-//===-- FloatingPoint.cpp - Floating point Reg -> Stack converter ---------===//
+//===-- X86FloatingPoint.cpp - Floating point Reg -> Stack converter ------===//
//
// The LLVM Compiler Infrastructure
//
#include "llvm/CodeGen/Passes.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
-#include "llvm/Function.h" // FIXME: remove when using MBB CFG!
-#include "llvm/Support/CFG.h" // FIXME: remove when using MBB CFG!
-#include "Support/Debug.h"
-#include "Support/DepthFirstIterator.h"
-#include "Support/Statistic.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/ADT/DepthFirstIterator.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/STLExtras.h"
#include <algorithm>
#include <set>
using namespace llvm;
std::swap(Stack[RegMap[RegOnTop]], Stack[StackTop-1]);
// Emit an fxch to update the runtime processors version of the state
- MachineInstr *MI = BuildMI(X86::FXCH, 1).addReg(STReg);
- MBB->insert(I, MI);
+ BuildMI(*MBB, I, X86::FXCH, 1).addReg(STReg);
NumFXCH++;
}
}
- void duplicateToTop(unsigned RegNo, unsigned AsReg,
- MachineBasicBlock::iterator &I) {
+ void duplicateToTop(unsigned RegNo, unsigned AsReg, MachineInstr *I) {
unsigned STReg = getSTReg(RegNo);
pushReg(AsReg); // New register on top of stack
- MachineInstr *MI = BuildMI(X86::FLDrr, 1).addReg(STReg);
- MBB->insert(I, MI);
+ BuildMI(*MBB, I, X86::FLDrr, 1).addReg(STReg);
}
// popStackAfter - Pop the current value off of the top of the FP stack
// after the specified instruction.
void popStackAfter(MachineBasicBlock::iterator &I);
+ // freeStackSlotAfter - Free the specified register from the register stack,
+ // so that it is no longer in a register. If the register is currently at
+ // the top of the stack, we just pop the current instruction, otherwise we
+ // store the current top-of-stack into the specified slot, then pop the top
+ // of stack.
+ void freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned Reg);
+
bool processBasicBlock(MachineFunction &MF, MachineBasicBlock &MBB);
void handleZeroArgFP(MachineBasicBlock::iterator &I);
void handleOneArgFP(MachineBasicBlock::iterator &I);
void handleOneArgFPRW(MachineBasicBlock::iterator &I);
void handleTwoArgFP(MachineBasicBlock::iterator &I);
+ void handleCompareFP(MachineBasicBlock::iterator &I);
+ void handleCondMovFP(MachineBasicBlock::iterator &I);
void handleSpecialFP(MachineBasicBlock::iterator &I);
};
}
/// register references into FP stack references.
///
bool FPS::runOnMachineFunction(MachineFunction &MF) {
+ // We only need to run this pass if there are any FP registers used in this
+ // function. If it is all integer, there is nothing for us to do!
+ const bool *PhysRegsUsed = MF.getUsedPhysregs();
+ bool FPIsUsed = false;
+
+ assert(X86::FP6 == X86::FP0+6 && "Register enums aren't sorted right!");
+ for (unsigned i = 0; i <= 6; ++i)
+ if (PhysRegsUsed[X86::FP0+i]) {
+ FPIsUsed = true;
+ break;
+ }
+
+ // Early exit.
+ if (!FPIsUsed) return false;
+
LV = &getAnalysis<LiveVariables>();
StackTop = 0;
- // Figure out the mapping of MBB's to BB's.
- //
- // FIXME: Eventually we should be able to traverse the MBB CFG directly, and
- // we will need to extend this when one llvm basic block can codegen to
- // multiple MBBs.
- //
- // FIXME again: Just use the mapping established by LiveVariables!
- //
- std::map<const BasicBlock*, MachineBasicBlock *> MBBMap;
- for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
- MBBMap[I->getBasicBlock()] = I;
-
// Process the function in depth first order so that we process at least one
// of the predecessors for every reachable block in the function.
- std::set<const BasicBlock*> Processed;
- const BasicBlock *Entry = MF.getFunction()->begin();
+ std::set<MachineBasicBlock*> Processed;
+ MachineBasicBlock *Entry = MF.begin();
bool Changed = false;
- for (df_ext_iterator<const BasicBlock*, std::set<const BasicBlock*> >
+ for (df_ext_iterator<MachineBasicBlock*, std::set<MachineBasicBlock*> >
I = df_ext_begin(Entry, Processed), E = df_ext_end(Entry, Processed);
I != E; ++I)
- Changed |= processBasicBlock(MF, *MBBMap[*I]);
-
- assert(MBBMap.size() == Processed.size() &&
- "Doesn't handle unreachable code yet!");
+ Changed |= processBasicBlock(MF, **I);
return Changed;
}
/// transforming FP instructions into their stack form.
