addr, const char* pattern, const char* mask);
     uintptr_t findPattern(const char* pattern, const char* mask, uintptr_t offset = 0);
     uintptr_t findPattern(const Pattern& pat, uintptr_t offset = 0);
diff --git a/src/hooks.asm b/src/hooks.asm
diff --git a/src/scriptProfiler.cpp b/src/scriptProfiler.cpp

Original file line number	Diff line number	Diff line change
`@@ -157,6 +157,8 @@ uint32_t getRandColor() {`
`157`	`157`
`158`	`158`	`AdapterBrofiler::AdapterBrofiler() {`
`159`	`159`	`type = AdapterType::Brofiler;`
	`160`	`+`
	`161`	`+ static Brofiler::ThreadScope mainThreadScope("Frame");`
`160`	`162`	`}`
`161`	`163`
`162`	`164`