#include "llvm/iPHINode.h"
#include "llvm/iMemory.h"
#include "llvm/Type.h"
+#include "llvm/DerivedTypes.h"
#include "llvm/Constants.h"
#include "llvm/Pass.h"
#include "llvm/CodeGen/MachineFunction.h"
using namespace MOTy; // Get Use, Def, UseAndDef
+
+/// BMI - A special BuildMI variant that takes an iterator to insert the
+/// instruction at as well as a basic block.
+/// this is the version for when you have a destination register in mind.
+inline static MachineInstrBuilder BMI(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &I,
+ MachineOpCode Opcode,
+ unsigned NumOperands,
+ unsigned DestReg) {
+ assert(I >= MBB->begin() && I <= MBB->end() && "Bad iterator!");
+ MachineInstr *MI = new MachineInstr(Opcode, NumOperands+1, true, true);
+ I = ++MBB->insert(I, MI);
+ return MachineInstrBuilder(MI).addReg(DestReg, MOTy::Def);
+}
+
+/// BMI - A special BuildMI variant that takes an iterator to insert the
+/// instruction at as well as a basic block.
+inline static MachineInstrBuilder BMI(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &I,
+ MachineOpCode Opcode,
+ unsigned NumOperands) {
+ assert(I > MBB->begin() && I <= MBB->end() && "Bad iterator!");
+ MachineInstr *MI = new MachineInstr(Opcode, NumOperands, true, true);
+ I = ++MBB->insert(I, MI);
+ return MachineInstrBuilder(MI);
+}
+
+
namespace {
struct ISel : public FunctionPass, InstVisitor<ISel> {
TargetMachine &TM;
unsigned CurReg;
std::map<Value*, unsigned> RegMap; // Mapping between Val's and SSA Regs
+ // MBBMap - Mapping between LLVM BB -> Machine BB
+ std::map<const BasicBlock*, MachineBasicBlock*> MBBMap;
+
ISel(TargetMachine &tm)
: TM(tm), F(0), BB(0), CurReg(MRegisterInfo::FirstVirtualRegister) {}
///
bool runOnFunction(Function &Fn) {
F = &MachineFunction::construct(&Fn, TM);
+
+ for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
+ F->getBasicBlockList().push_back(MBBMap[I] = new MachineBasicBlock(I));
+
+ // Emit instructions to load the arguments... The function's arguments
+ // look like this:
+ //
+ // [EBP] -- copy of old EBP
+ // [EBP + 4] -- return address
+ // [EBP + 8] -- first argument (leftmost lexically)
+ //
+ // So we want to start with counter = 2.
+ //
+ BB = &F->front();
+ unsigned ArgOffset = 8;
+ for (Function::aiterator I = Fn.abegin(), E = Fn.aend(); I != E;
+ ++I, ArgOffset += 4) {
+ unsigned Reg = getReg(*I);
+
+ // Load it out of the stack frame at EBP + 4*argPos.
+
+ // FIXME: This should load the argument of the appropriate size!!
+ addRegOffset(BuildMI(BB, X86::MOVmr32, 4, Reg), X86::EBP, ArgOffset);
+ }
+
+ // Instruction select everything except PHI nodes
visit(Fn);
+
+ // Select the PHI nodes
+ SelectPHINodes();
+
RegMap.clear();
+ MBBMap.clear();
CurReg = MRegisterInfo::FirstVirtualRegister;
F = 0;
return false; // We never modify the LLVM itself.
}
+ virtual const char *getPassName() const {
+ return "X86 Simple Instruction Selection";
+ }
+
/// visitBasicBlock - This method is called when we are visiting a new basic
/// block. This simply creates a new MachineBasicBlock to emit code into
/// and adds it to the current MachineFunction. Subsequent visit* for
/// instructions will be invoked for all instructions in the basic block.
///
void visitBasicBlock(BasicBlock &LLVM_BB) {
- BB = new MachineBasicBlock(&LLVM_BB);
- // FIXME: Use the auto-insert form when it's available
- F->getBasicBlockList().push_back(BB);
+ BB = MBBMap[&LLVM_BB];
}
+
+ /// SelectPHINodes - Insert machine code to generate phis. This is tricky
+ /// because we have to generate our sources into the source basic blocks,
+ /// not the current one.
+ ///
+ void SelectPHINodes();
+
// Visitation methods for various instructions. These methods simply emit
// fixed X86 code for each instruction.
