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

fastmem 16kb pages support

This commit is contained in:
RSDuck 2024-11-18 18:33:07 +01:00
parent bbf9290b5f
commit 2fdc0ff1c6
5 changed files with 60 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/ARMJIT.cpp
View File

@ -916,7 +916,7 @@ void ARMJIT::CompileBlock(ARM* cpu) noexcept
AddressRange* region = CodeMemRegions[addressRanges[j] >> 27];
if (!PageContainsCode(&region[(addressRanges[j] & 0x7FFF000) / 512]))
if (!PageContainsCode(&region[(addressRanges[j] & 0x7FFF000 & ~(Memory.PageSize - 1)) / 512], Memory.PageSize))
Memory.SetCodeProtection(addressRanges[j] >> 27, addressRanges[j] & 0x7FFFFFF, true);
AddressRange* range = &region[(addressRanges[j] & 0x7FFFFFF) / 512];
@ -969,7 +969,7 @@ void ARMJIT::InvalidateByAddr(u32 localAddr) noexcept
range->Blocks.Remove(i);
if (range->Blocks.Length == 0
&& !PageContainsCode(&region[(localAddr & 0x7FFF000) / 512]))
&& !PageContainsCode(&region[(localAddr & 0x7FFF000 & ~(Memory.PageSize - 1)) / 512], Memory.PageSize))
{
Memory.SetCodeProtection(localAddr >> 27, localAddr & 0x7FFFFFF, false);
}
@ -1003,7 +1003,7 @@ void ARMJIT::InvalidateByAddr(u32 localAddr) noexcept
if (otherRange->Blocks.Length == 0)
{
if (!PageContainsCode(&otherRegion[(addr & 0x7FFF000) / 512]))
if (!PageContainsCode(&otherRegion[(addr & 0x7FFF000 & ~(Memory.PageSize - 1)) / 512], Memory.PageSize))
Memory.SetCodeProtection(addr >> 27, addr & 0x7FFFFFF, false);
otherRange->Code = 0;

1
src/ARMJIT.h
View File

@ -72,6 +72,7 @@ private:
bool LiteralOptimizations = false;
bool BranchOptimizations = false;
bool FastMemory = false;
public:
melonDS::NDS& NDS;
TinyVector<u32> InvalidLiterals {};

4
src/ARMJIT_Internal.h
View File

@ -85,9 +85,9 @@ typedef void (*InterpreterFunc)(ARM* cpu);
extern InterpreterFunc InterpretARM[];
extern InterpreterFunc InterpretTHUMB[];
inline bool PageContainsCode(const AddressRange* range)
inline bool PageContainsCode(const AddressRange* range, u32 pageSize)
{
for (int i = 0; i < 8; i++)
for (int i = 0; i < pageSize / 512; i++)
{
if (range[i].Blocks.Length > 0)
return true;

