- avoid calling VMFillParams for simple stack frames

- split Setup into more functions

src/scripting/backend/vmbuilder.cpp

@@ -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

src/scripting/vm/jit.cpp

@@ -471,45 +471,15 @@ void JitCompiler::Setup()
 	konsts = sfunc->KonstS;
 	konsta = sfunc->KonstA;
 
-	regD.Resize(sfunc->NumRegD);
-	regF.Resize(sfunc->NumRegF);
-	regA.Resize(sfunc->NumRegA);
-	regS.Resize(sfunc->NumRegS);
-
-	for (int i = 0; i < sfunc->NumRegD; i++)
-	{
-		regname.Format("regD%d", i);
-		regD[i] = cc.newInt32(regname.GetChars());
-	}
-
-	for (int i = 0; i < sfunc->NumRegF; i++)
-	{
-		regname.Format("regF%d", i);
-		regF[i] = cc.newXmmSd(regname.GetChars());
-	}
-
-	for (int i = 0; i < sfunc->NumRegS; i++)
-	{
-		regname.Format("regS%d", i);
-		regS[i] = cc.newIntPtr(regname.GetChars());
-	}
-
-	for (int i = 0; i < sfunc->NumRegA; i++)
-	{
-		regname.Format("regA%d", i);
-		regA[i] = cc.newIntPtr(regname.GetChars());
-	}
-
 	labels.Resize(sfunc->CodeSize);
 
-	// VMCalls[0]++
+	CreateRegisters();
-	auto vmcallsptr = newTempIntPtr();
+	IncrementVMCalls();
-	auto vmcalls = newTempInt32();
+	SetupFrame();
-	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);
 
+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));
@@ -530,24 +500,33 @@ void JitCompiler::Setup()
 		auto slowinit = cc.newLabel();
 		auto endinit = cc.newLabel();
 
-#if 0 // this crashes sometimes
 		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;
 
-		// 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())
+		if (sfunc->ArgFlags[i] & (VARF_Out | VARF_Ref))
 		{
 			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)));
@@ -559,7 +538,7 @@ void JitCompiler::Setup()
 			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())
+		else if (type == TypeFloat64)
 		{
 			cc.movsd(regF[regf++], x86::qword_ptr(args, argsPos++ * sizeof(VMValue) + offsetof(VMValue, f)));
 		}
@@ -567,14 +546,32 @@ void JitCompiler::Setup()
 		{
 			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");
 
-		cc.jmp(endinit);
+	for (int i = regd; i < sfunc->NumRegD; i++)
-#endif
+		cc.xor_(regD[i], regD[i]);
-		cc.bind(slowinit);
+
+	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));
@@ -638,11 +635,12 @@ void JitCompiler::Setup()
 
 	for (int i = 0; i < sfunc->NumRegA; i++)
 		cc.mov(regA[i], x86::ptr(vmframe, offsetA + i * sizeof(void*)));
-
-		cc.bind(endinit);
 }
-	else
+
+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
@@ -679,7 +677,6 @@ void JitCompiler::Setup()
 	for (int i = 0; i < sfunc->NumRegA; i++)
 		cc.mov(regA[i], x86::ptr(vmframe, offsetA + i * sizeof(void*)));
 }
-}
 
 void JitCompiler::EmitPopFrame()
 {
@@ -700,6 +697,49 @@ void JitCompiler::EmitPopFrame()
 	}
 }
 
+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);
+	regS.Resize(sfunc->NumRegS);
+
+	for (int i = 0; i < sfunc->NumRegD; i++)
+	{
+		regname.Format("regD%d", i);
+		regD[i] = cc.newInt32(regname.GetChars());
+	}
+
+	for (int i = 0; i < sfunc->NumRegF; i++)
+	{
+		regname.Format("regF%d", i);
+		regF[i] = cc.newXmmSd(regname.GetChars());
+	}
+
+	for (int i = 0; i < sfunc->NumRegS; i++)
+	{
+		regname.Format("regS%d", i);
+		regS[i] = cc.newIntPtr(regname.GetChars());
+	}
+
+	for (int i = 0; i < sfunc->NumRegA; i++)
+	{
+		regname.Format("regA%d", i);
+		regA[i] = cc.newIntPtr(regname.GetChars());
+	}
+}
+
 void JitCompiler::EmitNullPointerThrow(int index, EVMAbortException reason)
 {
 	auto label = EmitThrowExceptionLabel(reason);

src/scripting/vm/jitintern.h

@@ -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();

src/scripting/vm/vm.h

@@ -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;