//
void visitSimpleBinary(BinaryOperator &B, unsigned OpcodeClass);
void visitAdd(BinaryOperator &B) { visitSimpleBinary(B, 0); }
void visitSub(BinaryOperator &B) { visitSimpleBinary(B, 1); }
+ void doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator &MBBI,
+ unsigned destReg, const Type *resultType,
+ unsigned op0Reg, unsigned op1Reg);
void visitMul(BinaryOperator &B);
void visitDiv(BinaryOperator &B) { visitDivRem(B); }
// Memory Instructions
void visitLoadInst(LoadInst &I);
void visitStoreInst(StoreInst &I);
-
+ void visitGetElementPtrInst(GetElementPtrInst &I);
+ void visitMallocInst(MallocInst &I);
+ void visitFreeInst(FreeInst &I);
+ void visitAllocaInst(AllocaInst &I);
+
// Other operators
void visitShiftInst(ShiftInst &I);
- void visitPHINode(PHINode &I);
+ void visitPHINode(PHINode &I) {} // PHI nodes handled by second pass
void visitCastInst(CastInst &I);
void visitInstruction(Instruction &I) {
abort();
}
+ /// promote32 - Make a value 32-bits wide, and put it somewhere.
+ void promote32 (const unsigned targetReg, Value *v);
+ // emitGEPOperation - Common code shared between visitGetElementPtrInst and
+ // constant expression GEP support.
+ //
+ void emitGEPOperation(MachineBasicBlock *BB, MachineBasicBlock::iterator&IP,
+ Value *Src, User::op_iterator IdxBegin,
+ User::op_iterator IdxEnd, unsigned TargetReg);
+
/// copyConstantToRegister - Output the instructions required to put the
/// specified constant into the specified register.
///
- void copyConstantToRegister(Constant *C, unsigned Reg);
+ void copyConstantToRegister(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &MBBI,
+ Constant *C, unsigned Reg);
+
+ /// makeAnotherReg - This method returns the next register number
+ /// we haven't yet used.
+ unsigned makeAnotherReg(const Type *Ty) {
+ // Add the mapping of regnumber => reg class to MachineFunction
+ F->addRegMap(CurReg, TM.getRegisterInfo()->getRegClassForType(Ty));
+ return CurReg++;
+ }
/// getReg - This method turns an LLVM value into a register number. This
/// is guaranteed to produce the same register number for a particular value
///
unsigned getReg(Value &V) { return getReg(&V); } // Allow references
unsigned getReg(Value *V) {
+ // Just append to the end of the current bb.
+ MachineBasicBlock::iterator It = BB->end();
+ return getReg(V, BB, It);
+ }
+ unsigned getReg(Value *V, MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &IPt) {
unsigned &Reg = RegMap[V];
if (Reg == 0) {
- Reg = CurReg++;
+ Reg = makeAnotherReg(V->getType());
RegMap[V] = Reg;
-
- // Add the mapping of regnumber => reg class to MachineFunction
- F->addRegMap(Reg,
- TM.getRegisterInfo()->getRegClassForType(V->getType()));
}
// If this operand is a constant, emit the code to copy the constant into
// the register here...
//
- if (Constant *C = dyn_cast<Constant>(V))
- copyConstantToRegister(C, Reg);
+ if (Constant *C = dyn_cast<Constant>(V)) {
+ copyConstantToRegister(MBB, IPt, C, Reg);
+ RegMap.erase(V); // Assign a new name to this constant if ref'd again
+ } else if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
+ // Move the address of the global into the register
+ BMI(MBB, IPt, X86::MOVir32, 1, Reg).addReg(GV);
+ RegMap.erase(V); // Assign a new name to this address if ref'd again
+ }
return Reg;
}
case Type::PointerTyID: return cInt; // Int's and pointers are class #2
case Type::LongTyID:
- case Type::ULongTyID: return cLong; // Longs are class #3
+ case Type::ULongTyID: //return cLong; // Longs are class #3
+ return cInt; // FIXME: LONGS ARE TREATED AS INTS!
+
case Type::FloatTyID: return cFloat; // Float is class #4
case Type::DoubleTyID: return cDouble; // Doubles are class #5
default:
}
}
+// getClassB - Just like getClass, but treat boolean values as bytes.
+static inline TypeClass getClassB(const Type *Ty) {
+ if (Ty == Type::BoolTy) return cByte;
+ return getClass(Ty);
+}
+
/// copyConstantToRegister - Output the instructions required to put the
/// specified constant into the specified register.
