- avoid calling VMFillParams for simple stack frames

- split Setup into more functions
This commit is contained in:
Magnus Norddahl 2018-11-16 01:13:25 +01:00
parent 6c31d2e965
commit 34b037c9db
4 changed files with 248 additions and 199 deletions

View File

@ -814,6 +814,7 @@ VMFunction *FFunctionBuildList::AddFunction(PNamespace *gnspc, const VersionInfo
if (it.Func->SymbolName != NAME_None)
{
it.Function->Proto = it.Func->Variants[0].Proto;
it.Function->ArgFlags = it.Func->Variants[0].ArgFlags;
}
mItems.Push(it);
@ -885,6 +886,7 @@ void FFunctionBuildList::Build()
if (sfunc->Proto == nullptr)
{
sfunc->Proto = NewPrototype(item.Proto->ReturnTypes, item.Func->Variants[0].Proto->ArgumentTypes);
sfunc->ArgFlags = item.Func->Variants[0].ArgFlags;
}
// Emit code

View File

@ -471,6 +471,245 @@ void JitCompiler::Setup()
konsts = sfunc->KonstS;
konsta = sfunc->KonstA;
labels.Resize(sfunc->CodeSize);
CreateRegisters();
IncrementVMCalls();
SetupFrame();
}
void JitCompiler::SetupFrame()
{
// the VM version reads this from the stack, but it is constant data
offsetParams = ((int)sizeof(VMFrame) + 15) & ~15;
offsetF = offsetParams + (int)(sfunc->MaxParam * sizeof(VMValue));
offsetS = offsetF + (int)(sfunc->NumRegF * sizeof(double));
offsetA = offsetS + (int)(sfunc->NumRegS * sizeof(FString));
offsetD = offsetA + (int)(sfunc->NumRegA * sizeof(void*));
offsetExtra = (offsetD + (int)(sfunc->NumRegD * sizeof(int32_t)) + 15) & ~15;
vmframe = cc.newIntPtr("vmframe");
if (sfunc->SpecialInits.Size() == 0 && sfunc->NumRegS == 0)
{
// This is a simple frame with no constructors or destructors. Allocate it on the stack ourselves.
auto vmstack = cc.newStack(sfunc->StackSize, 16, "vmstack");
cc.lea(vmframe, vmstack);
auto slowinit = cc.newLabel();
auto endinit = cc.newLabel();
cc.cmp(numargs, sfunc->NumArgs);
cc.jne(slowinit);
SetupSimpleFrame(vmstack);
cc.jmp(endinit);
cc.bind(slowinit);
SetupSimpleFrameMissingArgs(vmstack); // Does this ever happen?
cc.bind(endinit);
}
else
{
SetupFullVMFrame();
}
}
void JitCompiler::SetupSimpleFrame(asmjit::X86Mem vmstack)
{
using namespace asmjit;
int argsPos = 0;
int regd = 0, regf = 0, rega = 0;
for (unsigned int i = 0; i < sfunc->Proto->ArgumentTypes.Size(); i++)
{
const PType *type = sfunc->Proto->ArgumentTypes[i];
if (sfunc->ArgFlags[i] & (VARF_Out | VARF_Ref))
{
cc.mov(regA[rega++], x86::ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, a)));
}
else if (type == TypeVector2)
{
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
}
else if (type == TypeVector3)
{
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
}
else if (type == TypeFloat64)
{
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
}
else if (type == TypeString)
{
I_FatalError("JIT: Strings are not supported yet for simple frames");
}
else if (type->isIntCompatible())
{
cc.mov(regD[regd++], x86::dword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, i)));
}
else
{
