private void ForwardDeclare(Assembly assembly)
{
// Forward declare procedures
foreach (var proc in assembly.Procedures)
{
var x86proc = new X86Proc(GetSymbolName(proc));
procs[proc] = x86proc;
// Forward declare each basic block inside
foreach (var bb in proc.BasicBlocks)
{
basicBlocks[bb] = new X86BasicBlock(x86proc, formatOptions.Escape(bb.Name));
}
}
}
private void CompileInstr(Proc proc, Instr instr)
{
registerPool.FreeAll();
CommentInstr(instr);
switch (instr)
{
case Instr.Call call:
{
if (!(call.Procedure.Type is Type.Proc procTy))
{
throw new InvalidOperationException();
}
if (procTy.CallConv == CallConv.Cdecl)
{
// Push arguments backwards, track the offset of esp
var espOffset = 0;
int i = 0;
foreach (var arg in call.Arguments.Reverse())
{
var argValue = CompileValue(arg);
CommentInstr($"argument {i}");
WritePush(argValue);
espOffset += SizeOf(arg);
registerPool.FreeAll();
++i;
}
// If the return method is by address, we pass the address
var returnMethod = GetReturnMethod(procTy.Return);
if (returnMethod == ReturnMethod.Address)
{
var resultAddr = CompileToAddress(call.Result);
var resultAddrReg = registerPool.Allocate(DataWidth.dword);
CommentInstr("return target address");
WriteInstr(X86Op.Lea, resultAddrReg, resultAddr);
WriteInstr(X86Op.Push, resultAddrReg);
espOffset += DataWidth.dword.Size;
registerPool.FreeAll();
}
// Do the call
var procedure = CompileSingleValue(call.Procedure);
WriteInstr(X86Op.Call, procedure);
// Restore stack
WriteInstr(X86Op.Add, Register.esp, new Operand.Literal(DataWidth.dword, espOffset));
// Writing the result back
if (returnMethod == ReturnMethod.None || returnMethod == ReturnMethod.Address)
{
// No-op
// For size == 0 there's nothing to copy, for size > 4 the value is already copied
}
else
{
CommentInstr("copy return value");
// Just store what's written in eax or eax:edx
var resultSize = SizeOf(procTy.Return);
var result = CompileValue(call.Result);
if (result.Length == 1)
{
var eax = Register.AtSlot(0, DataWidth.GetFromSize(resultSize));
WriteInstr(X86Op.Mov, result[0], eax);
}
else
{
Debug.Assert(result.Length == 2);
var eax = Register.AtSlot(0, DataWidth.dword);
var edx = Register.AtSlot(2, DataWidth.GetFromSize(resultSize - 4));
WriteInstr(X86Op.Mov, result[0], eax);
WriteInstr(X86Op.Mov, result[1], edx);
}
}
}
else
{
throw new NotImplementedException();
}
}
break;
case Instr.Ret ret:
{
// Writing the return value
if (proc.CallConv == CallConv.Cdecl)
{
var returnMethod = GetReturnMethod(ret.Value.Type);
if (returnMethod == ReturnMethod.None)
{
// No-op
}
else if (returnMethod == ReturnMethod.Eax || returnMethod == ReturnMethod.EaxEdx)
{
// We write in eax or the eax:edx pair
// TODO: Floats are returned in a different register!
