Merge remote-tracking branch 'remotes/origin/master' into audio_interp

# Conflicts:
#	src/frontend/qt_sdl/main.cpp

CMakeLists.txt

@@ -67,6 +67,13 @@ endif()
 
 if (ENABLE_JIT)
 	add_definitions(-DJIT_ENABLED)
+
+	option(ENABLE_JIT_PROFILING "Enable JIT profiling with VTune" OFF)
+
+	if (ENABLE_JIT_PROFILING)
+		include(cmake/FindVTune.cmake)
+		add_definitions(-DJIT_PROFILING_ENABLED)
+	endif()
 endif()
 
 if (CMAKE_BUILD_TYPE STREQUAL Release)

cmake/FindVTune.cmake

@@ -0,0 +1,5 @@
+
+find_path(VTUNE_PATH "")
+
+include_directories("${VTUNE_PATH}/include")
+link_directories("${VTUNE_PATH}/lib64")

src/ARMJIT.cpp

@@ -43,9 +43,6 @@
 #include "Wifi.h"
 #include "NDSCart.h"
 
-#if defined(__APPLE__) && defined(__aarch64__)
-    #include <pthread.h>
-#endif
 
 #include "ARMJIT_x64/ARMJIT_Offsets.h"
 static_assert(offsetof(ARM, CPSR) == ARM_CPSR_offset, "");
@@ -320,9 +317,7 @@ void Init()
 
 void DeInit()
 {
-    #if defined(__APPLE__) && defined(__aarch64__)
+    JitEnableWrite();
-        pthread_jit_write_protect_np(false);
-    #endif
     ResetBlockCache();
     ARMJIT_Memory::DeInit();
 
@@ -331,9 +326,7 @@ void DeInit()
 
 void Reset()
 {
-    #if defined(__APPLE__) && defined(__aarch64__)
+    JitEnableWrite();
-        pthread_jit_write_protect_np(false);
-    #endif
     ResetBlockCache();
 
     ARMJIT_Memory::Reset();
@@ -918,13 +911,10 @@ void CompileBlock(ARM* cpu)
         block->StartAddrLocal = localAddr;
 
         FloodFillSetFlags(instrs, i - 1, 0xF);
-        #if defined(__APPLE__) && defined(__aarch64__)
+        
-            pthread_jit_write_protect_np(false);
+        JitEnableWrite();
-        #endif
         block->EntryPoint = JITCompiler->CompileBlock(cpu, thumb, instrs, i, hasMemoryInstr);
-        #if defined(__APPLE__) && defined(__aarch64__)
+        JitEnableExecute();
-            pthread_jit_write_protect_np(true);
-        #endif
 
         JIT_DEBUGPRINT("block start %p\n", block->EntryPoint);
     }
@@ -1168,4 +1158,20 @@ void ResetBlockCache()
     JITCompiler->Reset();
 }
 
+void JitEnableWrite()
+{
+    #if defined(__APPLE__) && defined(__aarch64__)
+        if (__builtin_available(macOS 11.0, *))
+            pthread_jit_write_protect_np(false);
+    #endif
+}
+
+void JitEnableExecute()
+{
+    #if defined(__APPLE__) && defined(__aarch64__)
+        if (__builtin_available(macOS 11.0, *))
+            pthread_jit_write_protect_np(true);
+    #endif
+}
+
 }

src/ARMJIT.h

@@ -24,6 +24,10 @@
 #include "ARM.h"
 #include "ARM_InstrInfo.h"
 
+#if defined(__APPLE__) && defined(__aarch64__)
+    #include <pthread.h>
+#endif
+
 namespace ARMJIT
 {
 
@@ -48,6 +52,8 @@ void ResetBlockCache();
 JitBlockEntry LookUpBlock(u32 num, u64* entries, u32 offset, u32 addr);
 bool SetupExecutableRegion(u32 num, u32 blockAddr, u64*& entry, u32& start, u32& size);
 
+void JitEnableWrite();
+void JitEnableExecute();
 }
 
 extern "C" void ARM_Dispatch(ARM* cpu, ARMJIT::JitBlockEntry entry);

src/ARMJIT_A64/ARMJIT_Compiler.cpp

@@ -33,10 +33,6 @@ extern char __start__;
 
