diff options
| -rw-r--r-- | src/ARMJIT_A64/ARMJIT_LoadStore.cpp | 27 | ||||
| -rw-r--r-- | src/ARMJIT_Memory.cpp | 48 | ||||
| -rw-r--r-- | src/ARMJIT_x64/ARMJIT_Compiler.cpp | 2 | ||||
| -rw-r--r-- | src/ARMJIT_x64/ARMJIT_LoadStore.cpp | 147 | ||||
| -rw-r--r-- | src/dolphin/Arm64Emitter.cpp | 5 | ||||
| -rw-r--r-- | src/dolphin/Arm64Emitter.h | 1 |
6 files changed, 139 insertions, 91 deletions
diff --git a/src/ARMJIT_A64/ARMJIT_LoadStore.cpp b/src/ARMJIT_A64/ARMJIT_LoadStore.cpp index c1b23a7..bdd9f43 100644 --- a/src/ARMJIT_A64/ARMJIT_LoadStore.cpp +++ b/src/ARMJIT_A64/ARMJIT_LoadStore.cpp @@ -474,19 +474,17 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc ANDI2R(W0, MapReg(rn), ~3); } - LoadStorePatch patch; if (compileFastPath) { ptrdiff_t fastPathStart = GetCodeOffset(); - ptrdiff_t firstLoadStoreOffset; - - bool firstLoadStore = true; + ptrdiff_t loadStoreOffsets[16]; MOVP2R(X1, Num == 0 ? ARMJIT_Memory::FastMem9Start : ARMJIT_Memory::FastMem7Start); ADD(X1, X1, X0); u32 offset = preinc ? 4 : 0; BitSet16::Iterator it = regs.begin(); + u32 i = 0; if (regsCount & 1) { @@ -499,11 +497,7 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc else if (store) LoadReg(reg, first); - if (firstLoadStore) - { - firstLoadStoreOffset = GetCodeOffset(); - firstLoadStore = false; - } + loadStoreOffsets[i++] = GetCodeOffset(); if (store) STR(INDEX_UNSIGNED, first, X1, offset); @@ -533,11 +527,7 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc else if (store) LoadReg(nextReg, second); - if (firstLoadStore) - { - firstLoadStoreOffset = GetCodeOffset(); - firstLoadStore = false; - } + loadStoreOffsets[i++] = GetCodeOffset(); if (store) STP(INDEX_SIGNED, first, second, X1, offset); @@ -552,12 +542,15 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc offset += 8; } + LoadStorePatch patch; patch.PatchSize = GetCodeOffset() - fastPathStart; - patch.PatchOffset = fastPathStart - firstLoadStoreOffset; SwapCodeRegion(); patch.PatchFunc = GetRXPtr(); - - LoadStorePatches[firstLoadStoreOffset] = patch; + for (i = 0; i < regsCount; i++) + { + patch.PatchOffset = fastPathStart - loadStoreOffsets[i]; + LoadStorePatches[loadStoreOffsets[i]] = patch; + } ABI_PushRegisters({30}); } diff --git a/src/ARMJIT_Memory.cpp b/src/ARMJIT_Memory.cpp index 4a8b8a5..4e1b0e0 100644 --- a/src/ARMJIT_Memory.cpp +++ b/src/ARMJIT_Memory.cpp @@ -73,7 +73,8 @@ u64 __nx_exception_stack_size = 0x8000; void __libnx_exception_handler(ThreadExceptionDump* ctx) { ARMJIT_Memory::FaultDescription desc; - desc.EmulatedFaultAddr = ctx->cpu_gprs[0].w; + u8* curArea = (u8*)(NDS::CurCPU == 0 ? ARMJIT_Memory::FastMem9Start : ARMJIT_Memory::FastMem7Start); + desc.EmulatedFaultAddr = (u8*)ctx->far.x - curArea; desc.FaultPC = ctx->pc.x; u64 integerRegisters[33]; @@ -109,10 +110,14 @@ void __libnx_exception_handler(ThreadExceptionDump* ctx) static LONG ExceptionHandler(EXCEPTION_POINTERS* exceptionInfo) { if (exceptionInfo->ExceptionRecord->ExceptionCode != EXCEPTION_ACCESS_VIOLATION) + { + printf("narg\n"); return EXCEPTION_CONTINUE_SEARCH; + } ARMJIT_Memory::FaultDescription desc; - desc.EmulatedFaultAddr = exceptionInfo->ContextRecord->Rcx; + u8* curArea = (u8*)(NDS::CurCPU == 