///
bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) {
- const TargetInstrInfo &TII = MF.getTarget().getInstrInfo();
+ const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
bool Changed = false;
MBB = &BB;
continue; // Efficiently ignore non-fp insts!
MachineInstr *PrevMI = 0;
- if (I != BB.begin()) {
- MachineBasicBlock::iterator tmp = I;
- --tmp;
- PrevMI = tmp;
- }
+ if (I != BB.begin())
+ PrevMI = prior(I);
++NumFP; // Keep track of # of pseudo instrs
DEBUG(std::cerr << "\nFPInst:\t";
- MI->print(std::cerr, MF.getTarget()));
+ MI->print(std::cerr, &(MF.getTarget())));
// Get dead variables list now because the MI pointer may be deleted as part
// of processing!
switch (Flags & X86II::FPTypeMask) {
case X86II::ZeroArgFP: handleZeroArgFP(I); break;
- case X86II::OneArgFP: handleOneArgFP(I); break; // fstp ST(0)
+ case X86II::OneArgFP: handleOneArgFP(I); break; // fstp ST(0)
case X86II::OneArgFPRW: handleOneArgFPRW(I); break; // ST(0) = fsqrt(ST(0))
- case X86II::TwoArgFP: handleTwoArgFP(I); break;
+ case X86II::TwoArgFP: handleTwoArgFP(I); break;
+ case X86II::CompareFP: handleCompareFP(I); break;
+ case X86II::CondMovFP: handleCondMovFP(I); break;
case X86II::SpecialFP: handleSpecialFP(I); break;
default: assert(0 && "Unknown FP Type!");
}
unsigned Reg = IB->second;
if (Reg >= X86::FP0 && Reg <= X86::FP6) {
DEBUG(std::cerr << "Register FP#" << Reg-X86::FP0 << " is dead!\n");
- ++I; // Insert fxch AFTER the instruction
- moveToTop(Reg-X86::FP0, I); // Insert fxch if necessary
- --I; // Move to fxch or old instruction
- popStackAfter(I); // Pop the top of the stack, killing value
+ freeStackSlotAfter(I, Reg-X86::FP0);
}
}
// Print out all of the instructions expanded to if -debug
- DEBUG(if (&*I == PrevMI) {
- std::cerr<< "Just deleted pseudo instruction\n";
- } else {
- MachineBasicBlock::iterator Start = I;
- // Rewind to first instruction newly inserted.
- while (Start != BB.begin() &&
- --Start != MachineBasicBlock::iterator(PrevMI));
- ++Start;
- std::cerr << "Inserted instructions:\n\t";
- Start->print(std::cerr, MF.getTarget());
- while (++Start != I); ++Start;
- }
- dumpStack();
- );
+ DEBUG(
+ MachineBasicBlock::iterator PrevI(PrevMI);
+ if (I == PrevI) {
+ std::cerr << "Just deleted pseudo instruction\n";
+ } else {
+ MachineBasicBlock::iterator Start = I;
+ // Rewind to first instruction newly inserted.
+ while (Start != BB.begin() && prior(Start) != PrevI) --Start;
+ std::cerr << "Inserted instructions:\n\t";
+ Start->print(std::cerr, &MF.getTarget());
+ while (++Start != next(I));
+ }
+ dumpStack();
+ );
Changed = true;
}
{ X86::FDIVRrST0, X86::FDIVRPrST0 },
{ X86::FDIVrST0 , X86::FDIVPrST0 },
- { X86::FISTr16 , X86::FISTPr16 },
- { X86::FISTr32 , X86::FISTPr32 },
+ { X86::FIST16m , X86::FISTP16m },
+ { X86::FIST32m , X86::FISTP32m },
{ X86::FMULrST0 , X86::FMULPrST0 },
- { X86::FSTr32 , X86::FSTPr32 },
- { X86::FSTr64 , X86::FSTPr64 },
+ { X86::FST32m , X86::FSTP32m },
+ { X86::FST64m , X86::FSTP64m },
{ X86::FSTrr , X86::FSTPrr },
{ X86::FSUBRrST0, X86::FSUBRPrST0 },
{ X86::FSUBrST0 , X86::FSUBPrST0 },
+ { X86::FUCOMIr , X86::FUCOMIPr },
+
{ X86::FUCOMPr , X86::FUCOMPPr },
{ X86::FUCOMr , X86::FUCOMPr },
};
I->RemoveOperand(0);
} else { // Insert an explicit pop
- MachineInstr *MI = BuildMI(X86::FSTPrr, 1).addReg(X86::ST0);
- I = MBB->insert(++I, MI);
+ I = BuildMI(*MBB, ++I, X86::FSTPrr, 1).addReg(X86::ST0);
}
}
+/// freeStackSlotAfter - Free the specified register from the register stack, so
+/// that it is no longer in a register. If the register is currently at the top
+/// of the stack, we just pop the current instruction, otherwise we store the
+/// current top-of-stack into the specified slot, then pop the top of stack.