///
-void ISel::copyConstantToRegister(Constant *C, unsigned R) {
- assert (!isa<ConstantExpr>(C) && "Constant expressions not yet handled!\n");
+void ISel::copyConstantToRegister(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &IP,
+ Constant *C, unsigned R) {
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
+ if (CE->getOpcode() == Instruction::GetElementPtr) {
+ emitGEPOperation(MBB, IP, CE->getOperand(0),
+ CE->op_begin()+1, CE->op_end(), R);
+ return;
+ }
+
+ std::cerr << "Offending expr: " << C << "\n";
+ assert (0 && "Constant expressions not yet handled!\n");
+ }
if (C->getType()->isIntegral()) {
- unsigned Class = getClass(C->getType());
+ unsigned Class = getClassB(C->getType());
assert(Class != 3 && "Type not handled yet!");
static const unsigned IntegralOpcodeTab[] = {
X86::MOVir8, X86::MOVir16, X86::MOVir32
};
- if (C->getType()->isSigned()) {
+ if (C->getType() == Type::BoolTy) {
+ BMI(MBB, IP, X86::MOVir8, 1, R).addZImm(C == ConstantBool::True);
+ } else if (C->getType()->isSigned()) {
ConstantSInt *CSI = cast<ConstantSInt>(C);
- BuildMI(BB, IntegralOpcodeTab[Class], 1, R).addSImm(CSI->getValue());
+ BMI(MBB, IP, IntegralOpcodeTab[Class], 1, R).addSImm(CSI->getValue());
} else {
ConstantUInt *CUI = cast<ConstantUInt>(C);
- BuildMI(BB, IntegralOpcodeTab[Class], 1, R).addZImm(CUI->getValue());
+ BMI(MBB, IP, IntegralOpcodeTab[Class], 1, R).addZImm(CUI->getValue());
}
+ } else if (isa<ConstantPointerNull>(C)) {
+ // Copy zero (null pointer) to the register.
+ BMI(MBB, IP, X86::MOVir32, 1, R).addZImm(0);
+ } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)) {
+ unsigned SrcReg = getReg(CPR->getValue(), MBB, IP);
+ BMI(MBB, IP, X86::MOVrr32, 1, R).addReg(SrcReg);
} else {
+ std::cerr << "Offending constant: " << C << "\n";
assert(0 && "Type not handled yet!");
}
}
+/// SelectPHINodes - Insert machine code to generate phis. This is tricky
+/// because we have to generate our sources into the source basic blocks, not
+/// the current one.
+///
+void ISel::SelectPHINodes() {
+ const Function &LF = *F->getFunction(); // The LLVM function...
+ for (Function::const_iterator I = LF.begin(), E = LF.end(); I != E; ++I) {
+ const BasicBlock *BB = I;
+ MachineBasicBlock *MBB = MBBMap[I];
+
+ // Loop over all of the PHI nodes in the LLVM basic block...
+ unsigned NumPHIs = 0;
+ for (BasicBlock::const_iterator I = BB->begin();
+ PHINode *PN = (PHINode*)dyn_cast<PHINode>(&*I); ++I) {
+ // Create a new machine instr PHI node, and insert it.
+ MachineInstr *MI = BuildMI(X86::PHI, PN->getNumOperands(), getReg(*PN));
+ MBB->insert(MBB->begin()+NumPHIs++, MI); // Insert it at the top of the BB
+
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
+ MachineBasicBlock *PredMBB = MBBMap[PN->getIncomingBlock(i)];
+
+ // Get the incoming value into a virtual register. If it is not already
+ // available in a virtual register, insert the computation code into
+ // PredMBB
+ //
+
+ MachineBasicBlock::iterator PI = PredMBB->begin();
+ while ((*PI)->getOpcode() == X86::PHI) ++PI;
+
+ MI->addRegOperand(getReg(PN->getIncomingValue(i), PredMBB, PI));
+ MI->addMachineBasicBlockOperand(PredMBB);
+ }
+ }
+ }
+}
+
+
/// SetCC instructions - Here we just emit boilerplate code to set a byte-sized
/// register, then move it to wherever the result should be.
// FIXME: assuming var1, var2 are in memory, if not, spill to
// stack first
case cFloat: // Floats
- BuildMI (BB, X86::FLDr4, 1).addReg (reg1);
- BuildMI (BB, X86::FLDr4, 1).addReg (reg2);
+ BuildMI (BB, X86::FLDr32, 1).addReg (reg1);
+ BuildMI (BB, X86::FLDr32, 1).addReg (reg2);
break;
case cDouble: // Doubles
- BuildMI (BB, X86::FLDr8, 1).addReg (reg1);
- BuildMI (BB, X86::FLDr8, 1).addReg (reg2);
+ BuildMI (BB, X86::FLDr64, 1).addReg (reg1);
+ BuildMI (BB, X86::FLDr64, 1).addReg (reg2);
break;
case cLong:
default:
BuildMI (BB, X86::MOVrr8, 1, getReg(I)).addReg(X86::AL);
}
+/// promote32 - Emit instructions to turn a narrow operand into a 32-bit-wide
+/// operand, in the specified target register.
