zdoom/gzdoom
0
0
Fork 0
mirror of https://github.com/ZDoom/gzdoom.git synced 2024-11-27 14:22:13 +00:00
Code Issues Projects Releases Packages Wiki Activity

- move the jit runtime to its own file

Browse source
This commit is contained in:
Magnus Norddahl 2018-11-26 10:46:09 +01:00
parent 90d79865a5
commit 8b852ce4d8
4 changed files with 302 additions and 296 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
src/CMakeLists.txt
View file

@ -1166,6 +1166,7 @@ set (PCH_SOURCES
scripting/vm/vmexec.cpp
scripting/vm/vmframe.cpp
scripting/vm/jit.cpp
scripting/vm/jit_runtime.cpp
scripting/vm/jit_call.cpp
scripting/vm/jit_flow.cpp
scripting/vm/jit_load.cpp

296
src/scripting/vm/jit.cpp
View file

@ -8,302 +8,6 @@ extern PStruct *TypeVector3;
static void OutputJitLog(const asmjit::StringLogger &logger);
static TArray<uint8_t*> JitBlocks;
static TArray<uint8_t*> JitFrames;
static size_t JitBlockPos = 0;
static size_t JitBlockSize = 0;
static asmjit::CodeInfo GetHostCodeInfo()
{
static bool firstCall = true;
static asmjit::CodeInfo codeInfo;
if (firstCall)
{
asmjit::JitRuntime rt;
codeInfo = rt.getCodeInfo();
firstCall = false;
}
return codeInfo;
}
static void *AllocJitMemory(size_t size)
{
using namespace asmjit;
if (JitBlockPos + size <= JitBlockSize)
{
uint8_t *p = JitBlocks[JitBlocks.Size() - 1];
p += JitBlockPos;
JitBlockPos += size;
return p;
}
else
{
size_t allocatedSize = 0;
void *p = OSUtils::allocVirtualMemory(1024 * 1024, &allocatedSize, OSUtils::kVMWritable | OSUtils::kVMExecutable);
if (!p)
return nullptr;
JitBlocks.Push((uint8_t*)p);
JitBlockSize = allocatedSize;
JitBlockPos = size;
return p;
}
}
#define UWOP_PUSH_NONVOL 0
#define UWOP_ALLOC_LARGE 1
#define UWOP_ALLOC_SMALL 2
#define UWOP_SET_FPREG 3
#define UWOP_SAVE_NONVOL 4
#define UWOP_SAVE_NONVOL_FAR 5
#define UWOP_SAVE_XMM128 8
#define UWOP_SAVE_XMM128_FAR 9
#define UWOP_PUSH_MACHFRAME 10
void JitRelease()
{
#ifdef _WIN64
for (auto p : JitFrames)
{
RtlDeleteFunctionTable((PRUNTIME_FUNCTION)p);
}
#endif
for (auto p : JitBlocks)
{
asmjit::OSUtils::releaseVirtualMemory(p, 1024 * 1024);
}
JitFrames.Clear();
JitBlocks.Clear();
JitBlockPos = 0;
JitBlockSize = 0;
}
static TArray<uint16_t> CreateUnwindInfo(asmjit::CCFunc *func)
{
using namespace asmjit;
FuncFrameLayout layout;
Error error = layout.init(func->getDetail(), func->getFrameInfo());
if (error != kErrorOk)
I_FatalError("FuncFrameLayout.init failed");
// We need a dummy emitter for instruction size calculations
CodeHolder code;
code.init(GetHostCodeInfo());
X86Assembler assembler(&code);
X86Emitter *emitter = assembler.asEmitter();
// Build UNWIND_CODE codes:
TArray<uint16_t> codes;
uint32_t opoffset, opcode, opinfo;
// Note: this must match exactly what X86Internal::emitProlog does