+void FPS::freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned FPRegNo) {
+ if (getStackEntry(0) == FPRegNo) { // already at the top of stack? easy.
+ popStackAfter(I);
+ return;
+ }
+
+ // Otherwise, store the top of stack into the dead slot, killing the operand
+ // without having to add in an explicit xchg then pop.
+ //
+ unsigned STReg = getSTReg(FPRegNo);
+ unsigned OldSlot = getSlot(FPRegNo);
+ unsigned TopReg = Stack[StackTop-1];
+ Stack[OldSlot] = TopReg;
+ RegMap[TopReg] = OldSlot;
+ RegMap[FPRegNo] = ~0;
+ Stack[--StackTop] = ~0;
+ I = BuildMI(*MBB, ++I, X86::FSTPrr, 1).addReg(STReg);
+}
+
+
static unsigned getFPReg(const MachineOperand &MO) {
assert(MO.isRegister() && "Expected an FP register!");
unsigned Reg = MO.getReg();
E = LV->killed_end(MI); KI != E; ++KI)
KillsSrc |= KI->second == X86::FP0+Reg;
- // FSTPr80 and FISTPr64 are strange because there are no non-popping versions.
+ // FSTP80r and FISTP64r are strange because there are no non-popping versions.
// If we have one _and_ we don't want to pop the operand, duplicate the value
// on the stack instead of moving it. This ensure that popping the value is
// always ok.
//
- if ((MI->getOpcode() == X86::FSTPr80 ||
- MI->getOpcode() == X86::FISTPr64) && !KillsSrc) {
+ if ((MI->getOpcode() == X86::FSTP80m ||
+ MI->getOpcode() == X86::FISTP64m) && !KillsSrc) {
duplicateToTop(Reg, 7 /*temp register*/, I);
} else {
moveToTop(Reg, I); // Move to the top of the stack...
}
MI->RemoveOperand(MI->getNumOperands()-1); // Remove explicit ST(0) operand
- if (MI->getOpcode() == X86::FSTPr80 || MI->getOpcode() == X86::FISTPr64) {
+ if (MI->getOpcode() == X86::FSTP80m || MI->getOpcode() == X86::FISTP64m) {
assert(StackTop > 0 && "Stack empty??");
--StackTop;
} else if (KillsSrc) { // Last use of operand?
}
-/// handleOneArgFPRW - fchs - ST(0) = -ST(0)
+/// handleOneArgFPRW: Handle instructions that read from the top of stack and
+/// replace the value with a newly computed value. These instructions may have
+/// non-fp operands after their FP operands.
+///
+/// Examples:
+/// R1 = fchs R2
+/// R1 = fadd R2, [mem]
///
void FPS::handleOneArgFPRW(MachineBasicBlock::iterator &I) {
MachineInstr *MI = I;
- assert(MI->getNumOperands() == 2 && "Can only handle fst* instructions!");
+ assert(MI->getNumOperands() >= 2 && "FPRW instructions must have 2 ops!!");
// Is this the last use of the source register?