// NOTE: Might be unnecessary allocation for edx
registerPool.Allocate(Register.eax, Register.edx);
var retValue = CompileValue(ret.Value);
if (retValue.Length == 1)
{
WriteInstr(X86Op.Mov, Register.eax, retValue[0]);
}
else
{
Debug.Assert(retValue.Length == 2);
WriteInstr(X86Op.Mov, Register.eax, retValue[0]);
WriteInstr(X86Op.Mov, Register.edx, retValue[1]);
}
}
else
{
// We receive a return address
// First we load that return address
var retAddrAddr = new Operand.Address(Register.ebp, 8);
var retAddr = registerPool.Allocate(DataWidth.dword);
CommentInstr("copy return value");
WriteInstr(X86Op.Mov, retAddr, retAddrAddr);
// Compile the value
var retValue = CompileValue(ret.Value);
// Copy
WriteCopy(retAddr, retValue);
}
}
else
{
throw new NotImplementedException();
}
// Write the epilogue, return
WriteProcEpilogue(proc);
CommentInstr(instr);
WriteInstr(X86Op.Ret);
}
break;
case Instr.Jmp jmp:
WriteInstr(X86Op.Jmp, basicBlocks[jmp.Target]);
break;
case Instr.JmpIf jmpIf:
{
var condParts = CompileValue(jmpIf.Condition);
Debug.Assert(condParts.Length > 0);
// Initial condition is the first bytes
var condHolder = registerPool.Allocate(condParts[0].GetWidth(sizeContext));
WriteInstr(X86Op.Mov, condHolder, condParts[0]);
// Remaining are or-ed to the result
foreach (var item in condParts.Skip(1))
{
WriteInstr(X86Op.Or, condHolder, item);
}
WriteInstr(X86Op.Test, condHolder, condHolder);
WriteInstr(X86Op.Jne, basicBlocks[jmpIf.Then]);
WriteInstr(X86Op.Jmp, basicBlocks[jmpIf.Else]);
}
break;
case Instr.Store store:
{
var target = CompileToAddress(store.Target);
var value = CompileValue(store.Value);
var address = registerPool.Allocate(DataWidth.dword);
WriteInstr(X86Op.Mov, address, target);
WriteCopy(address, value);
}
break;
case Instr.Load load:
{
var target = CompileToAddress(load.Result);
var source = CompileToAddress(load.Source);
var targetAddr = registerPool.Allocate(DataWidth.dword);
var sourceAddr = registerPool.Allocate(DataWidth.dword);
WriteInstr(X86Op.Lea, targetAddr, target);
WriteInstr(X86Op.Mov, sourceAddr, source);
WriteMemcopy(targetAddr, sourceAddr, SizeOf(load.Result));
}
break;
case Instr.Alloc alloc:
{
var size = SizeOf(alloc.Allocated);
WriteInstr(X86Op.Sub, Register.esp, new Operand.Literal(DataWidth.dword, size));
var result = CompileToAddress(alloc.Result);
WriteInstr(X86Op.Mov, result, Register.esp);
}
break;
case Instr.ElementPtr elementPtr:
{
var target = CompileToAddress(elementPtr.Result);
var value = CompileSingleValue(elementPtr.Value);
// We need it in a register
value = LoadToRegister(value);
var structTy = (Struct)((Type.Ptr)elementPtr.Value.Type).Subtype;
var index = elementPtr.Index.Value;
// Get offset, add it to the base address
var offset = OffsetOf(structTy, index);
WriteInstr(X86Op.Add, value, new Operand.Literal(DataWidth.dword, offset));
WriteMov(target, value);
}
break;
case Instr.Cast cast:
{
if (cast.Target is Type.Ptr && cast.Value.Type is Type.Ptr)
{
// Should be a no-op, simply copy
var target = CompileToAddress(cast.Result);
var src = CompileSingleValue(cast.Value);
WriteMov(target, src);
}
else
{
throw new InvalidOperationException();
}
}
break;
// Size-dependent operations
case Instr.Cmp cmp:
{
var targetParts = CompileValue(cmp.Result);
var leftParts = CompileValue(cmp.Left);
var rightParts = CompileValue(cmp.Right);
Debug.Assert(leftParts.Length == rightParts.Length);
var target = targetParts[0];
var targetWidth = target.GetWidth(sizeContext);
var truthy = new Operand.Literal(targetWidth, 1);