X86Gp zsp = emitter->zsp(); // ESP|RSP register.
X86Gp zbp = emitter->zsp(); // EBP|RBP register.
zbp.setId(X86Gp::kIdBp);
X86Gp gpReg = emitter->zsp(); // General purpose register (temporary).
X86Gp saReg = emitter->zsp(); // Stack-arguments base register.
uint32_t gpSaved = layout.getSavedRegs(X86Reg::kKindGp);
if (layout.hasPreservedFP())
{
// Emit: 'push zbp'
// 'mov zbp, zsp'.
gpSaved &= ~Utils::mask(X86Gp::kIdBp);
emitter->push(zbp);
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_PUSH_NONVOL;
opinfo = X86Gp::kIdBp;
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
emitter->mov(zbp, zsp);
}
if (gpSaved)
{
for (uint32_t i = gpSaved, regId = 0; i; i >>= 1, regId++)
{
if (!(i & 0x1)) continue;
// Emit: 'push gp' sequence.
gpReg.setId(regId);
emitter->push(gpReg);
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_PUSH_NONVOL;
opinfo = regId;
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
}
uint32_t stackArgsRegId = layout.getStackArgsRegId();
if (stackArgsRegId != Globals::kInvalidRegId && stackArgsRegId != X86Gp::kIdSp)
{
saReg.setId(stackArgsRegId);
if (!(layout.hasPreservedFP() && stackArgsRegId == X86Gp::kIdBp))
{
// Emit: 'mov saReg, zsp'.
emitter->mov(saReg, zsp);
}
}
if (layout.hasDynamicAlignment())
{
// Emit: 'and zsp, StackAlignment'.
emitter->and_(zsp, -static_cast<int32_t>(layout.getStackAlignment()));
}
if (layout.hasStackAdjustment())
{
// Emit: 'sub zsp, StackAdjustment'.
emitter->sub(zsp, layout.getStackAdjustment());
uint32_t stackadjust = layout.getStackAdjustment();
if (stackadjust <= 128)
{
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_ALLOC_SMALL;
opinfo = stackadjust / 8 - 1;
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
else if (stackadjust <= 512 * 1024 - 8)
{
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_ALLOC_LARGE;
opinfo = 0;
codes.Push(stackadjust / 8);
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
else
{
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_ALLOC_LARGE;
opinfo = 1;
codes.Push((uint16_t)(stackadjust >> 16));
codes.Push((uint16_t)stackadjust);
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
}
if (layout.hasDynamicAlignment() && layout.hasDsaSlotUsed())
{
// Emit: 'mov [zsp + dsaSlot], saReg'.
X86Mem saMem = x86::ptr(zsp, layout._dsaSlot);
emitter->mov(saMem, saReg);
}
uint32_t xmmSaved = layout.getSavedRegs(X86Reg::kKindVec);
if (xmmSaved)
{
X86Mem vecBase = x86::ptr(zsp, layout.getVecStackOffset());
X86Reg vecReg = x86::xmm(0);
bool avx = layout.isAvxEnabled();
bool aligned = layout.hasAlignedVecSR();
uint32_t vecInst = aligned ? (avx ? X86Inst::kIdVmovaps : X86Inst::kIdMovaps) : (avx ? X86Inst::kIdVmovups : X86Inst::kIdMovups);