unsigned Reg = getFPReg(MI->getOperand(1));
// ForwardST0Table - Map: A = B op C into: ST(0) = ST(0) op ST(i)
static const TableEntry ForwardST0Table[] = {
- { X86::FpADD, X86::FADDST0r },
- { X86::FpDIV, X86::FDIVST0r },
- { X86::FpMUL, X86::FMULST0r },
- { X86::FpSUB, X86::FSUBST0r },
- { X86::FpUCOM, X86::FUCOMr },
+ { X86::FpADD , X86::FADDST0r },
+ { X86::FpDIV , X86::FDIVST0r },
+ { X86::FpMUL , X86::FMULST0r },
+ { X86::FpSUB , X86::FSUBST0r },
};
// ReverseST0Table - Map: A = B op C into: ST(0) = ST(i) op ST(0)
static const TableEntry ReverseST0Table[] = {
- { X86::FpADD, X86::FADDST0r }, // commutative
- { X86::FpDIV, X86::FDIVRST0r },
- { X86::FpMUL, X86::FMULST0r }, // commutative
- { X86::FpSUB, X86::FSUBRST0r },
- { X86::FpUCOM, ~0 },
+ { X86::FpADD , X86::FADDST0r }, // commutative
+ { X86::FpDIV , X86::FDIVRST0r },
+ { X86::FpMUL , X86::FMULST0r }, // commutative
+ { X86::FpSUB , X86::FSUBRST0r },
};
// ForwardSTiTable - Map: A = B op C into: ST(i) = ST(0) op ST(i)
static const TableEntry ForwardSTiTable[] = {
- { X86::FpADD, X86::FADDrST0 }, // commutative
- { X86::FpDIV, X86::FDIVRrST0 },
- { X86::FpMUL, X86::FMULrST0 }, // commutative
- { X86::FpSUB, X86::FSUBRrST0 },
- { X86::FpUCOM, X86::FUCOMr },
+ { X86::FpADD , X86::FADDrST0 }, // commutative
+ { X86::FpDIV , X86::FDIVRrST0 },
+ { X86::FpMUL , X86::FMULrST0 }, // commutative
+ { X86::FpSUB , X86::FSUBRrST0 },
};
// ReverseSTiTable - Map: A = B op C into: ST(i) = ST(i) op ST(0)
static const TableEntry ReverseSTiTable[] = {
- { X86::FpADD, X86::FADDrST0 },
- { X86::FpDIV, X86::FDIVrST0 },
- { X86::FpMUL, X86::FMULrST0 },
- { X86::FpSUB, X86::FSUBrST0 },
- { X86::FpUCOM, ~0 },
+ { X86::FpADD , X86::FADDrST0 },
+ { X86::FpDIV , X86::FDIVrST0 },
+ { X86::FpMUL , X86::FMULrST0 },
+ { X86::FpSUB , X86::FSUBrST0 },
};
/// ST(i) = fsub ST(0), ST(i)
/// ST(0) = fsubr ST(0), ST(i)
/// ST(i) = fsubr ST(0), ST(i)
-///
-/// In addition to three address instructions, this also handles the FpUCOM
-/// instruction which only has two operands, but no destination. This
-/// instruction is also annoying because there is no "reverse" form of it
-/// available.
///
void FPS::handleTwoArgFP(MachineBasicBlock::iterator &I) {
ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
MachineInstr *MI = I;
unsigned NumOperands = MI->getNumOperands();
- assert(NumOperands == 3 ||
- (NumOperands == 2 && MI->getOpcode() == X86::FpUCOM) &&
- "Illegal TwoArgFP instruction!");
+ assert(NumOperands == 3 && "Illegal TwoArgFP instruction!");
unsigned Dest = getFPReg(MI->getOperand(0));
unsigned Op0 = getFPReg(MI->getOperand(NumOperands-2));
unsigned Op1 = getFPReg(MI->getOperand(NumOperands-1));
KillsOp1 |= (KI->second == X86::FP0+Op1);
}
- // If this is an FpUCOM instruction, we must make sure the first operand is on
- // the top of stack, the other one can be anywhere...
- if (MI->getOpcode() == X86::FpUCOM)
- moveToTop(Op0, I);
-
unsigned TOS = getStackEntry(0);
// One of our operands must be on the top of the stack. If neither is yet, we
Op0 = TOS = Dest;
KillsOp0 = true;
}
- } else if (!KillsOp0 && !KillsOp1 && MI->getOpcode() != X86::FpUCOM) {
+ } else if (!KillsOp0 && !KillsOp1) {
// If we DO have one of our operands at the top of the stack, but we don't
// have a dead operand, we must duplicate one of the operands to a new slot
// on the stack.
// Now we know that one of our operands is on the top of the stack, and at
// least one of our operands is killed by this instruction.
- assert((TOS == Op0 || TOS == Op1) &&
- (KillsOp0 || KillsOp1 || MI->getOpcode() == X86::FpUCOM) &&
+ assert((TOS == Op0 || TOS == Op1) && (KillsOp0 || KillsOp1) &&
"Stack conditions not set up right!");
// We decide which form to use based on what is on the top of the stack, and
unsigned NotTOS = (TOS == Op0) ? Op1 : Op0;
// Replace the old instruction with a new instruction
- MBB->remove(I);
- I = MBB->insert(I, BuildMI(Opcode, 1).addReg(getSTReg(NotTOS)));
+ MBB->remove(I++);
+ I = BuildMI(*MBB, I, Opcode, 1).addReg(getSTReg(NotTOS));
// If both operands are killed, pop one off of the stack in addition to
// overwriting the other one.
popStackAfter(I); // Pop the top of stack
}
- // Insert an explicit pop of the "updated" operand for FUCOM
- if (MI->getOpcode() == X86::FpUCOM) {
- if (KillsOp0 && !KillsOp1)
- popStackAfter(I); // If we kill the first operand, pop it!