+void
+ISel::promote32 (unsigned targetReg, Value *v)
+{
+ unsigned vReg = getReg (v);
+ unsigned Class = getClass (v->getType ());
+ bool isUnsigned = v->getType ()->isUnsigned ();
+ assert (((Class == cByte) || (Class == cShort) || (Class == cInt))
+ && "Unpromotable operand class in promote32");
+ switch (Class)
+ {
+ case cByte:
+ // Extend value into target register (8->32)
+ if (isUnsigned)
+ BuildMI (BB, X86::MOVZXr32r8, 1, targetReg).addReg (vReg);
+ else
+ BuildMI (BB, X86::MOVSXr32r8, 1, targetReg).addReg (vReg);
+ break;
+ case cShort:
+ // Extend value into target register (16->32)
+ if (isUnsigned)
+ BuildMI (BB, X86::MOVZXr32r16, 1, targetReg).addReg (vReg);
+ else
+ BuildMI (BB, X86::MOVSXr32r16, 1, targetReg).addReg (vReg);
+ break;
+ case cInt:
+ // Move value into target register (32->32)
+ BuildMI (BB, X86::MOVrr32, 1, targetReg).addReg (vReg);
+ break;
+ }
+}
/// 'ret' instruction - Here we are interested in meeting the x86 ABI. As such,
/// we have the following possibilities:
/// ret long, ulong : Move value into EAX/EDX and return
/// ret float/double : Top of FP stack
///
-void ISel::visitReturnInst (ReturnInst &I) {
- if (I.getNumOperands() == 0) {
- // Emit a 'ret' instruction
- BuildMI(BB, X86::RET, 0);
- return;
- }
-
- unsigned val = getReg(I.getOperand(0));
- unsigned Class = getClass(I.getOperand(0)->getType());
- bool isUnsigned = I.getOperand(0)->getType()->isUnsigned();
- switch (Class) {
- case cByte:
- // ret sbyte, ubyte: Extend value into EAX and return
- if (isUnsigned)
- BuildMI (BB, X86::MOVZXr32r8, 1, X86::EAX).addReg (val);
- else
- BuildMI (BB, X86::MOVSXr32r8, 1, X86::EAX).addReg (val);
- break;
- case cShort:
- // ret short, ushort: Extend value into EAX and return
- if (isUnsigned)
- BuildMI (BB, X86::MOVZXr32r16, 1, X86::EAX).addReg (val);
- else
- BuildMI (BB, X86::MOVSXr32r16, 1, X86::EAX).addReg (val);
- break;
- case cInt:
- // ret int, uint, ptr: Move value into EAX and return
- // MOV EAX, <val>
- BuildMI(BB, X86::MOVrr32, 1, X86::EAX).addReg(val);
- break;
-
- // ret float/double: top of FP stack
- // FLD <val>
- case cFloat: // Floats
- BuildMI(BB, X86::FLDr4, 1).addReg(val);
- break;
- case cDouble: // Doubles
- BuildMI(BB, X86::FLDr8, 1).addReg(val);
- break;
- case cLong:
- // ret long: use EAX(least significant 32 bits)/EDX (most
- // significant 32)...uh, I think so Brain, but how do i call
- // up the two parts of the value from inside this mouse
- // cage? *zort*
- default:
- visitInstruction(I);
- }
-
+void
+ISel::visitReturnInst (ReturnInst &I)
+{
+ if (I.getNumOperands () == 0)
+ {
+ // Emit a 'ret' instruction
+ BuildMI (BB, X86::RET, 0);
+ return;
+ }
+ Value *rv = I.getOperand (0);
+ unsigned Class = getClass (rv->getType ());
+ switch (Class)
+ {
+ // integral return values: extend or move into EAX and return.
+ case cByte:
+ case cShort:
+ case cInt:
+ promote32 (X86::EAX, rv);
+ break;
+ // ret float/double: top of FP stack
+ // FLD <val>
+ case cFloat: // Floats
+ BuildMI (BB, X86::FLDr32, 1).addReg (getReg (rv));
+ break;
+ case cDouble: // Doubles
+ BuildMI (BB, X86::FLDr64, 1).addReg (getReg (rv));
+ break;
+ case cLong:
+ // ret long: use EAX(least significant 32 bits)/EDX (most
+ // significant 32)...uh, I think so Brain, but how do i call
+ // up the two parts of the value from inside this mouse
+ // cage? *zort*
+ default:
+ visitInstruction (I);
+ }
// Emit a 'ret' instruction
- BuildMI(BB, X86::RET, 0);
+ BuildMI (BB, X86::RET, 0);
}
/// visitBranchInst - Handle conditional and unconditional branches here. Note
}
}
+/// visitCallInst - Push args on stack and do a procedure call instruction.
void
-ISel::visitCallInst (CallInst &CI)
+ISel::visitCallInst (CallInst & CI)
{
- visitInstruction (CI);
+ // keep a counter of how many bytes we pushed on the stack
+ unsigned bytesPushed = 0;
+
+ // Push the arguments on the stack in reverse order, as specified by
+ // the ABI.