var falsy = new Operand.Literal(targetWidth, 0);
// We need to branch to write the result
var labelNameBase = GetUniqueName("cmp_result");
Debug.Assert(currentProcedure != null);
var trueBB = new X86BasicBlock(currentProcedure, $"{labelNameBase}_T");
var falseBB = new X86BasicBlock(currentProcedure, $"{labelNameBase}_F");
var finallyBB = new X86BasicBlock(currentProcedure, $"{labelNameBase}_C");
// Add all these basic blocks to the current procedure
Debug.Assert(currentProcedure != null);
currentProcedure.BasicBlocks.Add(trueBB);
currentProcedure.BasicBlocks.Add(falseBB);
currentProcedure.BasicBlocks.Add(finallyBB);
bool signed = ((Type.Int)cmp.Left.Type).Signed;
// For each part pair we compare
// NOTE: It's important that we start from the most significant bytes here for the relational operators
int i = 0;
foreach (var(l, r) in leftParts.Zip(rightParts).Reverse())
{
// NOTE: first && signed means that we are looking at the MSB, so in case of a signed
// comparison, we need to use the signed compare jumps, unsigned in any other case
// TL;DR: fisrt && signed means we need signed jump operation
bool first = i == 0;
bool last = i == leftParts.Length - 1;
i += 1;
// We need to do the comparison
var tmp = WriteNonImmOrMemoryInstr(X86Op.Cmp, l, r);
if (tmp != null)
{
registerPool.Free(tmp);
}
if (last)
{
// If this was the last part to compare, we can just branch to the truthy part as the
// condition succeeded (or the falsy one on mismatch)
var inverseOp = ComparisonToJump(cmp.Comparison.Inverse, first && signed);
WriteInstr(inverseOp, falseBB);
WriteInstr(X86Op.Jmp, trueBB);
}
else
{
// These are not the last elements we compare
if (cmp.Comparison == Comparison.eq)
{
// We can jump to false on inequality
WriteInstr(X86Op.Jne, falseBB);
}
else if (cmp.Comparison == Comparison.ne)
{
// We can jump to true on inequality
WriteInstr(X86Op.Jne, trueBB);
}
else if (cmp.Comparison == Comparison.le || cmp.Comparison == Comparison.le_eq)
{
// If the part less, then we can jump to true, if greater, then to false
var lessOp = ComparisonToJump(Comparison.le, first && signed);
var greaterOp = ComparisonToJump(Comparison.gr, first && signed);
WriteInstr(lessOp, trueBB);
WriteInstr(greaterOp, falseBB);
}
else if (cmp.Comparison == Comparison.gr || cmp.Comparison == Comparison.gr_eq)
{
// If the part is greater, then we can jump to true, if less, then to false
var greaterOp = ComparisonToJump(Comparison.gr, first && signed);
var lessOp = ComparisonToJump(Comparison.le, first && signed);
WriteInstr(greaterOp, trueBB);
WriteInstr(lessOp, falseBB);
}
}
}
// On true block, we write the truthy value then jump to the continuation
currentBasicBlock = trueBB;
WriteInstr(X86Op.Mov, target, truthy);
WriteInstr(X86Op.Jmp, finallyBB);
// On false block, we write the falsy value then jump to the continuation
currentBasicBlock = falseBB;
WriteInstr(X86Op.Mov, target, falsy);
WriteInstr(X86Op.Jmp, finallyBB);
// We continue writing on the continuation
currentBasicBlock = finallyBB;
}
break;
case Instr.Add:
case Instr.Sub:
{
var arith = (ArithInstr)instr;
var target = CompileToAddress(arith.Result);
if (arith.Left.Type is Type.Ptr leftPtr)
{
// Pointer arithmetic
var left = CompileSingleValue(arith.Left);
var right = CompileSingleValue(arith.Right);
right = LoadToRegister(right);
WriteMov(target, left);
// We multiply the offset by the data size to get the expected behavior
WriteInstr(X86Op.Imul, right, new Operand.Literal(DataWidth.dword, SizeOf(leftPtr.Subtype)));