- else if (KillsOp1 && Op0 != Op1) {
- if (getStackEntry(0) == Op1) {
- popStackAfter(I); // If it's right at the top of stack, just pop it
- } else {
- // Otherwise, move the top of stack into the dead slot, killing the
- // operand without having to add in an explicit xchg then pop.
- //
- unsigned STReg = getSTReg(Op1);
- unsigned OldSlot = getSlot(Op1);
- unsigned TopReg = Stack[StackTop-1];
- Stack[OldSlot] = TopReg;
- RegMap[TopReg] = OldSlot;
- RegMap[Op1] = ~0;
- Stack[--StackTop] = ~0;
-
- MachineInstr *MI = BuildMI(X86::FSTPrr, 1).addReg(STReg);
- I = MBB->insert(++I, MI);
- }
- }
- }
-
// Update stack information so that we know the destination register is now on
// the stack.
- if (MI->getOpcode() != X86::FpUCOM) {
- unsigned UpdatedSlot = getSlot(updateST0 ? TOS : NotTOS);
- assert(UpdatedSlot < StackTop && Dest < 7);
- Stack[UpdatedSlot] = Dest;
- RegMap[Dest] = UpdatedSlot;
- }
+ unsigned UpdatedSlot = getSlot(updateST0 ? TOS : NotTOS);
+ assert(UpdatedSlot < StackTop && Dest < 7);
+ Stack[UpdatedSlot] = Dest;
+ RegMap[Dest] = UpdatedSlot;
delete MI; // Remove the old instruction
}
+/// handleCompareFP - Handle FUCOM and FUCOMI instructions, which have two FP
+/// register arguments and no explicit destinations.
+///
+void FPS::handleCompareFP(MachineBasicBlock::iterator &I) {
+ ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
+ ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable);
+ MachineInstr *MI = I;
+
+ unsigned NumOperands = MI->getNumOperands();
+ assert(NumOperands == 2 && "Illegal FUCOM* instruction!");
+ unsigned Op0 = getFPReg(MI->getOperand(NumOperands-2));
+ unsigned Op1 = getFPReg(MI->getOperand(NumOperands-1));
+ bool KillsOp0 = false, KillsOp1 = false;
+
+ for (LiveVariables::killed_iterator KI = LV->killed_begin(MI),
+ E = LV->killed_end(MI); KI != E; ++KI) {
+ KillsOp0 |= (KI->second == X86::FP0+Op0);
+ KillsOp1 |= (KI->second == X86::FP0+Op1);
+ }
+
+ // Make sure the first operand is on the top of stack, the other one can be
+ // anywhere.
+ moveToTop(Op0, I);
+
+ MI->getOperand(0).setReg(getSTReg(Op1));
+ MI->RemoveOperand(1);
+
+ // If any of the operands are killed by this instruction, free them.
+ if (KillsOp0) freeStackSlotAfter(I, Op0);
+ if (KillsOp1 && Op0 != Op1) freeStackSlotAfter(I, Op1);
+}
+
+/// handleCondMovFP - Handle two address conditional move instructions. These
+/// instructions move a st(i) register to st(0) iff a condition is true. These
+/// instructions require that the first operand is at the top of the stack, but
+/// otherwise don't modify the stack at all.
+void FPS::handleCondMovFP(MachineBasicBlock::iterator &I) {
+ MachineInstr *MI = I;
+
+ unsigned Op0 = getFPReg(MI->getOperand(0));
+ unsigned Op1 = getFPReg(MI->getOperand(1));
+
+ // The first operand *must* be on the top of the stack.
+ moveToTop(Op0, I);
+
+ // Change the second operand to the stack register that the operand is in.
+ MI->RemoveOperand(0);
+ MI->getOperand(0).setReg(getSTReg(Op1));
+
+ // If we kill the second operand, make sure to pop it from the stack.
+ if (Op0 != Op1)
+ for (LiveVariables::killed_iterator KI = LV->killed_begin(MI),
+ E = LV->killed_end(MI); KI != E; ++KI)
+ if (KI->second == X86::FP0+Op1) {
+ // Get this value off of the register stack.
+ freeStackSlotAfter(I, Op1);
+ break;
+ }
+}
+
/// handleSpecialFP - Handle special instructions which behave unlike other
/// floating point instructions. This is primarily intended for use by pseudo