cc.mov(regA[rega++], x86::ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, a)));
}
}
if (sfunc->NumArgs != argsPos || regd > sfunc->NumRegD || regf > sfunc->NumRegF || rega > sfunc->NumRegA)
I_FatalError("JIT: sfunc->NumArgs != argsPos || regd > sfunc->NumRegD || regf > sfunc->NumRegF || rega > sfunc->NumRegA");
for (int i = regd; i < sfunc->NumRegD; i++)
cc.xor_(regD[i], regD[i]);
for (int i = regf; i < sfunc->NumRegF; i++)
cc.xorpd(regF[i], regF[i]);
for (int i = rega; i < sfunc->NumRegA; i++)
cc.xor_(regA[i], regA[i]);
}
void JitCompiler::SetupSimpleFrameMissingArgs(asmjit::X86Mem vmstack)
{
using namespace asmjit;
auto sfuncptr = newTempIntPtr();
cc.mov(sfuncptr, imm_ptr(sfunc));
if (cc.is64Bit())
cc.mov(x86::qword_ptr(vmframe, offsetof(VMFrame, Func)), sfuncptr);
else
cc.mov(x86::dword_ptr(vmframe, offsetof(VMFrame, Func)), sfuncptr);
cc.mov(x86::byte_ptr(vmframe, offsetof(VMFrame, NumRegD)), sfunc->NumRegD);
cc.mov(x86::byte_ptr(vmframe, offsetof(VMFrame, NumRegF)), sfunc->NumRegF);
cc.mov(x86::byte_ptr(vmframe, offsetof(VMFrame, NumRegS)), sfunc->NumRegS);
cc.mov(x86::byte_ptr(vmframe, offsetof(VMFrame, NumRegA)), sfunc->NumRegA);
cc.mov(x86::word_ptr(vmframe, offsetof(VMFrame, MaxParam)), sfunc->MaxParam);
cc.mov(x86::word_ptr(vmframe, offsetof(VMFrame, NumParam)), 0);
// Zero initialize the variables (retardedly stupid to do here - should be done by the compiler backend!!)
unsigned int clearbegin = (unsigned int)offsetof(VMFrame, NumParam) + 2;
unsigned int clearend = sfunc->StackSize;
unsigned int sseend = clearbegin + (clearend - clearbegin) / 16 * 16;
if (clearbegin < sseend)
{
auto zerosse = newTempXmmPd();
cc.xorpd(zerosse, zerosse);
for (unsigned int i = clearbegin; i < sseend; i += 16)
cc.movupd(x86::ptr(vmframe, i), zerosse);
}
if (sseend < clearend)
{
auto zero32 = newTempInt32();
cc.xor_(zero32, zero32);
unsigned int dwordend = sseend + (clearend - sseend) / 4 * 4;
for (unsigned int i = sseend; i < dwordend; i += 4)
cc.mov(asmjit::x86::dword_ptr(vmframe, i), zero32);
for (unsigned int i = dwordend; i < clearend; i++)
cc.mov(asmjit::x86::byte_ptr(vmframe, i), zero32.r8Lo());
}
auto fillParams = CreateCall<void, VMFrame *, VMValue *, int>([](VMFrame *newf, VMValue *args, int numargs) {
try
{
VMFillParams(args, newf, numargs);
}
catch (...)
{
VMThrowException(std::current_exception());
}
});
fillParams->setArg(0, vmframe);
fillParams->setArg(1, args);
fillParams->setArg(2, numargs);
for (int i = 0; i < sfunc->NumRegD; i++)
cc.mov(regD[i], x86::dword_ptr(vmframe, offsetD + i * sizeof(int32_t)));
for (int i = 0; i < sfunc->NumRegF; i++)
cc.movsd(regF[i], x86::qword_ptr(vmframe, offsetF + i * sizeof(double)));
for (int i = 0; i < sfunc->NumRegS; i++)
cc.lea(regS[i], x86::ptr(vmframe, offsetS + i * sizeof(FString)));
for (int i = 0; i < sfunc->NumRegA; i++)
cc.mov(regA[i], x86::ptr(vmframe, offsetA + i * sizeof(void*)));
}
void JitCompiler::SetupFullVMFrame()
{
using namespace asmjit;