 #include <stdlib.h>
 
-#ifdef __APPLE__
-    #include <pthread.h>
-#endif
-
 using namespace Arm64Gen;
 
 extern "C" void ARM_Ret();
@@ -266,7 +262,7 @@ Compiler::Compiler()
     u64 alignedSize = (((u64)JitMem + sizeof(JitMem)) & ~(pageSize - 1)) - (u64)pageAligned;
     #ifdef __APPLE__
         pageAligned = (u8*)mmap(NULL, 1024*1024*16, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS | MAP_JIT,-1, 0);
-        pthread_jit_write_protect_np(false);
+        JitEnableWrite();
     #else
         mprotect(pageAligned, alignedSize, PROT_EXEC | PROT_READ | PROT_WRITE);
     #endif
@@ -685,6 +681,7 @@ void Compiler::Comp_BranchSpecialBehaviour(bool taken)
     {
         RegCache.PrepareExit();
 
+        if (ConstantCycles)
             ADD(RCycles, RCycles, ConstantCycles);
         QuickTailCall(X0, ARM_Ret);
     }
@@ -836,6 +833,7 @@ JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[]
 
     RegCache.Flush();
 
+    if (ConstantCycles)
         ADD(RCycles, RCycles, ConstantCycles);
     QuickTailCall(X0, ARM_Ret);
 

src/ARMJIT_A64/ARMJIT_Compiler.h

@@ -185,7 +185,7 @@ public:
     void T_Comp_BL_LONG_2();
     void T_Comp_BL_Merged();
 
-    s32 Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc, bool decrement, bool usermode);
+    s32 Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc, bool decrement, bool usermode, bool skipLoadingRn);
 
     void Comp_Mul_Mla(bool S, bool mla, Arm64Gen::ARM64Reg rd, Arm64Gen::ARM64Reg rm, Arm64Gen::ARM64Reg rs, Arm64Gen::ARM64Reg rn);
 

src/ARMJIT_A64/ARMJIT_LoadStore.cpp

@@ -465,7 +465,7 @@ void Compiler::T_Comp_MemSPRel()
     Comp_MemAccess(CurInstr.T_Reg(8), 13, Op2(offset), 32, load ? 0 : memop_Store);
 }
 
-s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc, bool decrement, bool usermode)
+s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc, bool decrement, bool usermode, bool skipLoadingRn)
 {
     IrregularCycles = true;
 
@@ -474,7 +474,8 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
     if (regsCount == 0)
         return 0; // actually not the right behaviour TODO: fix me
 
-    if (regsCount == 1 && !usermode && RegCache.LoadedRegs & (1 << *regs.begin()))
+    int firstReg = *regs.begin();
+    if (regsCount == 1 && !usermode && RegCache.LoadedRegs & (1 << firstReg) && !(firstReg == rn && skipLoadingRn))
     {
         int flags = 0;
         if (store)
@@ -483,7 +484,7 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
             flags |= memop_SubtractOffset;
         Op2 offset = preinc ? Op2(4) : Op2(0);
 
-        Comp_MemAccess(*regs.begin(), rn, offset, 32, flags);
+        Comp_MemAccess(firstReg, rn, offset, 32, flags);
 
         return decrement ? -4 : 4;
     }
@@ -539,12 +540,16 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
             loadStoreOffsets[i++] = GetCodeOffset();
 
             if (store)
+            {
                 STR(INDEX_UNSIGNED, first, X1, offset);
-            else
+            }
+            else if (!(reg == rn && skipLoadingRn))
+            {
                 LDR(INDEX_UNSIGNED, first, X1, offset);
 
-            if (!(RegCache.LoadedRegs & (1 << reg)) && !store)
+                if (!(RegCache.LoadedRegs & (1 << reg)))
                     SaveReg(reg, first);
+            }
 
             offset += 4;
         }
@@ -558,13 +563,23 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
 