0 ? ARMJIT_Memory::FastMem9Start : ARMJIT_Memory::FastMem7Start); + desc.EmulatedFaultAddr = (u8*)exceptionInfo->ExceptionRecord->ExceptionInformation[1] - curArea; desc.FaultPC = exceptionInfo->ContextRecord->Rip; s32 offset = 0; @@ -122,6 +127,7 @@ static LONG ExceptionHandler(EXCEPTION_POINTERS* exceptionInfo) return EXCEPTION_CONTINUE_EXECUTION; } + printf("miauz\n"); return EXCEPTION_CONTINUE_SEARCH; } @@ -277,7 +283,8 @@ void SetCodeProtectionRange(u32 addr, u32 size, u32 num, int protection) winProtection = PAGE_READONLY; else winProtection = PAGE_READWRITE; - VirtualProtect(dst, size, winProtection, &oldProtection); + bool success = VirtualProtect(dst, size, winProtection, &oldProtection); + assert(success); #else int posixProt; if (protection == 0) @@ -348,9 +355,10 @@ void SetCodeProtection(int region, u32 offset, bool protect) { Mapping& mapping = Mappings[region][i]; +// if (offset < mapping.LocalOffset || offset >= mapping.LocalOffset + mapping.Size) +// continue; + u32 effectiveAddr = mapping.Addr + (offset - mapping.LocalOffset); - if (offset < mapping.LocalOffset || offset >= mapping.LocalOffset + mapping.Size) - continue; if (mapping.Num == 0 && region != memregion_DTCM && effectiveAddr >= NDS::ARM9->DTCMBase @@ -401,8 +409,8 @@ void RemapDTCM(u32 newBase, u32 newSize) printf("mapping %d %x %x %x %x\n", region, mapping.Addr, mapping.Size, mapping.Num, mapping.LocalOffset); - bool oldOverlap = NDS::ARM9->DTCMSize > 0 && !(oldDTCMBase >= end || oldDTCBEnd < start); - bool newOverlap = newSize > 0 && !(newBase >= end || newEnd < start); + bool oldOverlap = NDS::ARM9->DTCMSize > 0 && !(oldDTCMBase >= end || oldDTCBEnd <= start); + bool newOverlap = newSize > 0 && !(newBase >= end || newEnd <= start); if (mapping.Num == 0 && (oldOverlap || newOverlap)) { @@ -449,24 +457,22 @@ void RemapNWRAM(int num) void RemapSWRAM() { printf("remapping SWRAM\n"); + for (int i = 0; i < Mappings[memregion_WRAM7].Length;) + { + Mapping& mapping = Mappings[memregion_WRAM7][i]; + if (mapping.Addr + mapping.Size < 0x03800000) + { + mapping.Unmap(memregion_WRAM7); + Mappings[memregion_WRAM7].Remove(i); + } + else + i++; + } for (int i = 0; i < Mappings[memregion_SharedWRAM].Length; i++) { Mappings[memregion_SharedWRAM][i].Unmap(memregion_SharedWRAM); } Mappings[memregion_SharedWRAM].Clear(); - for (int i = 0; i < Mappings[memregion_WRAM7].Length; i++) - { - Mappings[memregion_WRAM7][i].Unmap(memregion_WRAM7); - } - Mappings[memregion_WRAM7].Clear(); - for (int j = 0; j < 3; j++) - { - for (int i = 0; i < Mappings[memregion_NewSharedWRAM_A + j].Length; i++) - { - Mappings[memregion_NewSharedWRAM_A + j][i].Unmap(memregion_NewSharedWRAM_A + j); - } - Mappings[memregion_NewSharedWRAM_A + j].Clear(); - } } bool MapAtAddress(u32 addr) @@ -687,8 +693,6 @@ bool IsFastmemCompatible(int region) || region == memregion_NewSharedWRAM_C) return false; #endif - if (region == memregion_DTCM) - return false; return OffsetsPerRegion[region] != UINT32_MAX; } diff --git a/src/ARMJIT_x64/ARMJIT_Compiler.cpp b/src/ARMJIT_x64/ARMJIT_Compiler.cpp index d8bdd56..6905010 100644 --- a/src/ARMJIT_x64/ARMJIT_Compiler.cpp +++ b/src/ARMJIT_x64/ARMJIT_Compiler.cpp @@ -633,7 +633,7 @@ void Compiler::Comp_SpecialBranchBehaviour(bool taken) JitBlockEntry Compiler::CompileBlock(ARM* cpu, bool thumb, FetchedInstr instrs[], int instrsCount) { - if (NearSize - (NearCode - NearStart) < 1024 * 32) // guess... + if (NearSize - (GetCodePtr() - NearStart) < 1024 * 32) // guess... { printf("near reset\n"); ResetBlockCache(); diff --git a/src/ARMJIT_x64/ARMJIT_LoadStore.cpp b/src/ARMJIT_x64/ARMJIT_LoadStore.cpp index ff15729..aa84105 100644 --- a/src/ARMJIT_x64/ARMJIT_LoadStore.cpp +++ b/src/ARMJIT_x64/ARMJIT_LoadStore.cpp @@ -216,9 +216,17 @@ void Compiler::Comp_MemAccess(int rd, int rn, const Op2& op2, int size, int flag if (size == 32) { - AND(32, R(RSCRATCH3), Imm8(0x3)); - SHL(32, R(RSCRATCH3), Imm8(3)); - ROR_(32, rdMapped, R(RSCRATCH3)); + if (addrIsStatic) + { + if (staticAddress & 0x3) + ROR_(32, rdMapped, Imm8((staticAddress & 0x3) * 8)); + } + else + { + AND(32, R(RSCRATCH3), Imm8(0x3)); + SHL(32, R(RSCRATCH3), Imm8(3)); + ROR_(32, rdMapped, R(RSCRATCH3)); + } } } @@ -233,78 +241,115 @@ void Compiler::Comp_MemAccess(int rd, int rn, const Op2& op2, int size, int flag { PushRegs(false); - if (Num == 0) + void* func = NULL; + if (addrIsStatic) + func = ARMJIT_Memory::GetFuncForAddr(CurCPU, staticAddress, flags & memop_Store, size); + + if (func) { - MOV(64, R(ABI_PARAM2), R(RCPU)); + AND(32, R(RSCRATCH3), Imm8(addressMask)); + if (ABI_PARAM1 != RSCRATCH3) MOV(32, R(ABI_PARAM1), R(RSCRATCH3)); if (flags & memop_Store) - { - MOV(32, R(ABI_PARAM3), rdMapped); + MOV(32, R(ABI_PARAM2), rdMapped); + + ABI_CallFunction((void (*)())func); + + PopRegs(false); - switch (size | NDS::ConsoleType) + if (!(flags & memop_Store)) + { + if (size == 32) { - case 32: CALL((void*)&SlowWrite9<u32, 0>); break; - case 16: CALL((void*)&SlowWrite9<u16, 0>); break; - case 8: CALL((void*)&SlowWrite9<u8, 0>); break; - case 33: CALL((void*)&SlowWrite9<u32, 1>); break; - case 17: CALL((void*)&SlowWrite9<u16, 1>); break; - case 9: CALL((void*)&SlowWrite9<u8, 1>); break; + MOV(32, rdMapped, R(RSCRATCH)); + if (staticAddress & 0x3) + ROR_(32, rdMapped, Imm8((staticAddress & 0x3) * 8)); } - } - else - { - switch (size | NDS::ConsoleType) + else { - case 32: CALL((void*)&SlowRead9<u32, 0>); break; - case 16: CALL((void*)&SlowRead9<u16, 0>); break; - case 8: CALL((void*)&SlowRead9<u8, 0>); break; - case 33: CALL((void*)&SlowRead9<u32, 1>); break; - case 17: CALL((void*)&SlowRead9<u16, 1>); break; - case 9: CALL((void*)&SlowRead9<u8, 1>); break; + if (flags & memop_SignExtend) + MOVSX(32, size, rdMapped.GetSimpleReg(), R(RSCRATCH)); + else + MOVZX(32, size, rdMapped.GetSimpleReg(), R(RSCRATCH)); } } } else { - if (ABI_PARAM1 != RSCRATCH3) - MOV(32, R(ABI_PARAM1), R(RSCRATCH3)); - if (flags & memop_Store) + if (Num == 0) { - MOV(32, R(ABI_PARAM2), rdMapped); + MOV(64, R(ABI_PARAM2), R(RCPU)); + if (ABI_PARAM1 != RSCRATCH3) + MOV(32, R(ABI_PARAM1), R(RSCRATCH3)); + if (flags & memop_Store) + { + MOV(32, R(ABI_PARAM3), rdMapped); - switch (size | NDS::ConsoleType) + switch (size | NDS::ConsoleType) + { + case 32: CALL((void*)&SlowWrite9<u32, 0>); break; + case 16: CALL((void*)&SlowWrite9<u16, 0>); break; + case 8: CALL((void*)&SlowWrite9<u8, 0>); break; + case 