ShuriZma
/
melonDS
mirror of https://github.com/melonDS-emu/melonDS.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Make it buildable on aarch64

This commit is contained in:
RSDuck 2020-07-23 19:07:33 +00:00
parent e63bd7e38c
commit 961b4252e2
5 changed files with 129 additions and 71 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

62
src/ARMJIT_A64/ARMJIT_Compiler.cpp
View File

@ -312,6 +312,8 @@ Compiler::Compiler()
RET();
}
for (int consoleType = 0; consoleType < 2; consoleType++)
{
for (int num = 0; num < 2; num++)
{
for (int size = 0; size < 3; size++)
@ -319,7 +321,7 @@ Compiler::Compiler()
for (int reg = 0; reg < 8; reg++)
{
ARM64Reg rdMapped = (ARM64Reg)(W19 + reg);
PatchedStoreFuncs[num][size][reg] = GetRXPtr();
PatchedStoreFuncs[consoleType][num][size][reg] = GetRXPtr();
if (num == 0)
{
MOV(X1, RCPU);
@ -330,32 +332,63 @@ Compiler::Compiler()
MOV(W1, rdMapped);
}
ABI_PushRegisters({30});
if (consoleType == 0)
{
switch ((8 << size) | num)
{
case 32: QuickCallFunction(X3, SlowWrite9<u32>); break;
case 33: QuickCallFunction(X3, SlowWrite7<u32>); break;
case 16: QuickCallFunction(X3, SlowWrite9<u16>); break;
case 17: QuickCallFunction(X3, SlowWrite7<u16>); break;
case 8: QuickCallFunction(X3, SlowWrite9<u8>); break;
case 9: QuickCallFunction(X3, SlowWrite7<u8>); break;
case 32: QuickCallFunction(X3, SlowWrite9<u32, 0>); break;
case 33: QuickCallFunction(X3, SlowWrite7<u32, 0>); break;
case 16: QuickCallFunction(X3, SlowWrite9<u16, 0>); break;
case 17: QuickCallFunction(X3, SlowWrite7<u16, 0>); break;
case 8: QuickCallFunction(X3, SlowWrite9<u8, 0>); break;
case 9: QuickCallFunction(X3, SlowWrite7<u8, 0>); break;
}
}
else
{
switch ((8 << size) | num)
{
case 32: QuickCallFunction(X3, SlowWrite9<u32, 1>); break;
case 33: QuickCallFunction(X3, SlowWrite7<u32, 1>); break;
case 16: QuickCallFunction(X3, SlowWrite9<u16, 1>); break;
case 17: QuickCallFunction(X3, SlowWrite7<u16, 1>); break;
case 8: QuickCallFunction(X3, SlowWrite9<u8, 1>); break;
case 9: QuickCallFunction(X3, SlowWrite7<u8, 1>); break;
}
}
ABI_PopRegisters({30});
RET();
for (int signextend = 0; signextend < 2; signextend++)
{
PatchedLoadFuncs[num][size][signextend][reg] = GetRXPtr();
PatchedLoadFuncs[consoleType][num][size][signextend][reg] = GetRXPtr();
if (num == 0)
MOV(X1, RCPU);
ABI_PushRegisters({30});
if (consoleType == 0)
{
switch ((8 << size) | num)
{
case 32: QuickCallFunction(X3, SlowRead9<u32>); break;
case 33: QuickCallFunction(X3, SlowRead7<u32>); break;
case 16: QuickCallFunction(X3, SlowRead9<u16>); break;
case 17: QuickCallFunction(X3, SlowRead7<u16>); break;
case 8: QuickCallFunction(X3, SlowRead9<u8>); break;
case 9: QuickCallFunction(X3, SlowRead7<u8>); break;
case 32: QuickCallFunction(X3, SlowRead9<u32, 0>); break;
case 33: QuickCallFunction(X3, SlowRead7<u32, 0>); break;
case 16: QuickCallFunction(X3, SlowRead9<u16, 0>); break;
case 17: QuickCallFunction(X3, SlowRead7<u16, 0>); break;
case 8: QuickCallFunction(X3, SlowRead9<u8, 0>); break;
case 9: QuickCallFunction(X3, SlowRead7<u8, 0>); break;
}
}
else
{
switch ((8 << size) | num)
{
case 32: QuickCallFunction(X3, SlowRead9<u32, 1>); break;
case 33: QuickCallFunction(X3, SlowRead7<u32, 1>); break;
case 16: QuickCallFunction(X3, SlowRead9<u16, 1>); break;
case 17: QuickCallFunction(X3, SlowRead7<u16, 1>); break;
case 8: QuickCallFunction(X3, SlowRead9<u8, 1>); break;
case 9: QuickCallFunction(X3, SlowRead7<u8, 1>); break;
}
}
ABI_PopRegisters({30});
if (size == 32)
@ -369,6 +402,7 @@ Compiler::Compiler()
}
}
}
}
FlushIcache();