WriteInstr(arith is Instr.Add ? X86Op.Add : X86Op.Sub, target, right);
}
else if (arith.Left.Type is Type.Int && arith.Right.Type is Type.Int)
{
// Integer arithmetic
var targetAddr = registerPool.Allocate(DataWidth.dword);
WriteInstr(X86Op.Lea, targetAddr, target);
var left = CompileValue(arith.Left);
var right = CompileValue(arith.Right);
Debug.Assert(left.Length == right.Length);
var firstOp = arith is Instr.Add ? X86Op.Add : X86Op.Sub;
var remOps = arith is Instr.Add ? X86Op.Adc : X86Op.Sbb;
int offset = 0;
bool first = true;
foreach (var(l, r) in left.Zip(right))
{
var width = l.GetWidth(sizeContext);
var addr = new Operand.Address(targetAddr, offset);
var displ = new Operand.Indirect(width, addr);
WriteMov(displ, l);
if (first)
{
// We use add or sub
WriteNonImmOrMemoryInstr(firstOp, displ, r);
}
else
{
// We use adc or sbb
WriteNonImmOrMemoryInstr(remOps, displ, r);
}
offset += width.Size;
first = false;
}
}
else
{
throw new InvalidOperationException();
}
}
break;
case Instr.Mul mul:
{
if (SizeOf(mul.Left.Type) > 4)
{
// For now we skip this
throw new NotSupportedException("Multiplication of > 4 byte operands is not supported!");
}
var target = CompileToAddress(mul.Result);
var left = CompileSingleValue(mul.Left);
var right = CompileSingleValue(mul.Right);
var tmp = registerPool.Allocate(DataWidth.dword);
WriteMov(tmp, left);
WriteInstr(X86Op.Imul, tmp, right);
WriteMov(target, tmp);
}
break;
case Instr.Div:
case Instr.Mod:
{
var arith = (ArithInstr)instr;
if (SizeOf(arith.Left.Type) > 4)
{
// For now we skip this
throw new NotSupportedException("Division of > 4 byte operands is not supported!");
}
// NOTE: This is different from multiplication
registerPool.Allocate(Register.eax, Register.edx);
var target = CompileToAddress(arith.Result);
var left = CompileSingleValue(arith.Left);
var right = CompileSingleValue(arith.Right);
right = LoadToRegister(right);
WriteInstr(X86Op.Mov, Register.edx, new Operand.Literal(DataWidth.dword, 0));
WriteInstr(X86Op.Mov, Register.eax, left);
WriteInstr(X86Op.Idiv, right);
WriteInstr(X86Op.Mov, target, arith is Instr.Div ? Register.eax : Register.edx);
}
break;
case Instr.BitAnd:
case Instr.BitOr:
case Instr.BitXor:
{
var bitw = (BitwiseInstr)instr;
var target = CompileToAddress(bitw.Result);
var leftParts = CompileValue(bitw.Left);
var rightParts = CompileValue(bitw.Right);
var op = bitw switch
{
Instr.BitAnd => X86Op.And,
Instr.BitOr => X86Op.Or,
Instr.BitXor => X86Op.Xor,
_ => throw new NotImplementedException(),
};
// We can just apply the operation for each part
Debug.Assert(leftParts.Length == rightParts.Length);
var targetAddr = registerPool.Allocate(DataWidth.dword);
WriteInstr(X86Op.Lea, targetAddr, target);
int offset = 0;
foreach (var(l, r) in leftParts.Zip(rightParts))
{
var width = l.GetWidth(sizeContext);
var addr = new Operand.Address(targetAddr, offset);
var displ = new Operand.Indirect(width, addr);
WriteMov(displ, l);
WriteInstr(op, displ, r);
offset += width.Size;
}
}
break;
case Instr.Shl:
case Instr.Shr:
{
var bitsh = (BitShiftInstr)instr;
if (SizeOf(bitsh.Shifted.Type) > 4 || SizeOf(bitsh.Amount.Type) > 4)
{
// For now we skip this
throw new NotSupportedException("Shifting of > 4 byte operands is not supported!");
}
var target = CompileToAddress(bitsh.Result);
var left = CompileSingleValue(bitsh.Shifted);
var right = CompileSingleValue(bitsh.Amount);
left = LoadToRegister(left);
WriteInstr(bitsh is Instr.Shl ? X86Op.Shl : X86Op.Shr, left, right);
WriteMov(target, left);
}
break;
default: throw new NotImplementedException();
}
}