stack = cc.newIntPtr("stack");
auto allocFrame = CreateCall<VMFrameStack *, VMScriptFunction *, VMValue *, int>([](VMScriptFunction *func, VMValue *args, int numargs) -> VMFrameStack* {
try
{
VMFrameStack *stack = &GlobalVMStack;
VMFrame *newf = stack->AllocFrame(func);
CurrentJitExceptInfo->vmframes++;
VMFillParams(args, newf, numargs);
return stack;
}
catch (...)
{
VMThrowException(std::current_exception());
return nullptr;
}
});
allocFrame->setRet(0, stack);
allocFrame->setArg(0, imm_ptr(sfunc));
allocFrame->setArg(1, args);
allocFrame->setArg(2, numargs);
cc.mov(vmframe, x86::ptr(stack)); // stack->Blocks
cc.mov(vmframe, x86::ptr(vmframe, VMFrameStack::OffsetLastFrame())); // Blocks->LastFrame
for (int i = 0; i < sfunc->NumRegD; i++)
cc.mov(regD[i], x86::dword_ptr(vmframe, offsetD + i * sizeof(int32_t)));
for (int i = 0; i < sfunc->NumRegF; i++)
cc.movsd(regF[i], x86::qword_ptr(vmframe, offsetF + i * sizeof(double)));
for (int i = 0; i < sfunc->NumRegS; i++)
cc.lea(regS[i], x86::ptr(vmframe, offsetS + i * sizeof(FString)));
for (int i = 0; i < sfunc->NumRegA; i++)
cc.mov(regA[i], x86::ptr(vmframe, offsetA + i * sizeof(void*)));
}
void JitCompiler::EmitPopFrame()
{
if (sfunc->SpecialInits.Size() != 0 || sfunc->NumRegS != 0)
{
auto popFrame = CreateCall<void, VMFrameStack *>([](VMFrameStack *stack) {
try
{
stack->PopFrame();
CurrentJitExceptInfo->vmframes--;
}
catch (...)
{
VMThrowException(std::current_exception());
}
});
popFrame->setArg(0, stack);
}
}
void JitCompiler::IncrementVMCalls()
{
// VMCalls[0]++
auto vmcallsptr = newTempIntPtr();
auto vmcalls = newTempInt32();
cc.mov(vmcallsptr, asmjit::imm_ptr(VMCalls));
cc.mov(vmcalls, asmjit::x86::dword_ptr(vmcallsptr));
cc.add(vmcalls, (int)1);
cc.mov(asmjit::x86::dword_ptr(vmcallsptr), vmcalls);
}
void JitCompiler::CreateRegisters()
{
regD.Resize(sfunc->NumRegD);
regF.Resize(sfunc->NumRegF);
regA.Resize(sfunc->NumRegA);
@ -499,205 +738,6 @@ void JitCompiler::Setup()
regname.Format("regA%d", i);
regA[i] = cc.newIntPtr(regname.GetChars());
}
labels.Resize(sfunc->CodeSize);
// VMCalls[0]++
auto vmcallsptr = newTempIntPtr();
auto vmcalls = newTempInt32();
cc.mov(vmcallsptr, imm_ptr(VMCalls));
cc.mov(vmcalls, x86::dword_ptr(vmcallsptr));
cc.add(vmcalls, (int)1);
cc.mov(x86::dword_ptr(vmcallsptr), vmcalls);
// the VM version reads this from the stack, but it is constant data
offsetParams = ((int)sizeof(VMFrame) + 15) & ~15;
offsetF = offsetParams + (int)(sfunc->MaxParam * sizeof(VMValue));
offsetS = offsetF + (int)(sfunc->NumRegF * sizeof(double));
offsetA = offsetS + (int)(sfunc->NumRegS * sizeof(FString));
offsetD = offsetA + (int)(sfunc->NumRegA * sizeof(void*));
offsetExtra = (offsetD + (int)(sfunc->NumRegD * sizeof(int32_t)) + 15) & ~15;
vmframe = cc.newIntPtr("vmframe");
if (sfunc->SpecialInits.Size() == 0 && sfunc->NumRegS == 0)
{
// This is a simple frame with no constructors or destructors. Allocate it on the stack ourselves.