+ for (unsigned i = CI.getNumOperands()-1; i >= 1; --i)
+ {
+ Value *v = CI.getOperand (i);
+ switch (getClass (v->getType ()))
+ {
+ case cByte:
+ case cShort:
+ // Promote V to 32 bits wide, and move the result into EAX,
+ // then push EAX.
+ promote32 (X86::EAX, v);
+ BuildMI (BB, X86::PUSHr32, 1).addReg (X86::EAX);
+ bytesPushed += 4;
+ break;
+ case cInt:
+ case cFloat: {
+ unsigned Reg = getReg(v);
+ BuildMI (BB, X86::PUSHr32, 1).addReg(Reg);
+ bytesPushed += 4;
+ break;
+ }
+ default:
+ // FIXME: long/ulong/double args not handled.
+ visitInstruction (CI);
+ break;
+ }
+ }
+
+ if (Function *F = CI.getCalledFunction()) {
+ // Emit a CALL instruction with PC-relative displacement.
+ BuildMI(BB, X86::CALLpcrel32, 1).addPCDisp(F);
+ } else {
+ unsigned Reg = getReg(CI.getCalledValue());
+ BuildMI(BB, X86::CALLr32, 1).addReg(Reg);
+ }
+
+ // Adjust the stack by `bytesPushed' amount if non-zero
+ if (bytesPushed > 0)
+ BuildMI (BB, X86::ADDri32,2,X86::ESP).addReg(X86::ESP).addZImm(bytesPushed);
+
+ // If there is a return value, scavenge the result from the location the call
+ // leaves it in...
+ //
+ if (CI.getType() != Type::VoidTy) {
+ unsigned resultTypeClass = getClass (CI.getType ());
+ switch (resultTypeClass) {
+ case cByte:
+ case cShort:
+ case cInt: {
+ // Integral results are in %eax, or the appropriate portion
+ // thereof.
+ static const unsigned regRegMove[] = {
+ X86::MOVrr8, X86::MOVrr16, X86::MOVrr32
+ };
+ static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX };
+ BuildMI (BB, regRegMove[resultTypeClass], 1,
+ getReg (CI)).addReg (AReg[resultTypeClass]);
+ break;
+ }
+ case cFloat:
+ // Floating-point return values live in %st(0) (i.e., the top of
+ // the FP stack.) The general way to approach this is to do a
+ // FSTP to save the top of the FP stack on the real stack, then
+ // do a MOV to load the top of the real stack into the target
+ // register.
+ visitInstruction (CI); // FIXME: add the right args for the calls below
+ // BuildMI (BB, X86::FSTPm32, 0);
+ // BuildMI (BB, X86::MOVmr32, 0);
+ break;
+ default:
+ std::cerr << "Cannot get return value for call of type '"
+ << *CI.getType() << "'\n";
+ visitInstruction(CI);
+ }
+ }
}
/// visitSimpleBinary - Implement simple binary operators for integral types...
BuildMI(BB, Opcode, 2, getReg(B)).addReg(Op0r).addReg(Op1r);
}
-/// visitMul - Multiplies are not simple binary operators because they must deal
-/// with the EAX register explicitly.
-///
-void ISel::visitMul(BinaryOperator &I) {
- unsigned Class = getClass(I.getType());
- if (Class > 2) // FIXME: Handle longs
- visitInstruction(I);
-
+/// doMultiply - Emit appropriate instructions to multiply together
+/// the registers op0Reg and op1Reg, and put the result in destReg.
+/// The type of the result should be given as resultType.
+void ISel::doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator &MBBI,
+ unsigned destReg, const Type *resultType,
+ unsigned op0Reg, unsigned op1Reg) {
+ unsigned Class = getClass (resultType);
+
+ // FIXME:
+ assert (Class <= 2 && "Someday, we will learn how to multiply"
+ "longs and floating-point numbers. This is not that day.");
+
static const unsigned Regs[] ={ X86::AL , X86::AX , X86::EAX };
static const unsigned MulOpcode[]={ X86::MULrr8, X86::MULrr16, X86::MULrr32 };
static const unsigned MovOpcode[]={ X86::MOVrr8, X86::MOVrr16, X86::MOVrr32 };
-
unsigned Reg = Regs[Class];
- unsigned Op0Reg = getReg(I.getOperand(0));
- unsigned Op1Reg = getReg(I.getOperand(1));
- // Put the first operand into one of the A registers...