6
src/ARMJIT_A64/ARMJIT_Compiler.h
View File

@ -247,9 +247,9 @@ public:
std::unordered_map<ptrdiff_t, LoadStorePatch> LoadStorePatches;
// [Num][Size][Sign Extend][Output register]
void* PatchedLoadFuncs[2][3][2][8];
void* PatchedStoreFuncs[2][3][8];
// [Console Type][Num][Size][Sign Extend][Output register]
void* PatchedLoadFuncs[2][2][3][2][8];
void* PatchedStoreFuncs[2][2][3][8];
RegisterCache<Compiler, Arm64Gen::ARM64Reg> RegCache;

6
src/ARMJIT_A64/ARMJIT_Linkage.s
View File

@ -2,9 +2,9 @@
.text
#define RCPSR W27
#define RCycles W28
#define RCPU X29
#define RCPSR w27
#define RCycles w28
#define RCPU x29
.p2align 4,,15

63
src/ARMJIT_A64/ARMJIT_LoadStore.cpp
View File

@ -174,8 +174,8 @@ void Compiler::Comp_MemAccess(int rd, int rn, Op2 offset, int size, int flags)
LoadStorePatch patch;
patch.PatchFunc = flags & memop_Store
? PatchedStoreFuncs[Num][__builtin_ctz(size) - 3][rdMapped - W19]
: PatchedLoadFuncs[Num][__builtin_ctz(size) - 3][!!(flags & memop_SignExtend)][rdMapped - W19];
? PatchedStoreFuncs[NDS::ConsoleType][Num][__builtin_ctz(size) - 3][rdMapped - W19]
: PatchedLoadFuncs[NDS::ConsoleType][Num][__builtin_ctz(size) - 3][!!(flags & memop_SignExtend)][rdMapped - W19];
assert(rdMapped - W19 >= 0 && rdMapped - W19 < 8);
MOVP2R(X7, Num == 0 ? ARMJIT_Memory::FastMem9Start : ARMJIT_Memory::FastMem7Start);
@ -241,20 +241,26 @@ void Compiler::Comp_MemAccess(int rd, int rn, Op2 offset, int size, int flags)
if (flags & memop_Store)
{
MOV(W2, rdMapped);
switch (size)
switch (size | NDS::ConsoleType)
{
case 32: QuickCallFunction(X3, SlowWrite9<u32>); break;
case 16: QuickCallFunction(X3, SlowWrite9<u16>); break;
case 8: QuickCallFunction(X3, SlowWrite9<u8>); break;
case 32: QuickCallFunction(X3, SlowWrite9<u32, 0>); break;
case 33: QuickCallFunction(X3, SlowWrite9<u32, 1>); break;
case 16: QuickCallFunction(X3, SlowWrite9<u16, 0>); break;
case 17: QuickCallFunction(X3, SlowWrite9<u16, 1>); break;
case 8: QuickCallFunction(X3, SlowWrite9<u8, 0>); break;
case 9: QuickCallFunction(X3, SlowWrite9<u8, 1>); break;
}
}
else
{
switch (size)
{
case 32: QuickCallFunction(X3, SlowRead9<u32>); break;
case 16: QuickCallFunction(X3, SlowRead9<u16>); break;
case 8: QuickCallFunction(X3, SlowRead9<u8>); break;
case 32: QuickCallFunction(X3, SlowRead9<u32, 0>); break;
case 33: QuickCallFunction(X3, SlowRead9<u32, 1>); break;
case 16: QuickCallFunction(X3, SlowRead9<u16, 0>); break;
case 17: QuickCallFunction(X3, SlowRead9<u16, 0>); break;
case 8: QuickCallFunction(X3, SlowRead9<u8, 0>); break;
case 9: QuickCallFunction(X3, SlowRead9<u8, 1>); break;
}
}
}
@ -265,18 +271,24 @@ void Compiler::Comp_MemAccess(int rd, int rn, Op2 offset, int size, int flags)
