JIT A64: fixes

also update Switch code for latest libnx

3 files changed, 38 insertions, 31 deletions

diff --git a/src/ARMJIT_A64/ARMJIT_Compiler.cpp b/src/ARMJIT_A64/ARMJIT_Compiler.cpp
index 5fe3fe7..880a6fc 100644
--- a/src/ARMJIT_A64/ARMJIT_Compiler.cpp
+++ b/src/ARMJIT_A64/ARMJIT_Compiler.cpp
@@ -187,7 +187,8 @@ Compiler::Compiler()
     JitRWBase = aligned_alloc(0x1000, JitMemSize);
 
     JitRXStart = (u8*)&__start__ - JitMemSize - 0x1000;
-    JitRWStart = virtmemReserve(JitMemSize);
+    virtmemLock();
+    JitRWStart = virtmemFindAslr(JitMemSize, 0x1000);
     MemoryInfo info = {0};
     u32 pageInfo = {0};
     int i = 0;
@@ -214,6 +215,8 @@ Compiler::Compiler()
     succeded = R_SUCCEEDED(svcMapProcessMemory(JitRWStart, envGetOwnProcessHandle(), (u64)JitRXStart, JitMemSize));
     assert(succeded);
 
+    virtmemUnlock();
+
     SetCodeBase((u8*)JitRWStart, (u8*)JitRXStart);
     JitMemMainSize = JitMemSize;
 #else
@@ -426,7 +429,6 @@ Compiler::~Compiler()
     {
         bool succeded = R_SUCCEEDED(svcUnmapProcessMemory(JitRWStart, envGetOwnProcessHandle(), (u64)JitRXStart, JitMemSize));
         assert(succeded);
-        virtmemFree(JitRWStart, JitMemSize);
         succeded = R_SUCCEEDED(svcUnmapProcessCodeMemory(envGetOwnProcessHandle(), (u64)JitRXStart, (u64)JitRWBase, JitMemSize));
         assert(succeded);
         free(JitRWBase);
@@ -753,7 +755,8 @@ JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[]
                         FixupBranch skipNop = B();
                         SetJumpTarget(skipExecute);
 
-                        Comp_AddCycles_C();
+                        if (IrregularCycles)
+                            Comp_AddCycles_C(true);
 
                         Comp_BranchSpecialBehaviour(false);
 
diff --git a/src/ARMJIT_A64/ARMJIT_LoadStore.cpp b/src/ARMJIT_A64/ARMJIT_LoadStore.cpp
index 2c14dc6..3d30759 100644
--- a/src/ARMJIT_A64/ARMJIT_LoadStore.cpp
+++ b/src/ARMJIT_A64/ARMJIT_LoadStore.cpp
@@ -472,31 +472,24 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
     bool compileFastPath = Config::JIT_FastMemory
         && store && !usermode && (CurInstr.Cond() < 0xE || ARMJIT_Memory::IsFastmemCompatible(expectedTarget));
 
-    if (decrement)
     {
-        s32 offset = -regsCount * 4 + (preinc ? 0 : 4);
+        s32 offset = decrement
+            ? -regsCount * 4 + (preinc ? 0 : 4)
+            : (preinc ? 4 : 0);
+
         if (offset)
-        {
             ADDI2R(W0, MapReg(rn), offset);
-            ANDI2R(W0, W0, ~3);
-        }
-        else
-        {
+        else if (compileFastPath)
             ANDI2R(W0, MapReg(rn), ~3);
-        }
-    }
-    else
-    {
-        ANDI2R(W0, MapReg(rn), ~3);
-        if (preinc)
-            ADD(W0, W0, 4);
+        else
+            MOV(W0, MapReg(rn));
     }
 
     u8* patchFunc;
     if (compileFastPath)
     {
         ptrdiff_t fastPathStart = GetCodeOffset();
-        ptrdiff_t loadStoreOffsets[16];
+        ptrdiff_t loadStoreOffsets[8];
 