             ARM64Reg first = W3, second = W4;
             if (RegCache.LoadedRegs & (1 << reg))
+            {
+                if (!(reg == rn && skipLoadingRn))
                     first = MapReg(reg);
+            }
             else if (store)
+            {
                 LoadReg(reg, first);
+            }
             if (RegCache.LoadedRegs & (1 << nextReg))
+            {
+                if (!(nextReg == rn && skipLoadingRn))
                     second = MapReg(nextReg);
+            }
             else if (store)
+            {
                 LoadReg(nextReg, second);
+            }
 
             loadStoreOffsets[i++] = GetCodeOffset();
             if (store)
@@ -705,6 +720,8 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
                 LDR(INDEX_UNSIGNED, W3, SP, i * 8);
                 MOVI2R(W1, reg - 8);
                 BL(WriteBanked);
+                if (!(reg == rn && skipLoadingRn))
+                {
                     FixupBranch alreadyWritten = CBNZ(W4);
                     if (RegCache.LoadedRegs & (1 << reg))
                         MOV(MapReg(reg), W3);
@@ -712,13 +729,14 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
                         SaveReg(reg, W3);
                     SetJumpTarget(alreadyWritten);
                 }
+            }
             else if (!usermode && nextReg != regs.end())
             {
                 ARM64Reg first = W3, second = W4;
                 
-                if (RegCache.LoadedRegs & (1 << reg))
+                if (RegCache.LoadedRegs & (1 << reg) && !(reg == rn && skipLoadingRn))
                     first = MapReg(reg);
-                if (RegCache.LoadedRegs & (1 << *nextReg))
+                if (RegCache.LoadedRegs & (1 << *nextReg) && !(*nextReg == rn && skipLoadingRn))
                     second = MapReg(*nextReg);
 
                 LDP(INDEX_SIGNED, EncodeRegTo64(first), EncodeRegTo64(second), SP, i * 8);
@@ -732,10 +750,13 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
                 i++;
             }
             else if (RegCache.LoadedRegs & (1 << reg))
+            {
+                if (!(reg == rn && skipLoadingRn))
                 {
                     ARM64Reg mapped = MapReg(reg);
                     LDR(INDEX_UNSIGNED, mapped, SP, i * 8);
                 }
+            }
             else
             {
                 LDR(INDEX_UNSIGNED, W3, SP, i * 8);
@@ -778,13 +799,13 @@ void Compiler::A_Comp_LDM_STM()
 
     ARM64Reg rn = MapReg(CurInstr.A_Reg(16));
 
-    s32 offset = Comp_MemAccessBlock(CurInstr.A_Reg(16), regs, !load, pre, !add, usermode);
-
     if (load && writeback && regs[CurInstr.A_Reg(16)])
         writeback = Num == 0
-            ? (!(regs & ~BitSet16(1 << CurInstr.A_Reg(16)))) || (regs & ~BitSet16((2 << CurInstr.A_Reg(16)) - 1))
+            && (!(regs & ~BitSet16(1 << CurInstr.A_Reg(16)))) || (regs & ~BitSet16((2 << CurInstr.A_Reg(16)) - 1));
-            : false;
+
-    if (writeback)
+    s32 offset = Comp_MemAccessBlock(CurInstr.A_Reg(16), regs, !load, pre, !add, usermode, load && writeback);
+
+    if (writeback && offset)
     {
         if (offset > 0)
             ADD(rn, rn, offset);
@@ -806,12 +827,15 @@ void Compiler::T_Comp_PUSH_POP()
     }
 
     ARM64Reg sp = MapReg(13);
-    s32 offset = Comp_MemAccessBlock(13, regs, !load, !load, !load, false);
+    s32 offset = Comp_MemAccessBlock(13, regs, !load, !load, !load, false, false);
 
+    if (offset)
+    {
         if (offset > 0)
             ADD(sp, sp, offset);
         else
             SUB(sp, sp, -offset);
+    }
 }
 
 void Compiler::T_Comp_LDMIA_STMIA()
@@ -821,9 +845,11 @@ void Compiler::T_Comp_LDMIA_STMIA()
     bool load = CurInstr.Instr & (1 << 11);
     u32 regsCount = regs.Count();
 
-    s32 offset = Comp_MemAccessBlock(CurInstr.T_Reg(8), regs, !load, false, false, false);
+    bool writeback = !load || !regs[CurInstr.T_Reg(8)];
 