63
src/ARMJIT_Memory.cpp
View File

@ -275,10 +275,14 @@ enum
memstate_MappedProtected,
};
#define CHECK_ALIGNED(value) assert(((value) & (PageSize-1)) == 0)
bool ARMJIT_Memory::MapIntoRange(u32 addr, u32 num, u32 offset, u32 size) noexcept
{
CHECK_ALIGNED(addr);
CHECK_ALIGNED(offset);
CHECK_ALIGNED(size);
u8* dst = (u8*)(num == 0 ? FastMem9Start : FastMem7Start) + addr;
#ifdef __SWITCH__
Result r = (svcMapProcessMemory(dst, envGetOwnProcessHandle(),
@ -333,6 +337,10 @@ bool ARMJIT_Memory::MapIntoRange(u32 addr, u32 num, u32 offset, u32 size) noexce
bool ARMJIT_Memory::UnmapFromRange(u32 addr, u32 num, u32 offset, u32 size) noexcept
{
CHECK_ALIGNED(addr);
CHECK_ALIGNED(offset);
CHECK_ALIGNED(size);
u8* dst = (u8*)(num == 0 ? FastMem9Start : FastMem7Start) + addr;
#ifdef __SWITCH__
Result r = svcUnmapProcessMemory(dst, envGetOwnProcessHandle(),
@ -388,6 +396,9 @@ bool ARMJIT_Memory::UnmapFromRange(u32 addr, u32 num, u32 offset, u32 size) noex
#ifndef __SWITCH__
void ARMJIT_Memory::SetCodeProtectionRange(u32 addr, u32 size, u32 num, int protection) noexcept
{
CHECK_ALIGNED(addr);
CHECK_ALIGNED(size);
u8* dst = (u8*)(num == 0 ? FastMem9Start : FastMem7Start) + addr;
#if defined(_WIN32)
DWORD winProtection, oldProtection;
@ -432,14 +443,14 @@ void ARMJIT_Memory::Mapping::Unmap(int region, melonDS::NDS& nds) noexcept
else
{
u32 segmentOffset = offset;
u8 status = statuses[(Addr + offset) >> 12];
while (statuses[(Addr + offset) >> 12] == status
u8 status = statuses[(Addr + offset) >> PageShift];
while (statuses[(Addr + offset) >> PageShift] == status
&& offset < Size
&& (!skipDTCM || Addr + offset != dtcmStart))
{
assert(statuses[(Addr + offset) >> 12] != memstate_Unmapped);
statuses[(Addr + offset) >> 12] = memstate_Unmapped;
offset += 0x1000;
assert(statuses[(Addr + offset) >> PageShift] != memstate_Unmapped);
statuses[(Addr + offset) >> PageShift] = memstate_Unmapped;
offset += PageSize;
}
#ifdef __SWITCH__
@ -483,7 +494,7 @@ void ARMJIT_Memory::Mapping::Unmap(int region, melonDS::NDS& nds) noexcept
void ARMJIT_Memory::SetCodeProtection(int region, u32 offset, bool protect) noexcept
{
offset &= ~0xFFF;
offset &= ~(PageSize - 1);
//printf("set code protection %d %x %d\n", region, offset, protect);
for (int i = 0; i < Mappings[region].Length; i++)
@ -501,9 +512,9 @@ void ARMJIT_Memory::SetCodeProtection(int region, u32 offset, bool protect) noex
u8* states = (u8*)(mapping.Num == 0 ? MappingStatus9 : MappingStatus7);
//printf("%x %d %x %x %x %d\n", effectiveAddr, mapping.Num, mapping.Addr, mapping.LocalOffset, mapping.Size, states[effectiveAddr >> 12]);
assert(states[effectiveAddr >> 12] == (protect ? memstate_MappedRW : memstate_MappedProtected));
states[effectiveAddr >> 12] = protect ? memstate_MappedProtected : memstate_MappedRW;
//printf("%x %d %x %x %x %d\n", effectiveAddr, mapping.Num, mapping.Addr, mapping.LocalOffset, mapping.Size, states[effectiveAddr >> PageShift]);
assert(states[effectiveAddr >> PageShift] == (protect ? memstate_MappedRW : memstate_MappedProtected));
states[effectiveAddr >> PageShift] = protect ? memstate_MappedProtected : memstate_MappedRW;
#if defined(__SWITCH__)
bool success;
@ -513,7 +524,7 @@ void ARMJIT_Memory::SetCodeProtection(int region, u32 offset, bool protect) noex
success = MapIntoRange(effectiveAddr, mapping.Num, OffsetsPerRegion[region] + offset, 0x1000);
assert(success);
#else
SetCodeProtectionRange(effectiveAddr, 0x1000, mapping.Num, protect ? 1 : 2);
SetCodeProtectionRange(effectiveAddr, PageSize, mapping.Num, protect ? 1 : 2);
#endif
}
}
@ -642,6 +653,15 @@ bool ARMJIT_Memory::MapAtAddress(u32 addr) noexcept
&& dtcmEnd >= mirrorStart
&& dtcmStart < mirrorStart + mirrorSize)
{
if (dtcmSize < PageSize)
{
// we could technically mask out the DTCM by setting a hole to access permissions
// but realistically there isn't much of a point in mapping less than 16kb of DTCM
// so it isn't worth more complex support
Log(LogLevel::Info, "DTCM size smaller than 16kb skipping mapping entirely");
return false;
}
bool success;
if (dtcmStart > mirrorStart)
{
@ -677,14 +697,14 @@ bool ARMJIT_Memory::MapAtAddress(u32 addr) noexcept
else
{
u32 sectionOffset = offset;
bool hasCode = isExecutable && PageContainsCode(&range[offset / 512]);
bool hasCode = isExecutable && PageContainsCode(&range[offset / 512], PageSize);
while (offset < mirrorSize
&& (!isExecutable || PageContainsCode(&range[offset / 512]) == hasCode)
&& (!isExecutable || PageContainsCode(&range[offset / 512], PageSize) == hasCode)
&& (!skipDTCM || mirrorStart + offset != NDS.ARM9.DTCMBase))
{
assert(states[(mirrorStart + offset) >> 12] == memstate_Unmapped);
states[(mirrorStart + offset) >> 12] = hasCode ? memstate_MappedProtected : memstate_MappedRW;
offset += 0x1000;
assert(states[(mirrorStart + offset) >> PageShift] == memstate_Unmapped);
states[(mirrorStart + offset) >> PageShift] = hasCode ? memstate_MappedProtected : memstate_MappedRW;
offset += PageSize;
}
u32 sectionSize = offset - sectionOffset;
@ -714,6 +734,9 @@ bool ARMJIT_Memory::MapAtAddress(u32 addr) noexcept
return true;
}
u32 ARMJIT_Memory::PageSize = 0;
u32 ARMJIT_Memory::PageShift = 0;
bool ARMJIT_Memory::IsFastMemSupported()
{
#ifdef __APPLE__
@ -727,9 +750,13 @@ bool ARMJIT_Memory::IsFastMemSupported()
ARMJIT_Global::Init();
isSupported = virtualAlloc2Ptr != nullptr;
ARMJIT_Global::DeInit();
PageSize = RegularPageSize;
#else
isSupported = __sysconf(_SC_PAGESIZE) == 0x1000;
PageSize = __sysconf(_SC_PAGESIZE);
isSupported = PageShift == RegularPageSize || PageSize == LargePageSize;
#endif
PageShift = __builtin_ctz(PageSize);
initialised = true;
}
return isSupported;
@ -795,7 +822,7 @@ bool ARMJIT_Memory::FaultHandler(FaultDescription& faultDesc, melonDS::NDS& nds)
u8* memStatus = nds.CurCPU == 0 ? nds.JIT.Memory.MappingStatus9 : nds.JIT.Memory.MappingStatus7;
if (memStatus[faultDesc.EmulatedFaultAddr >> 12] == memstate_Unmapped)
if (memStatus[faultDesc.EmulatedFaultAddr >> PageShift] == memstate_Unmapped)
rewriteToSlowPath = !nds.JIT.Memory.MapAtAddress(faultDesc.EmulatedFaultAddr);
if (rewriteToSlowPath)