uint32_t vecSize = 16;
for (uint32_t i = xmmSaved, regId = 0; i; i >>= 1, regId++)
{
if (!(i & 0x1)) continue;
// Emit 'movaps|movups [zsp + X], xmm0..15'.
vecReg.setId(regId);
emitter->emit(vecInst, vecBase, vecReg);
vecBase.addOffsetLo32(static_cast<int32_t>(vecSize));
if (vecBase.getOffsetLo32() / vecSize < (1 << 16))
{
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_SAVE_XMM128;
opinfo = regId;
codes.Push(vecBase.getOffsetLo32() / vecSize);
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
else
{
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_SAVE_XMM128_FAR;
opinfo = regId;
codes.Push((uint16_t)(vecBase.getOffsetLo32() >> 16));
codes.Push((uint16_t)vecBase.getOffsetLo32());
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
}
}
// Build the UNWIND_INFO structure:
uint16_t version = 1, flags = 0, frameRegister = 0, frameOffset = 0;
uint16_t sizeOfProlog = (uint16_t)assembler.getOffset();
uint16_t countOfCodes = (uint16_t)codes.Size();
TArray<uint16_t> info;
info.Push(version | (flags << 3) | (sizeOfProlog << 8));
info.Push(countOfCodes | (frameRegister << 8) | (frameOffset << 12));
for (unsigned int i = codes.Size(); i > 0; i--)
info.Push(codes[i - 1]);
if (codes.Size() % 2 == 1)
info.Push(0);
return info;
}
static void *AddJitFunction(asmjit::CodeHolder* code, asmjit::CCFunc *func)
{
using namespace asmjit;
size_t codeSize = code->getCodeSize();
if (codeSize == 0)
return nullptr;
#ifdef _WIN64
TArray<uint16_t> unwindInfo = CreateUnwindInfo(func);
size_t unwindInfoSize = unwindInfo.Size() * sizeof(uint16_t);
size_t functionTableSize = sizeof(RUNTIME_FUNCTION);
#else
size_t unwindInfoSize = 0;
size_t functionTableSize = 0;
#endif
codeSize = (codeSize + 15) / 16 * 16;
uint8_t *p = (uint8_t *)AllocJitMemory(codeSize + unwindInfoSize + functionTableSize);
if (!p)
return nullptr;
size_t relocSize = code->relocate(p);
if (relocSize == 0)
return nullptr;
size_t unwindStart = relocSize;
unwindStart = (unwindStart + 15) / 16 * 16;
JitBlockPos -= codeSize - unwindStart;
#ifdef _WIN64
uint8_t *baseaddr = JitBlocks.Last();
uint8_t *startaddr = p;
uint8_t *endaddr = p + relocSize;
uint8_t *unwindptr = p + unwindStart;
memcpy(unwindptr, &unwindInfo[0], unwindInfoSize);
RUNTIME_FUNCTION *table = (RUNTIME_FUNCTION*)(unwindptr + unwindInfoSize);
table[0].BeginAddress = (DWORD)(ptrdiff_t)(startaddr - baseaddr);
table[0].EndAddress = (DWORD)(ptrdiff_t)(endaddr - baseaddr);
table[0].UnwindInfoAddress = (DWORD)(ptrdiff_t)(unwindptr - baseaddr);
BOOLEAN result = RtlAddFunctionTable(table, 1, (DWORD64)baseaddr);
JitFrames.Push((uint8_t*)table);
if (result == 0)
I_FatalError("RtlAddFunctionTable failed");
#endif
return p;
}
JitFuncPtr JitCompile(VMScriptFunction *sfunc)
{
#if 0