-    if (!load || !regs[CurInstr.T_Reg(8)])
+    s32 offset = Comp_MemAccessBlock(CurInstr.T_Reg(8), regs, !load, false, false, false, load && writeback);
+
+    if (writeback && offset)
     {
         if (offset > 0)
             ADD(rb, rb, offset);

src/ARMJIT_x64/ARMJIT_Compiler.cpp

@@ -22,6 +22,7 @@
 #include "../Config.h"
 
 #include <assert.h>
+#include <stdarg.h>
 
 #include "../dolphin/CommonFuncs.h"
 
@@ -298,6 +299,10 @@ Compiler::Compiler()
         SetJumpTarget(und);
         MOV(32, R(RSCRATCH3), MComplex(RCPU, RSCRATCH2, SCALE_4, offsetof(ARM, R_UND)));
         RET();
+
+#ifdef JIT_PROFILING_ENABLED
+        CreateMethod("ReadBanked", ReadBanked);
+#endif
     }
     {
         // RSCRATCH  mode
@@ -341,6 +346,10 @@ Compiler::Compiler()
         MOV(32, MComplex(RCPU, RSCRATCH2, SCALE_4, offsetof(ARM, R_UND)), R(RSCRATCH3));
         CLC();
         RET();
+
+#ifdef JIT_PROFILING_ENABLED
+        CreateMethod("WriteBanked", WriteBanked);
+#endif
     }
 
     for (int consoleType = 0; consoleType < 2; consoleType++)
@@ -401,6 +410,10 @@ Compiler::Compiler()
                     ABI_PopRegistersAndAdjustStack(CallerSavedPushRegs, 8);
                     RET();
 
+#ifdef JIT_PROFILING_ENABLED
+                    CreateMethod("FastMemStorePatch%d_%d_%d", PatchedStoreFuncs[consoleType][num][size][reg], num, size, reg);
+#endif
+
                     for (int signextend = 0; signextend < 2; signextend++)
                     {
                         PatchedLoadFuncs[consoleType][num][size][signextend][reg] = GetWritableCodePtr();
@@ -439,6 +452,10 @@ Compiler::Compiler()
                         else
                             MOVZX(32, 8 << size, rdMapped, R(RSCRATCH));
                         RET();
+
+#ifdef JIT_PROFILING_ENABLED
+                        CreateMethod("FastMemLoadPatch%d_%d_%d_%d", PatchedLoadFuncs[consoleType][num][size][signextend][reg], num, size, reg, signextend);
+#endif
                     }
                 }
             }
@@ -663,11 +680,34 @@ void Compiler::Comp_SpecialBranchBehaviour(bool taken)
     {
         RegCache.PrepareExit();
 
+        if (ConstantCycles)
             ADD(32, MDisp(RCPU, offsetof(ARM, Cycles)), Imm32(ConstantCycles));
         JMP((u8*)&ARM_Ret, true);
     }
 }
 
+#ifdef JIT_PROFILING_ENABLED
+void Compiler::CreateMethod(const char* namefmt, void* start, ...)
+{
+    if (iJIT_IsProfilingActive())
+    {
+        va_list args;
+        va_start(args, start);
+        char name[64];
+        vsprintf(name, namefmt, args);
+        va_end(args);
+
+        iJIT_Method_Load method = {0};
+        method.method_id = iJIT_GetNewMethodID();
+        method.method_name = name;
+        method.method_load_address = start;
+        method.method_size = GetWritableCodePtr() - (u8*)start;
+
+        iJIT_NotifyEvent(iJVM_EVENT_TYPE_METHOD_LOAD_FINISHED, (void*)&method);
+    }
+}
+#endif
+
 JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[], int instrsCount, bool hasMemoryInstr)
 {
     if (NearSize - (GetCodePtr() - NearStart) < 1024 * 32) // guess...
@@ -802,9 +842,14 @@ JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[]
 
     RegCache.Flush();
 
+    if (ConstantCycles)
         ADD(32, MDisp(RCPU, offsetof(ARM, Cycles)), Imm32(ConstantCycles));
     JMP((u8*)ARM_Ret, true);
 