10
src/ARMJIT_Memory.h
View File

@ -49,9 +49,12 @@ class Compiler;
class ARMJIT;
#endif
static constexpr u32 LargePageSize = 0x4000;
static constexpr u32 RegularPageSize = 0x1000;
constexpr u32 RoundUp(u32 size) noexcept
{
return size;
return (size + LargePageSize - 1) & ~(LargePageSize - 1);
}
static constexpr u32 MemBlockMainRAMOffset = 0;
@ -139,6 +142,9 @@ public:
static void RegisterFaultHandler();
static void UnregisterFaultHandler();
static u32 PageSize;
static u32 PageShift;
private:
friend class Compiler;
struct Mapping
@ -164,6 +170,7 @@ private:
void* FastMem9Start;
void* FastMem7Start;
u8* MemoryBase = nullptr;
#if defined(__SWITCH__)
VirtmemReservation* FastMem9Reservation, *FastMem7Reservation;
u8* MemoryBaseCodeMem;
@ -178,6 +185,7 @@ private:
static LONG ExceptionHandler(EXCEPTION_POINTERS* exceptionInfo);
HANDLE MemoryFile = INVALID_HANDLE_VALUE;
#else
static void SigsegvHandler(int sig, siginfo_t* info, void* rawContext);
int MemoryFile = -1;
#endif