         MOVP2R(X1, Num == 0 ? ARMJIT_Memory::FastMem9Start : ARMJIT_Memory::FastMem7Start);
         ADD(X1, X1, X0);
@@ -547,16 +540,19 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
                 LoadReg(nextReg, second);
 
             loadStoreOffsets[i++] = GetCodeOffset();
-
             if (store)
+            {
                 STP(INDEX_SIGNED, first, second, X1, offset);
+            }
             else
+            {
                 LDP(INDEX_SIGNED, first, second, X1, offset);
-
-            if (!(RegCache.LoadedRegs & (1 << reg)) && !store)
-                SaveReg(reg, first);
-            if (!(RegCache.LoadedRegs & (1 << nextReg)) && !store)
-                SaveReg(nextReg, second);
+            
+                if (!(RegCache.LoadedRegs & (1 << reg)))
+                    SaveReg(reg, first);
+                if (!(RegCache.LoadedRegs & (1 << nextReg)))
+                    SaveReg(nextReg, second);
+            }
 
             offset += 8;
         }
@@ -566,7 +562,8 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
         SwapCodeRegion();
         patchFunc = (u8*)GetRXPtr();
         patch.PatchFunc = patchFunc;
-        for (i = 0; i < regsCount; i++)
+        u32 numLoadStores = i;
+        for (i = 0; i < numLoadStores; i++)
         {
             patch.PatchOffset = fastPathStart - loadStoreOffsets[i];
             LoadStorePatches[loadStoreOffsets[i]] = patch;
diff --git a/src/ARMJIT_Memory.cpp b/src/ARMJIT_Memory.cpp
index 33d6bcf..5539880 100644
--- a/src/ARMJIT_Memory.cpp
+++ b/src/ARMJIT_Memory.cpp
@@ -275,6 +275,7 @@ u8 MappingStatus9[1 << (32-12)];
 u8 MappingStatus7[1 << (32-12)];
 
 #if defined(__SWITCH__)
+VirtmemReservation* FastMem9Reservation, *FastMem7Reservation;
 u8* MemoryBase;
 u8* MemoryBaseCodeMem;
 #elif defined(_WIN32)
@@ -676,7 +677,8 @@ void Init()
 {
 #if defined(__SWITCH__)
     MemoryBase = (u8*)aligned_alloc(0x1000, MemoryTotalSize);
-    MemoryBaseCodeMem = (u8*)virtmemReserve(MemoryTotalSize);
+    virtmemLock();
+    MemoryBaseCodeMem = (u8*)virtmemFindCodeMemory(MemoryTotalSize, 0x1000);
 
     bool succeded = R_SUCCEEDED(svcMapProcessCodeMemory(envGetOwnProcessHandle(), (u64)MemoryBaseCodeMem, 
         (u64)MemoryBase, MemoryTotalSize));
@@ -686,11 +688,15 @@ void Init()
     assert(succeded);
 
     // 8 GB of address space, just don't ask...
-    FastMem9Start = virtmemReserve(AddrSpaceSize);
+    FastMem9Start = virtmemFindAslr(AddrSpaceSize, 0x1000);
     assert(FastMem9Start);
-    FastMem7Start = virtmemReserve(AddrSpaceSize);
+    FastMem7Start = virtmemFindAslr(AddrSpaceSize, 0x1000);
     assert(FastMem7Start);
 
+    FastMem9Reservation = virtmemAddReservation(FastMem9Start, AddrSpaceSize);
+    FastMem7Reservation = virtmemAddReservation(FastMem7Start, AddrSpaceSize);
+    virtmemUnlock();
+
     u8* basePtr = MemoryBaseCodeMem;
 #elif defined(_WIN32)
     ExceptionHandlerHandle = AddVectoredExceptionHandler(1, ExceptionHandler);
@@ -775,11 +781,12 @@ void Init()
 void DeInit()
 {
 #if defined(__SWITCH__)
-    virtmemFree(FastMem9Start, AddrSpaceSize);
-    virtmemFree(FastMem7Start, AddrSpaceSize);
+    virtmemLock();
+    virtmemRemoveReservation(FastMem9Reservation);
+    virtmemRemoveReservation(FastMem7Reservation);
+    virtmemUnlock();
 
     svcUnmapProcessCodeMemory(envGetOwnProcessHandle(), (u64)MemoryBaseCodeMem, (u64)MemoryBase, MemoryTotalSize);
-    virtmemFree(MemoryBaseCodeMem, MemoryTotalSize);
     free(MemoryBase);
 #elif defined(__APPLE__)
     char* fastmemPidName = new char[snprintf(NULL, 0, "melondsfastmem%d", getpid()) + 1];