+#ifdef JIT_PROFILING_ENABLED
+    CreateMethod("JIT_Block_%d_%d_%08X", (void*)res, Num, Thumb, instrs[0].Addr);
+#endif
+
     /*FILE* codeout = fopen("codeout", "a");
     fprintf(codeout, "beginning block argargarg__ %x!!!", instrs[0].Addr);
     fwrite((u8*)res, GetWritableCodePtr() - (u8*)res, 1, codeout);

src/ARMJIT_x64/ARMJIT_Compiler.h

@@ -25,6 +25,10 @@
 #include "../ARMJIT_Internal.h"
 #include "../ARMJIT_RegisterCache.h"
 
+#ifdef JIT_PROFILING_ENABLED
+#include <jitprofiling.h>
+#endif
+
 #include <unordered_map>
 
 namespace ARMJIT
@@ -163,7 +167,7 @@ public:
         memop_SubtractOffset = 1 << 4
     };
     void Comp_MemAccess(int rd, int rn, const Op2& op2, int size, int flags);
-    s32 Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc, bool decrement, bool usermode);
+    s32 Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc, bool decrement, bool usermode, bool skipLoadingRn);
     bool Comp_MemLoadLiteral(int size, bool signExtend, int rd, u32 addr);
 
     void Comp_ArithTriOp(void (Compiler::*op)(int, const Gen::OpArg&, const Gen::OpArg&), 
@@ -230,6 +234,10 @@ public:
 
     u8* RewriteMemAccess(u8* pc);
 
+#ifdef JIT_PROFILING_ENABLED
+    void CreateMethod(const char* namefmt, void* start, ...);
+#endif
+
     u8* FarCode;
     u8* NearCode;
     u32 FarSize;

src/ARMJIT_x64/ARMJIT_LoadStore.cpp

@@ -399,14 +399,15 @@ void Compiler::Comp_MemAccess(int rd, int rn, const Op2& op2, int size, int flag
     }
 }
 
-s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc, bool decrement, bool usermode)
+s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc, bool decrement, bool usermode, bool skipLoadingRn)
 {
     int regsCount = regs.Count();
 
     if (regsCount == 0)
         return 0; // actually not the right behaviour TODO: fix me
 
-    if (regsCount == 1 && !usermode && RegCache.LoadedRegs & (1 << *regs.begin()))
+    int firstReg = *regs.begin();
+    if (regsCount == 1 && !usermode && RegCache.LoadedRegs & (1 << firstReg) && !(firstReg == rn && skipLoadingRn))
     {
         int flags = 0;
         if (store)
@@ -415,7 +416,7 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
             flags |= memop_SubtractOffset;
         Op2 offset = preinc ? Op2(4) : Op2(0);
 
-        Comp_MemAccess(*regs.begin(), rn, offset, 32, flags);
+        Comp_MemAccess(firstReg, rn, offset, 32, flags);
 
         return decrement ? -4 : 4;
     }
@@ -482,7 +483,10 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
             {
                 if (RegCache.LoadedRegs & (1 << reg))
                 {
+                    if (!(reg == rn && skipLoadingRn))
                         MOV(32, MapReg(reg), mem);
+                    else
+                        MOV(32, R(RSCRATCH), mem); // just touch the memory
                 }
                 else
                 {
@@ -548,6 +552,8 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
                 MOV(32, R(RSCRATCH2), Imm32(reg - 8));
                 POP(RSCRATCH3);
                 CALL(WriteBanked);
+                if (!(reg == rn && skipLoadingRn))
+                {
                     FixupBranch sucessfulWritten = J_CC(CC_NC);
                     if (RegCache.LoadedRegs & (1 << reg))
                             MOV(32, R(RegCache.Mapping[reg]), R(RSCRATCH3));
@@ -555,6 +561,7 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
                         SaveReg(reg, RSCRATCH3);
                     SetJumpTarget(sucessfulWritten);
                 }
+            }
             else if (!(RegCache.LoadedRegs & (1 << reg)))
             {
                 assert(reg != 15);
@@ -562,6 +569,10 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
                 POP(RSCRATCH);
                 SaveReg(reg, RSCRATCH);
             }
+            else if (reg == rn && skipLoadingRn)
+            {
+                ADD(64, R(RSP), Imm8(8));
+            }
             else
             {
                 POP(MapReg(reg).GetSimpleReg());
@@ -748,14 +759,14 @@ void Compiler::A_Comp_LDM_STM()
 
     OpArg rn = MapReg(CurInstr.A_Reg(16));
 