- BuildMI(BB, MovOpcode[Class], 1, Reg).addReg(Op0Reg);
+ // Emit a MOV to put the first operand into the appropriately-sized
+ // subreg of EAX.
+ BMI(MBB, MBBI, MovOpcode[Class], 1, Reg).addReg (op0Reg);
- // Emit the appropriate multiply instruction...
- BuildMI(BB, MulOpcode[Class], 1).addReg(Op1Reg);
+ // Emit the appropriate multiply instruction.
+ BMI(MBB, MBBI, MulOpcode[Class], 1).addReg (op1Reg);
- // Put the result into the destination register...
- BuildMI(BB, MovOpcode[Class], 1, getReg(I)).addReg(Reg);
+ // Emit another MOV to put the result into the destination register.
+ BMI(MBB, MBBI, MovOpcode[Class], 1, destReg).addReg (Reg);
+}
+
+/// visitMul - Multiplies are not simple binary operators because they must deal
+/// with the EAX register explicitly.
+///
+void ISel::visitMul(BinaryOperator &I) {
+ unsigned DestReg = getReg(I);
+ unsigned Op0Reg = getReg(I.getOperand(0));
+ unsigned Op1Reg = getReg(I.getOperand(1));
+ MachineBasicBlock::iterator MBBI = BB->end();
+ doMultiply(BB, MBBI, DestReg, I.getType(), Op0Reg, Op1Reg);
}
if (isSigned) {
// Emit a sign extension instruction...
- BuildMI(BB, ExtOpcode[Class], 1, ExtReg).addReg(Reg);
+ BuildMI(BB, ExtOpcode[Class], 0);
} else {
// If unsigned, emit a zeroing instruction... (reg = xor reg, reg)
BuildMI(BB, ClrOpcode[Class], 2, ExtReg).addReg(ExtReg).addReg(ExtReg);
}
-/// visitPHINode - Turn an LLVM PHI node into an X86 PHI node...
-///
-void ISel::visitPHINode(PHINode &PN) {
- MachineInstr *MI = BuildMI(BB, X86::PHI, PN.getNumOperands(), getReg(PN));
+/// visitCastInst - Here we have various kinds of copying with or without
+/// sign extension going on.
+void
+ISel::visitCastInst (CastInst &CI)
+{
+ const Type *targetType = CI.getType ();
+ Value *operand = CI.getOperand (0);
+ unsigned int operandReg = getReg (operand);
+ const Type *sourceType = operand->getType ();
+ unsigned int destReg = getReg (CI);
+ //
+ // Currently we handle:
+ //
+ // 1) cast * to bool
+ //
+ // 2) cast {sbyte, ubyte} to {sbyte, ubyte}
+ // cast {short, ushort} to {ushort, short}
+ // cast {int, uint, ptr} to {int, uint, ptr}
+ //
+ // 3) cast {sbyte, ubyte} to {ushort, short}
+ // cast {sbyte, ubyte} to {int, uint, ptr}
+ // cast {short, ushort} to {int, uint, ptr}
+ //
+ // 4) cast {int, uint, ptr} to {short, ushort}
+ // cast {int, uint, ptr} to {sbyte, ubyte}
+ // cast {short, ushort} to {sbyte, ubyte}
+
+ // 1) Implement casts to bool by using compare on the operand followed
+ // by set if not zero on the result.
+ if (targetType == Type::BoolTy)
+ {
+ BuildMI (BB, X86::CMPri8, 2).addReg (operandReg).addZImm (0);
+ BuildMI (BB, X86::SETNEr, 1, destReg);
+ return;
+ }
- for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
- // FIXME: This will put constants after the PHI nodes in the block, which
- // is invalid. They should be put inline into the PHI node eventually.
- //
- MI->addRegOperand(getReg(PN.getIncomingValue(i)));
- MI->addPCDispOperand(PN.getIncomingBlock(i));
+ // 2) Implement casts between values of the same type class (as determined
+ // by getClass) by using a register-to-register move.
+ unsigned srcClass = getClassB (sourceType);
+ unsigned targClass = getClass (targetType);
+ static const unsigned regRegMove[] = {
+ X86::MOVrr8, X86::MOVrr16, X86::MOVrr32
+ };
+ if ((srcClass < cLong) && (targClass < cLong) && (srcClass == targClass))
+ {
+ BuildMI (BB, regRegMove[srcClass], 1, destReg).addReg (operandReg);
+ return;
+ }
+ // 3) Handle cast of SMALLER int to LARGER int using a move with sign
+ // extension or zero extension, depending on whether the source type
+ // was signed.