auto vmstack = cc.newStack(sfunc->StackSize, 16, "vmstack");
cc.lea(vmframe, vmstack);
auto slowinit = cc.newLabel();
auto endinit = cc.newLabel();
#if 0 // this crashes sometimes
cc.cmp(numargs, sfunc->NumArgs);
cc.jne(slowinit);
// Is there a better way to know the type than this?
int argsPos = 0;
int regd = 0, regf = 0, rega = 0;
for (unsigned int i = 0; i < sfunc->Proto->ArgumentTypes.Size(); i++)
{
const PType *type = sfunc->Proto->ArgumentTypes[i];
if (type->isPointer())
{
cc.mov(regA[rega++], x86::ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, a)));
}
else if (type->isIntCompatible())
{
cc.mov(regD[regd++], x86::dword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, i)));
}
else if (type == TypeVector2)
{
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
}
else if (type == TypeVector3)
{
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
}
else if (type->isFloat())
{
cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
}
else if (type == TypeString)
{
I_FatalError("JIT: Strings are not supported yet for simple frames");
}
}
if (sfunc->NumArgs != argsPos || regd > sfunc->NumRegD || regf > sfunc->NumRegF || rega > sfunc->NumRegA)
I_FatalError("JIT: sfunc->NumArgs != argsPos || regd > sfunc->NumRegD || regf > sfunc->NumRegF || rega > sfunc->NumRegA");
cc.jmp(endinit);
#endif
cc.bind(slowinit);
auto sfuncptr = newTempIntPtr();
cc.mov(sfuncptr, imm_ptr(sfunc));
if (cc.is64Bit())
cc.mov(x86::qword_ptr(vmframe, offsetof(VMFrame, Func)), sfuncptr);
else
cc.mov(x86::dword_ptr(vmframe, offsetof(VMFrame, Func)), sfuncptr);
cc.mov(x86::byte_ptr(vmframe, offsetof(VMFrame, NumRegD)), sfunc->NumRegD);
cc.mov(x86::byte_ptr(vmframe, offsetof(VMFrame, NumRegF)), sfunc->NumRegF);
cc.mov(x86::byte_ptr(vmframe, offsetof(VMFrame, NumRegS)), sfunc->NumRegS);
cc.mov(x86::byte_ptr(vmframe, offsetof(VMFrame, NumRegA)), sfunc->NumRegA);
cc.mov(x86::word_ptr(vmframe, offsetof(VMFrame, MaxParam)), sfunc->MaxParam);
cc.mov(x86::word_ptr(vmframe, offsetof(VMFrame, NumParam)), 0);
// Zero initialize the variables (retardedly stupid to do here - should be done by the compiler backend!!)
unsigned int clearbegin = (unsigned int)offsetof(VMFrame, NumParam) + 2;
unsigned int clearend = sfunc->StackSize;
unsigned int sseend = clearbegin + (clearend - clearbegin) / 16 * 16;
if (clearbegin < sseend)
{
auto zerosse = newTempXmmPd();
cc.xorpd(zerosse, zerosse);
for (unsigned int i = clearbegin; i < sseend; i += 16)
cc.movupd(x86::ptr(vmframe, i), zerosse);
}
if (sseend < clearend)
{
auto zero32 = newTempInt32();
cc.xor_(zero32, zero32);
unsigned int dwordend = sseend + (clearend - sseend) / 4 * 4;
for (unsigned int i = sseend; i < dwordend; i += 4)
cc.mov(asmjit::x86::dword_ptr(vmframe, i), zero32);
for (unsigned int i = dwordend; i < clearend; i++)
cc.mov(asmjit::x86::byte_ptr(vmframe, i), zero32.r8Lo());
}
auto fillParams = CreateCall<void, VMFrame *, VMValue *, int>([](VMFrame *newf, VMValue *args, int numargs) {