MOV(W1, rdMapped);
switch (size)
{
case 32: QuickCallFunction(X3, SlowWrite7<u32>); break;
case 16: QuickCallFunction(X3, SlowWrite7<u16>); break;
case 8: QuickCallFunction(X3, SlowWrite7<u8>); break;
case 32: QuickCallFunction(X3, SlowWrite7<u32, 0>); break;
case 33: QuickCallFunction(X3, SlowWrite7<u32, 1>); break;
case 16: QuickCallFunction(X3, SlowWrite7<u16, 0>); break;
case 17: QuickCallFunction(X3, SlowWrite7<u16, 1>); break;
case 8: QuickCallFunction(X3, SlowWrite7<u8, 0>); break;
case 9: QuickCallFunction(X3, SlowWrite7<u8, 1>); break;
}
}
else
{
switch (size)
{
case 32: QuickCallFunction(X3, SlowRead7<u32>); break;
case 16: QuickCallFunction(X3, SlowRead7<u16>); break;
case 8: QuickCallFunction(X3, SlowRead7<u8>); break;
case 32: QuickCallFunction(X3, SlowRead7<u32, 0>); break;
case 33: QuickCallFunction(X3, SlowRead7<u32, 1>); break;
case 16: QuickCallFunction(X3, SlowRead7<u16, 0>); break;
case 17: QuickCallFunction(X3, SlowRead7<u16, 1>); break;
case 8: QuickCallFunction(X3, SlowRead7<u8, 0>); break;
case 9: QuickCallFunction(X3, SlowRead7<u8, 1>); break;
}
}
}
@ -465,15 +477,25 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
if (decrement)
{
SUB(W0, MapReg(rn), regsCount * 4);
s32 offset = -regsCount * 4 + (preinc ? 0 : 4);
if (offset)
{
ADDI2R(W0, MapReg(rn), offset);
ANDI2R(W0, W0, ~3);
preinc ^= true;
}
else
{
ANDI2R(W0, MapReg(rn), ~3);
}
}
else
{
ANDI2R(W0, MapReg(rn), ~3);
if (preinc)
ADD(W0, W0, 4);
}
u8* patchFunc;
if (compileFastPath)
{
ptrdiff_t fastPathStart = GetCodeOffset();
@ -482,7 +504,7 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
MOVP2R(X1, Num == 0 ? ARMJIT_Memory::FastMem9Start : ARMJIT_Memory::FastMem7Start);
ADD(X1, X1, X0);
u32 offset = preinc ? 4 : 0;
u32 offset = 0;
BitSet16::Iterator it = regs.begin();
u32 i = 0;
@ -545,7 +567,8 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
LoadStorePatch patch;
patch.PatchSize = GetCodeOffset() - fastPathStart;
SwapCodeRegion();
patch.PatchFunc = GetRXPtr();
patchFunc = (u8*)GetRXPtr();
patch.PatchFunc = patchFunc;
for (i = 0; i < regsCount; i++)
{
patch.PatchOffset = fastPathStart - loadStoreOffsets[i];
@ -705,7 +728,7 @@ s32 Compiler::Comp_MemAccessBlock(int rn, BitSet16 regs, bool store, bool preinc
ABI_PopRegisters({30});
RET();
FlushIcacheSection((u8*)patch.PatchFunc, (u8*)GetRXPtr());
FlushIcacheSection(patchFunc, (u8*)GetRXPtr());
SwapCodeRegion();
}

1
src/CMakeLists.txt
View File

@ -90,6 +90,7 @@ if (ENABLE_JIT)
ARMJIT_A64/ARMJIT_Linkage.s
)
set_source_files_properties(ARMJIT_A64/ARMJIT_Linkage.s PROPERTIES COMPILE_FLAGS "-x assembler-with-cpp")
endif()
endif()