+ if ((srcClass < cLong) && (targClass < cLong) && (srcClass < targClass))
+ {
+ static const unsigned ops[] = {
+ X86::MOVSXr16r8, X86::MOVSXr32r8, X86::MOVSXr32r16,
+ X86::MOVZXr16r8, X86::MOVZXr32r8, X86::MOVZXr32r16
+ };
+ unsigned srcSigned = sourceType->isSigned ();
+ BuildMI (BB, ops[3 * srcSigned + srcClass + targClass - 1], 1,
+ destReg).addReg (operandReg);
+ return;
+ }
+ // 4) Handle cast of LARGER int to SMALLER int using a move to EAX
+ // followed by a move out of AX or AL.
+ if ((srcClass < cLong) && (targClass < cLong) && (srcClass > targClass))
+ {
+ static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX };
+ BuildMI (BB, regRegMove[srcClass], 1,
+ AReg[srcClass]).addReg (operandReg);
+ BuildMI (BB, regRegMove[targClass], 1, destReg).addReg (AReg[srcClass]);
+ return;
+ }
+ // Anything we haven't handled already, we can't (yet) handle at all.
+ //
+ // FP to integral casts can be handled with FISTP to store onto the
+ // stack while converting to integer, followed by a MOV to load from
+ // the stack into the result register. Integral to FP casts can be
+ // handled with MOV to store onto the stack, followed by a FILD to
+ // load from the stack while converting to FP. For the moment, I
+ // can't quite get straight in my head how to borrow myself some
+ // stack space and write on it. Otherwise, this would be trivial.
+ visitInstruction (CI);
+}
+
+// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
+// returns zero when the input is not exactly a power of two.
+static unsigned ExactLog2(unsigned Val) {
+ if (Val == 0) return 0;
+ unsigned Count = 0;
+ while (Val != 1) {
+ if (Val & 1) return 0;
+ Val >>= 1;
+ ++Count;
}
+ return Count+1;
}
+/// visitGetElementPtrInst - I don't know, most programs don't have
+/// getelementptr instructions, right? That means we can put off
+/// implementing this, right? Right. This method emits machine
+/// instructions to perform type-safe pointer arithmetic. I am
+/// guessing this could be cleaned up somewhat to use fewer temporary
+/// registers.
void
-ISel::visitCastInst (CastInst &CI)
+ISel::visitGetElementPtrInst (GetElementPtrInst &I)
{
- visitInstruction (CI);
+ unsigned outputReg = getReg (I);
+ MachineBasicBlock::iterator MI = BB->end();
+ emitGEPOperation(BB, MI, I.getOperand(0),
+ I.op_begin()+1, I.op_end(), outputReg);
}
+void ISel::emitGEPOperation(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &IP,
+ Value *Src, User::op_iterator IdxBegin,
+ User::op_iterator IdxEnd, unsigned TargetReg) {
+ const TargetData &TD = TM.getTargetData();
+ const Type *Ty = Src->getType();
+ unsigned basePtrReg = getReg(Src, MBB, IP);
+
+ // GEPs have zero or more indices; we must perform a struct access
+ // or array access for each one.
+ for (GetElementPtrInst::op_iterator oi = IdxBegin,
+ oe = IdxEnd; oi != oe; ++oi) {
+ Value *idx = *oi;
+ unsigned nextBasePtrReg = makeAnotherReg(Type::UIntTy);
+ if (const StructType *StTy = dyn_cast <StructType> (Ty)) {
+ // It's a struct access. idx is the index into the structure,
+ // which names the field. This index must have ubyte type.
+ const ConstantUInt *CUI = cast <ConstantUInt> (idx);
+ assert (CUI->getType () == Type::UByteTy
+ && "Funny-looking structure index in GEP");
+ // Use the TargetData structure to pick out what the layout of
+ // the structure is in memory. Since the structure index must
+ // be constant, we can get its value and use it to find the
+ // right byte offset from the StructLayout class's list of
+ // structure member offsets.
+ unsigned idxValue = CUI->getValue ();
+ unsigned memberOffset =
+ TD.getStructLayout (StTy)->MemberOffsets[idxValue];
+ // Emit an ADD to add memberOffset to the basePtr.
+ BMI(MBB, IP, X86::ADDri32, 2,
+ nextBasePtrReg).addReg (basePtrReg).addZImm (memberOffset);
+ // The next type is the member of the structure selected by the
+ // index.
+ Ty = StTy->getElementTypes ()[idxValue];
+ } else if (const SequentialType *SqTy = cast <SequentialType>(Ty)) {
+ // It's an array or pointer access: [ArraySize x ElementType].
+
+ // idx is the index into the array. Unlike with structure
+ // indices, we may not know its actual value at code-generation
+ // time.