try
{
VMFillParams(args, newf, numargs);
}
catch (...)
{
VMThrowException(std::current_exception());
}
});
fillParams->setArg(0, vmframe);
fillParams->setArg(1, args);
fillParams->setArg(2, numargs);
for (int i = 0; i < sfunc->NumRegD; i++)
cc.mov(regD[i], x86::dword_ptr(vmframe, offsetD + i * sizeof(int32_t)));
for (int i = 0; i < sfunc->NumRegF; i++)
cc.movsd(regF[i], x86::qword_ptr(vmframe, offsetF + i * sizeof(double)));
for (int i = 0; i < sfunc->NumRegS; i++)
cc.lea(regS[i], x86::ptr(vmframe, offsetS + i * sizeof(FString)));
for (int i = 0; i < sfunc->NumRegA; i++)
cc.mov(regA[i], x86::ptr(vmframe, offsetA + i * sizeof(void*)));
cc.bind(endinit);
}
else
{
stack = cc.newIntPtr("stack");
auto allocFrame = CreateCall<VMFrameStack *, VMScriptFunction *, VMValue *, int>([](VMScriptFunction *func, VMValue *args, int numargs) -> VMFrameStack* {
try
{
VMFrameStack *stack = &GlobalVMStack;
VMFrame *newf = stack->AllocFrame(func);
CurrentJitExceptInfo->vmframes++;
VMFillParams(args, newf, numargs);
return stack;
}
catch (...)
{
VMThrowException(std::current_exception());
return nullptr;
}
});
allocFrame->setRet(0, stack);
allocFrame->setArg(0, imm_ptr(sfunc));
allocFrame->setArg(1, args);
allocFrame->setArg(2, numargs);
cc.mov(vmframe, x86::ptr(stack)); // stack->Blocks
cc.mov(vmframe, x86::ptr(vmframe, VMFrameStack::OffsetLastFrame())); // Blocks->LastFrame
for (int i = 0; i < sfunc->NumRegD; i++)
cc.mov(regD[i], x86::dword_ptr(vmframe, offsetD + i * sizeof(int32_t)));
for (int i = 0; i < sfunc->NumRegF; i++)
cc.movsd(regF[i], x86::qword_ptr(vmframe, offsetF + i * sizeof(double)));
for (int i = 0; i < sfunc->NumRegS; i++)
cc.lea(regS[i], x86::ptr(vmframe, offsetS + i * sizeof(FString)));
for (int i = 0; i < sfunc->NumRegA; i++)
cc.mov(regA[i], x86::ptr(vmframe, offsetA + i * sizeof(void*)));
}
}
void JitCompiler::EmitPopFrame()
{
if (sfunc->SpecialInits.Size() != 0 || sfunc->NumRegS != 0)
{
auto popFrame = CreateCall<void, VMFrameStack *>([](VMFrameStack *stack) {
try
{
stack->PopFrame();
CurrentJitExceptInfo->vmframes--;
}
catch (...)
{
VMThrowException(std::current_exception());
}
});
popFrame->setArg(0, stack);
}
}
void JitCompiler::EmitNullPointerThrow(int index, EVMAbortException reason)

View File

@ -39,6 +39,12 @@ private:
#undef xx
void Setup();
void CreateRegisters();
void IncrementVMCalls();
void SetupFrame();
void SetupSimpleFrame(asmjit::X86Mem vmstack);
void SetupSimpleFrameMissingArgs(asmjit::X86Mem vmstack);
void SetupFullVMFrame();
void BindLabels();
void EmitOpcode();
void EmitPopFrame();

View File

@ -329,6 +329,7 @@ public:
FString PrintableName; // so that the VM can print meaningful info if something in this function goes wrong.
class PPrototype *Proto;
TArray<uint32_t> ArgFlags; // Should be the same length as Proto->ArgumentTypes
int(*ScriptCall)(VMFunction *func, VMValue *params, int numparams, VMReturn *ret, int numret) = nullptr;