33: CALL((void*)&SlowWrite9<u32, 1>); break; + case 17: CALL((void*)&SlowWrite9<u16, 1>); break; + case 9: CALL((void*)&SlowWrite9<u8, 1>); break; + } + } + else { - case 32: CALL((void*)&SlowWrite7<u32, 0>); break; - case 16: CALL((void*)&SlowWrite7<u16, 0>); break; - case 8: CALL((void*)&SlowWrite7<u8, 0>); break; - case 33: CALL((void*)&SlowWrite7<u32, 1>); break; - case 17: CALL((void*)&SlowWrite7<u16, 1>); break; - case 9: CALL((void*)&SlowWrite7<u8, 1>); break; + switch (size | NDS::ConsoleType) + { + case 32: CALL((void*)&SlowRead9<u32, 0>); break; + case 16: CALL((void*)&SlowRead9<u16, 0>); break; + case 8: CALL((void*)&SlowRead9<u8, 0>); break; + case 33: CALL((void*)&SlowRead9<u32, 1>); break; + case 17: CALL((void*)&SlowRead9<u16, 1>); break; + case 9: CALL((void*)&SlowRead9<u8, 1>); break; + } } } else { - switch (size | NDS::ConsoleType) + if (ABI_PARAM1 != RSCRATCH3) + MOV(32, R(ABI_PARAM1), R(RSCRATCH3)); + if (flags & memop_Store) { - case 32: CALL((void*)&SlowRead7<u32, 0>); break; - case 16: CALL((void*)&SlowRead7<u16, 0>); break; - case 8: CALL((void*)&SlowRead7<u8, 0>); break; - case 33: CALL((void*)&SlowRead7<u32, 1>); break; - case 17: CALL((void*)&SlowRead7<u16, 1>); break; - case 9: CALL((void*)&SlowRead7<u8, 1>); break; + MOV(32, R(ABI_PARAM2), rdMapped); + + switch (size | NDS::ConsoleType) + { + case 32: CALL((void*)&SlowWrite7<u32, 0>); break; + case 16: CALL((void*)&SlowWrite7<u16, 0>); break; + case 8: CALL((void*)&SlowWrite7<u8, 0>); break; + case 33: CALL((void*)&SlowWrite7<u32, 1>); break; + case 17: CALL((void*)&SlowWrite7<u16, 1>); break; + case 9: CALL((void*)&SlowWrite7<u8, 1>); break; + } + } + else + { + switch (size | NDS::ConsoleType) + { + case 32: CALL((void*)&SlowRead7<u32, 0>); break; + case 16: CALL((void*)&SlowRead7<u16, 0>); break; + case 8: CALL((void*)&SlowRead7<u8, 0>); break; + case 33: CALL((void*)&SlowRead7<u32, 1>); break; + case 17: CALL((void*)&SlowRead7<u16, 1>); break; + case 9: CALL((void*)&SlowRead7<u8, 1>); break; + } } } - } - PopRegs(false); - - if (!(flags & memop_Store)) - { - if (flags & memop_SignExtend) - MOVSX(32, size, rdMapped.GetSimpleReg(), R(RSCRATCH)); - else - MOVZX(32, size, rdMapped.GetSimpleReg(), R(RSCRATCH)); + PopRegs(false); + + if (!(flags & memop_Store)) + { + if (flags & memop_SignExtend) + MOVSX(32, size, rdMapped.GetSimpleReg(), R(RSCRATCH)); + else + MOVZX(32, size, rdMapped.GetSimpleReg(), R(RSCRATCH)); + } } } diff --git a/src/dolphin/Arm64Emitter.cpp b/src/dolphin/Arm64Emitter.cpp index dd2416b..97c93ba 100644 --- a/src/dolphin/Arm64Emitter.cpp +++ b/src/dolphin/Arm64Emitter.cpp @@ -340,6 +340,11 @@ void* ARM64XEmitter::GetRXPtr() return m_rxbase + m_code; } +u8* ARM64XEmitter::GetRXBase() +{ + return m_rxbase; +} + void ARM64XEmitter::ReserveCodeSpace(u32 bytes) { for (u32 i = 0; i < bytes / 4; i++) diff --git a/src/dolphin/Arm64Emitter.h b/src/dolphin/Arm64Emitter.h index 3d9d4ba..3efc838 100644 --- a/src/dolphin/Arm64Emitter.h +++ b/src/dolphin/Arm64Emitter.h @@ -556,6 +556,7 @@ public: const u8* GetRWPtr(); u8* GetWriteableRWPtr(); void* GetRXPtr(); + u8* GetRXBase(); void FlushIcache(); void FlushIcacheSection(u8* start, u8* end); |