298
src/scripting/vm/jit_runtime.cpp Normal file
View file

@ -0,0 +1,298 @@
#include "jit.h"
#include "jitintern.h"
static TArray<uint8_t*> JitBlocks;
static TArray<uint8_t*> JitFrames;
static size_t JitBlockPos = 0;
static size_t JitBlockSize = 0;
asmjit::CodeInfo GetHostCodeInfo()
{
static bool firstCall = true;
static asmjit::CodeInfo codeInfo;
if (firstCall)
{
asmjit::JitRuntime rt;
codeInfo = rt.getCodeInfo();
firstCall = false;
}
return codeInfo;
}
static void *AllocJitMemory(size_t size)
{
using namespace asmjit;
if (JitBlockPos + size <= JitBlockSize)
{
uint8_t *p = JitBlocks[JitBlocks.Size() - 1];
p += JitBlockPos;
JitBlockPos += size;
return p;
}
else
{
size_t allocatedSize = 0;
void *p = OSUtils::allocVirtualMemory(1024 * 1024, &allocatedSize, OSUtils::kVMWritable | OSUtils::kVMExecutable);
if (!p)
return nullptr;
JitBlocks.Push((uint8_t*)p);
JitBlockSize = allocatedSize;
JitBlockPos = size;
return p;
}
}
#define UWOP_PUSH_NONVOL 0
#define UWOP_ALLOC_LARGE 1
#define UWOP_ALLOC_SMALL 2
#define UWOP_SET_FPREG 3
#define UWOP_SAVE_NONVOL 4
#define UWOP_SAVE_NONVOL_FAR 5
#define UWOP_SAVE_XMM128 8
#define UWOP_SAVE_XMM128_FAR 9
#define UWOP_PUSH_MACHFRAME 10
void JitRelease()
{
#ifdef _WIN64
for (auto p : JitFrames)
{
RtlDeleteFunctionTable((PRUNTIME_FUNCTION)p);
}
#endif
for (auto p : JitBlocks)
{
asmjit::OSUtils::releaseVirtualMemory(p, 1024 * 1024);
}
JitFrames.Clear();
JitBlocks.Clear();
JitBlockPos = 0;
JitBlockSize = 0;
}
static TArray<uint16_t> CreateUnwindInfo(asmjit::CCFunc *func)
{
using namespace asmjit;
FuncFrameLayout layout;
Error error = layout.init(func->getDetail(), func->getFrameInfo());
if (error != kErrorOk)
I_FatalError("FuncFrameLayout.init failed");
// We need a dummy emitter for instruction size calculations
CodeHolder code;
code.init(GetHostCodeInfo());
X86Assembler assembler(&code);
X86Emitter *emitter = assembler.asEmitter();
// Build UNWIND_CODE codes:
TArray<uint16_t> codes;
uint32_t opoffset, opcode, opinfo;
// Note: this must match exactly what X86Internal::emitProlog does
X86Gp zsp = emitter->zsp(); // ESP|RSP register.
X86Gp zbp = emitter->zsp(); // EBP|RBP register.
zbp.setId(X86Gp::kIdBp);
X86Gp gpReg = emitter->zsp(); // General purpose register (temporary).
X86Gp saReg = emitter->zsp(); // Stack-arguments base register.
uint32_t gpSaved = layout.getSavedRegs(X86Reg::kKindGp);
if (layout.hasPreservedFP())
{
// Emit: 'push zbp'
// 'mov zbp, zsp'.
gpSaved &= ~Utils::mask(X86Gp::kIdBp);
emitter->push(zbp);
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_PUSH_NONVOL;
opinfo = X86Gp::kIdBp;
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
emitter->mov(zbp, zsp);
}
if (gpSaved)
{
for (uint32_t i = gpSaved, regId = 0; i; i >>= 1, regId++)
{
if (!(i & 0x1)) continue;
// Emit: 'push gp' sequence.
gpReg.setId(regId);
emitter->push(gpReg);
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_PUSH_NONVOL;
opinfo = regId;
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
}
uint32_t stackArgsRegId = layout.getStackArgsRegId();
if (stackArgsRegId != Globals::kInvalidRegId && stackArgsRegId != X86Gp::kIdSp)
{
saReg.setId(stackArgsRegId);
if (!(layout.hasPreservedFP() && stackArgsRegId == X86Gp::kIdBp))
{
// Emit: 'mov saReg, zsp'.
emitter->mov(saReg, zsp);
}
}
if (layout.hasDynamicAlignment())
{
// Emit: 'and zsp, StackAlignment'.
emitter->and_(zsp, -static_cast<int32_t>(layout.getStackAlignment()));
}
if (layout.hasStackAdjustment())
{
// Emit: 'sub zsp, StackAdjustment'.
emitter->sub(zsp, layout.getStackAdjustment());
uint32_t stackadjust = layout.getStackAdjustment();
if (stackadjust <= 128)
{
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_ALLOC_SMALL;
opinfo = stackadjust / 8 - 1;
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
else if (stackadjust <= 512 * 1024 - 8)
{