+ assert(idx->getType() == Type::LongTy && "Bad GEP array index!");
+
+ // We want to add basePtrReg to (idxReg * sizeof ElementType). First, we
+ // must find the size of the pointed-to type (Not coincidentally, the next
+ // type is the type of the elements in the array).
+ Ty = SqTy->getElementType();
+ unsigned elementSize = TD.getTypeSize(Ty);
+
+ // If idxReg is a constant, we don't need to perform the multiply!
+ if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(idx)) {
+ if (CSI->isNullValue()) {
+ BMI(MBB, IP, X86::MOVrr32, 1, nextBasePtrReg).addReg(basePtrReg);
+ } else {
+ unsigned Offset = elementSize*CSI->getValue();
+
+ BMI(MBB, IP, X86::ADDri32, 2,
+ nextBasePtrReg).addReg(basePtrReg).addZImm(Offset);
+ }
+ } else if (elementSize == 1) {
+ // If the element size is 1, we don't have to multiply, just add
+ unsigned idxReg = getReg(idx, MBB, IP);
+ BMI(MBB, IP, X86::ADDrr32, 2,
+ nextBasePtrReg).addReg(basePtrReg).addReg(idxReg);
+ } else {
+ unsigned idxReg = getReg(idx, MBB, IP);
+ unsigned OffsetReg = makeAnotherReg(Type::UIntTy);
+ if (unsigned Shift = ExactLog2(elementSize)) {
+ // If the element size is exactly a power of 2, use a shift to get it.
+
+ BMI(MBB, IP, X86::SHLir32, 2,
+ OffsetReg).addReg(idxReg).addZImm(Shift-1);
+ } else {
+ // Most general case, emit a multiply...
+ unsigned elementSizeReg = makeAnotherReg(Type::LongTy);
+ BMI(MBB, IP, X86::MOVir32, 1, elementSizeReg).addZImm(elementSize);
+
+ // Emit a MUL to multiply the register holding the index by
+ // elementSize, putting the result in OffsetReg.
+ doMultiply(MBB, IP, OffsetReg, Type::LongTy, idxReg, elementSizeReg);
+ }
+ // Emit an ADD to add OffsetReg to the basePtr.
+ BMI(MBB, IP, X86::ADDrr32, 2,
+ nextBasePtrReg).addReg (basePtrReg).addReg (OffsetReg);
+ }
+ }
+ // Now that we are here, further indices refer to subtypes of this
+ // one, so we don't need to worry about basePtrReg itself, anymore.
+ basePtrReg = nextBasePtrReg;
+ }
+ // After we have processed all the indices, the result is left in
+ // basePtrReg. Move it to the register where we were expected to
+ // put the answer. A 32-bit move should do it, because we are in
+ // ILP32 land.
+ BMI(MBB, IP, X86::MOVrr32, 1, TargetReg).addReg (basePtrReg);
+}
+
+
+/// visitMallocInst - I know that personally, whenever I want to remember
+/// something, I have to clear off some space in my brain.
+void
+ISel::visitMallocInst (MallocInst &I)
+{
+ // We assume that by this point, malloc instructions have been
+ // lowered to calls, and dlsym will magically find malloc for us.
+ // So we do not want to see malloc instructions here.
+ visitInstruction (I);
+}
+
+
+/// visitFreeInst - same story as MallocInst
+void
+ISel::visitFreeInst (FreeInst &I)
+{
+ // We assume that by this point, free instructions have been
+ // lowered to calls, and dlsym will magically find free for us.
+ // So we do not want to see free instructions here.
+ visitInstruction (I);
+}
+
+
+/// visitAllocaInst - I want some stack space. Come on, man, I said I
+/// want some freakin' stack space.
+void
+ISel::visitAllocaInst (AllocaInst &I)
+{
+ // Find the data size of the alloca inst's getAllocatedType.
+ const Type *allocatedType = I.getAllocatedType ();
+ const TargetData &TD = TM.DataLayout;
+ unsigned allocatedTypeSize = TD.getTypeSize (allocatedType);
+ // Keep stack 32-bit aligned.
+ unsigned int allocatedTypeWords = allocatedTypeSize / 4;
+ if (allocatedTypeSize % 4 != 0) { allocatedTypeWords++; }
+ // Subtract size from stack pointer, thereby allocating some space.
+ BuildMI(BB, X86::SUBri32, 2,
+ X86::ESP).addReg(X86::ESP).addZImm(allocatedTypeWords * 4);
+ // Put a pointer to the space into the result register, by copying
+ // the stack pointer.
+ BuildMI (BB, X86::MOVrr32, 1, getReg (I)).addReg (X86::ESP);
+}
+
+
/// createSimpleX86InstructionSelector - This pass converts an LLVM function
/// into a machine code representation is a very simple peep-hole fashion. The
/// generated code sucks but the implementation is nice and simple.