-    s32 offset = Comp_MemAccessBlock(CurInstr.A_Reg(16), regs, !load, pre, !add, usermode);
-
     if (load && writeback && regs[CurInstr.A_Reg(16)])
         writeback = Num == 0
-            ? (!(regs & ~BitSet16(1 << CurInstr.A_Reg(16)))) || (regs & ~BitSet16((2 << CurInstr.A_Reg(16)) - 1))
+            && (!(regs & ~BitSet16(1 << CurInstr.A_Reg(16)))) || (regs & ~BitSet16((2 << CurInstr.A_Reg(16)) - 1));
-            : false;
+
-    if (writeback)
+    s32 offset = Comp_MemAccessBlock(CurInstr.A_Reg(16), regs, !load, pre, !add, usermode, load && writeback);
-        ADD(32, rn, offset >= INT8_MIN && offset < INT8_MAX ? Imm8(offset) : Imm32(offset));
+
+    if (writeback && offset)
+        ADD(32, rn, Imm32(offset));
 }
 
 void Compiler::T_Comp_MemImm()
@@ -825,8 +836,9 @@ void Compiler::T_Comp_PUSH_POP()
     }
 
     OpArg sp = MapReg(13);
-    s32 offset = Comp_MemAccessBlock(13, regs, !load, !load, !load, false);
+    s32 offset = Comp_MemAccessBlock(13, regs, !load, !load, !load, false, false);
 
+    if (offset)
         ADD(32, sp, Imm8(offset)); // offset will be always be in range since PUSH accesses 9 regs max
 }
 
@@ -836,9 +848,11 @@ void Compiler::T_Comp_LDMIA_STMIA()
     OpArg rb = MapReg(CurInstr.T_Reg(8));
     bool load = CurInstr.Instr & (1 << 11);
 
-    s32 offset = Comp_MemAccessBlock(CurInstr.T_Reg(8), regs, !load, false, false, false);
+    bool writeback = !load || !regs[CurInstr.T_Reg(8)];
 
-    if (!load || !regs[CurInstr.T_Reg(8)])
+    s32 offset = Comp_MemAccessBlock(CurInstr.T_Reg(8), regs, !load, false, false, false, load && writeback);
+
+    if (writeback && offset)
         ADD(32, rb, Imm8(offset));
 }
 

src/CMakeLists.txt

@@ -124,3 +124,7 @@ else()
 		target_link_libraries(core rt)
 	endif()
 endif()
+
+if (ENABLE_JIT_PROFILING)
+	target_link_libraries(core jitprofiling)
+endif()

src/DSi.h

@@ -60,6 +60,7 @@ bool Init();
 void DeInit();
 void Reset();
 
+void SetupDirectBoot();
 void SoftReset();
 
 bool LoadBIOS();

src/NDS.cpp

@@ -164,7 +164,6 @@ bool Running;
 
 bool RunningGame;
 
-
 void DivDone(u32 param);
 void SqrtDone(u32 param);
 void RunTimer(u32 tid, s32 cycles);
@@ -342,8 +341,13 @@ void SetupDirectBoot()
 {
     if (ConsoleType == 1)
     {
-        printf("!! DIRECT BOOT NOT SUPPORTED IN DSI MODE\n");
+        // With the BIOS select in SCFG_BIOS and the initialization od
-        return;
+        // SCFG_BIOS depending on the Header->UnitType, we can now boot 
+        // directly in the roms.
+        // There are some more SCFG Settings that change depending on
+        // the unit type, so this is experimental
+        printf("!! DIRECT BOOT NOT STABLE IN DSI MODE\n");
+        DSi::SetupDirectBoot();
     }
 
     u32 bootparams[8];

src/frontend/qt_sdl/main.cpp

@@ -106,6 +106,8 @@ u32 micExtBufferWritePos;
 u32 micWavLength;
 s16* micWavBuffer;
 
+void micCallback(void* data, Uint8* stream, int len);
+
 
 void audioCallback(void* data, Uint8* stream, int len)
 {
@@ -143,6 +145,40 @@ void audioCallback(void* data, Uint8* stream, int len)
 }
 