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_ALLOC_LARGE;
opinfo = 0;
codes.Push(stackadjust / 8);
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
else
{
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_ALLOC_LARGE;
opinfo = 1;
codes.Push((uint16_t)(stackadjust >> 16));
codes.Push((uint16_t)stackadjust);
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
}
if (layout.hasDynamicAlignment() && layout.hasDsaSlotUsed())
{
// Emit: 'mov [zsp + dsaSlot], saReg'.
X86Mem saMem = x86::ptr(zsp, layout._dsaSlot);
emitter->mov(saMem, saReg);
}
uint32_t xmmSaved = layout.getSavedRegs(X86Reg::kKindVec);
if (xmmSaved)
{
X86Mem vecBase = x86::ptr(zsp, layout.getVecStackOffset());
X86Reg vecReg = x86::xmm(0);
bool avx = layout.isAvxEnabled();
bool aligned = layout.hasAlignedVecSR();
uint32_t vecInst = aligned ? (avx ? X86Inst::kIdVmovaps : X86Inst::kIdMovaps) : (avx ? X86Inst::kIdVmovups : X86Inst::kIdMovups);
uint32_t vecSize = 16;
for (uint32_t i = xmmSaved, regId = 0; i; i >>= 1, regId++)
{
if (!(i & 0x1)) continue;
// Emit 'movaps|movups [zsp + X], xmm0..15'.
vecReg.setId(regId);
emitter->emit(vecInst, vecBase, vecReg);
vecBase.addOffsetLo32(static_cast<int32_t>(vecSize));
if (vecBase.getOffsetLo32() / vecSize < (1 << 16))
{
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_SAVE_XMM128;
opinfo = regId;
codes.Push(vecBase.getOffsetLo32() / vecSize);
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
else
{
opoffset = (uint32_t)assembler.getOffset();
opcode = UWOP_SAVE_XMM128_FAR;
opinfo = regId;
codes.Push((uint16_t)(vecBase.getOffsetLo32() >> 16));
codes.Push((uint16_t)vecBase.getOffsetLo32());
codes.Push(opoffset | (opcode << 8) | (opinfo << 12));
}
}
}
// Build the UNWIND_INFO structure:
uint16_t version = 1, flags = 0, frameRegister = 0, frameOffset = 0;
uint16_t sizeOfProlog = (uint16_t)assembler.getOffset();
uint16_t countOfCodes = (uint16_t)codes.Size();
TArray<uint16_t> info;
info.Push(version | (flags << 3) | (sizeOfProlog << 8));
info.Push(countOfCodes | (frameRegister << 8) | (frameOffset << 12));
for (unsigned int i = codes.Size(); i > 0; i--)
info.Push(codes[i - 1]);
if (codes.Size() % 2 == 1)
info.Push(0);
return info;
}
void *AddJitFunction(asmjit::CodeHolder* code, asmjit::CCFunc *func)
{
using namespace asmjit;
size_t codeSize = code->getCodeSize();
if (codeSize == 0)
return nullptr;
#ifdef _WIN64
TArray<uint16_t> unwindInfo = CreateUnwindInfo(func);
size_t unwindInfoSize = unwindInfo.Size() * sizeof(uint16_t);
size_t functionTableSize = sizeof(RUNTIME_FUNCTION);
#else
size_t unwindInfoSize = 0;
size_t functionTableSize = 0;
#endif
codeSize = (codeSize + 15) / 16 * 16;
uint8_t *p = (uint8_t *)AllocJitMemory(codeSize + unwindInfoSize + functionTableSize);
if (!p)
return nullptr;
size_t relocSize = code->relocate(p);
if (relocSize == 0)
return nullptr;
size_t unwindStart = relocSize;
unwindStart = (unwindStart + 15) / 16 * 16;
JitBlockPos -= codeSize - unwindStart;
#ifdef _WIN64
uint8_t *baseaddr = JitBlocks.Last();
uint8_t *startaddr = p;
uint8_t *endaddr = p + relocSize;
uint8_t *unwindptr = p + unwindStart;
memcpy(unwindptr, &unwindInfo[0], unwindInfoSize);
RUNTIME_FUNCTION *table = (RUNTIME_FUNCTION*)(unwindptr + unwindInfoSize);
table[0].BeginAddress = (DWORD)(ptrdiff_t)(startaddr - baseaddr);
table[0].EndAddress = (DWORD)(ptrdiff_t)(endaddr - baseaddr);
table[0].UnwindInfoAddress = (DWORD)(ptrdiff_t)(unwindptr - baseaddr);
BOOLEAN result = RtlAddFunctionTable(table, 1, (DWORD64)baseaddr);
JitFrames.Push((uint8_t*)table);
if (result == 0)
I_FatalError("RtlAddFunctionTable failed");
#endif
return p;
}

3
src/scripting/vm/jitintern.h
View file

@ -307,3 +307,6 @@ public:
throw AsmJitException(err, message);
}
};
void *AddJitFunction(asmjit::CodeHolder* code, asmjit::CCFunc *func);
asmjit::CodeInfo GetHostCodeInfo();