 
+void micOpen()
+{
+    if (Config::MicInputType != 1)
+    {
+        micDevice = 0;
+        return;
+    }
+
+    SDL_AudioSpec whatIwant, whatIget;
+    memset(&whatIwant, 0, sizeof(SDL_AudioSpec));
+    whatIwant.freq = 44100;
+    whatIwant.format = AUDIO_S16LSB;
+    whatIwant.channels = 1;
+    whatIwant.samples = 1024;
+    whatIwant.callback = micCallback;
+    micDevice = SDL_OpenAudioDevice(NULL, 1, &whatIwant, &whatIget, 0);
+    if (!micDevice)
+    {
+        printf("Mic init failed: %s\n", SDL_GetError());
+    }
+    else
+    {
+        SDL_PauseAudioDevice(micDevice, 0);
+    }
+}
+
+void micClose()
+{
+    if (micDevice)
+        SDL_CloseAudioDevice(micDevice);
+
+    micDevice = 0;
+}
+
 void micLoadWav(const char* name)
 {
     SDL_AudioSpec format;
@@ -625,7 +661,7 @@ void EmuThread::emuRun()
     // checkme
     emit windowEmuStart();
     if (audioDevice) SDL_PauseAudioDevice(audioDevice, 0);
-    if (micDevice)   SDL_PauseAudioDevice(micDevice, 0);
+    micOpen();
 }
 
 void EmuThread::emuPause()
@@ -638,7 +674,7 @@ void EmuThread::emuPause()
     while (EmuStatus != 2);
 
     if (audioDevice) SDL_PauseAudioDevice(audioDevice, 1);
-    if (micDevice)   SDL_PauseAudioDevice(micDevice, 1);
+    micClose();
 }
 
 void EmuThread::emuUnpause()
@@ -651,7 +687,7 @@ void EmuThread::emuUnpause()
     EmuRunning = PrevEmuStatus;
 
     if (audioDevice) SDL_PauseAudioDevice(audioDevice, 0);
-    if (micDevice)   SDL_PauseAudioDevice(micDevice, 0);
+    micOpen();
 }
 
 void EmuThread::emuStop()
@@ -660,7 +696,7 @@ void EmuThread::emuStop()
     EmuPause = 0;
 
     if (audioDevice) SDL_PauseAudioDevice(audioDevice, 1);
-    if (micDevice)   SDL_PauseAudioDevice(micDevice, 1);
+    micClose();
 }
 
 void EmuThread::emuFrameStep()
@@ -2401,6 +2437,8 @@ void MainWindow::onUpdateAudioSettings()
 
 void MainWindow::onAudioSettingsFinished(int res)
 {
+    micClose();
+
     SPU::SetInterpolation(Config::AudioInterp);
 
     if (Config::MicInputType == 3)
@@ -2418,6 +2456,8 @@ void MainWindow::onAudioSettingsFinished(int res)
         else
             Frontend::Mic_SetExternalBuffer(NULL, 0);
     }
+
+    micOpen();
 }
 
 void MainWindow::onOpenWifiSettings()
@@ -2762,21 +2802,7 @@ int main(int argc, char** argv)
         SDL_PauseAudioDevice(audioDevice, 1);
     }
 
-    memset(&whatIwant, 0, sizeof(SDL_AudioSpec));
+    micDevice = 0;
-    whatIwant.freq = 44100;
-    whatIwant.format = AUDIO_S16LSB;
-    whatIwant.channels = 1;
-    whatIwant.samples = 1024;
-    whatIwant.callback = micCallback;
-    micDevice = SDL_OpenAudioDevice(NULL, 1, &whatIwant, &whatIget, 0);
-    if (!micDevice)
-    {
-        printf("Mic init failed: %s\n", SDL_GetError());
-    }
-    else
-    {
-        SDL_PauseAudioDevice(micDevice, 1);
-    }
 
 
     memset(micExtBuffer, 0, sizeof(micExtBuffer));
@@ -2847,7 +2873,7 @@ int main(int argc, char** argv)
     Frontend::DeInit_ROM();
 
     if (audioDevice) SDL_CloseAudioDevice(audioDevice);
-    if (micDevice)   SDL_CloseAudioDevice(micDevice);
+    micClose();
 
     SDL_DestroyCond(audioSync);
     SDL_DestroyMutex(audioSyncLock);