//===----------------------------------------------------------------------===//
#include "llvm/Target/TargetLowering.h"
+#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/MRegisterInfo.h"
+#include "llvm/DerivedTypes.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/Support/MathExtras.h"
TargetLowering::TargetLowering(TargetMachine &tm)
: TM(tm), TD(TM.getTargetData()) {
- assert(ISD::BUILTIN_OP_END <= 128 &&
+ assert(ISD::BUILTIN_OP_END <= 156 &&
"Fixed size array in TargetLowering is not large enough!");
// All operations default to being supported.
memset(OpActions, 0, sizeof(OpActions));
- IsLittleEndian = TD.isLittleEndian();
- ShiftAmountTy = SetCCResultTy = PointerTy = getValueType(TD.getIntPtrType());
+ IsLittleEndian = TD->isLittleEndian();
+ ShiftAmountTy = SetCCResultTy = PointerTy = getValueType(TD->getIntPtrType());
ShiftAmtHandling = Undefined;
memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
- maxStoresPerMemSet = maxStoresPerMemCpy = maxStoresPerMemMove = 8;
+ memset(TargetDAGCombineArray, 0,
+ sizeof(TargetDAGCombineArray)/sizeof(TargetDAGCombineArray[0]));
+ maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
allowUnalignedMemoryAccesses = false;
UseUnderscoreSetJmpLongJmp = false;
IntDivIsCheap = false;
// Set MVT::Vector to always be Expanded
SetValueTypeAction(MVT::Vector, Expand, *this, TransformToType,
ValueTypeActions);
+
+ // Loop over all of the legal vector value types, specifying an identity type
+ // transformation.
+ for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
+ i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
+ if (isTypeLegal((MVT::ValueType)i))
+ TransformToType[i] = (MVT::ValueType)i;
+ }
assert(isTypeLegal(MVT::f64) && "Target does not support FP?");
TransformToType[MVT::f64] = MVT::f64;
return NULL;
}
+/// getPackedTypeBreakdown - Packed types are broken down into some number of
+/// legal scalar types. For example, <8 x float> maps to 2 MVT::v2f32 values
+/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
+///
+/// This method returns the number and type of the resultant breakdown.
+///
+unsigned TargetLowering::getPackedTypeBreakdown(const PackedType *PTy,
+ MVT::ValueType &PTyElementVT,
+ MVT::ValueType &PTyLegalElementVT) const {
+ // Figure out the right, legal destination reg to copy into.
+ unsigned NumElts = PTy->getNumElements();
+ MVT::ValueType EltTy = getValueType(PTy->getElementType());
+
+ unsigned NumVectorRegs = 1;
+
+ // Divide the input until we get to a supported size. This will always
+ // end with a scalar if the target doesn't support vectors.
+ while (NumElts > 1 && !isTypeLegal(getVectorType(EltTy, NumElts))) {
+ NumElts >>= 1;
+ NumVectorRegs <<= 1;
+ }
+
+ MVT::ValueType VT;
+ if (NumElts == 1) {
+ VT = EltTy;
+ } else {
+ VT = getVectorType(EltTy, NumElts);
+ }
+ PTyElementVT = VT;
-
-/// MaskedValueIsZero - Return true if 'Op & Mask' is known to be zero. We use
-/// this predicate to simplify operations downstream. Op and Mask are known to
-/// be the same type.
-bool TargetLowering::MaskedValueIsZero(const SDOperand &Op,
- uint64_t Mask) const {
- unsigned SrcBits;
- if (Mask == 0) return true;
+ MVT::ValueType DestVT = getTypeToTransformTo(VT);
+ PTyLegalElementVT = DestVT;
+ if (DestVT < VT) {
+ // Value is expanded, e.g. i64 -> i16.
+ return NumVectorRegs*(MVT::getSizeInBits(VT)/MVT::getSizeInBits(DestVT));
+ } else {
+ // Otherwise, promotion or legal types use the same number of registers as
+ // the vector decimated to the appropriate level.
+ return NumVectorRegs;
+ }
- // If we know the result of a setcc has the top bits zero, use this info.
+ return DestVT;
+}
+
+//===----------------------------------------------------------------------===//
+// Optimization Methods
+//===----------------------------------------------------------------------===//
+
+/// ShrinkDemandedConstant - Check to see if the specified operand of the
+/// specified instruction is a constant integer. If so, check to see if there
+/// are any bits set in the constant that are not demanded. If so, shrink the
+/// constant and return true.
+bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDOperand Op,
+ uint64_t Demanded) {
+ // FIXME: ISD::SELECT, ISD::SELECT_CC
+ switch(Op.getOpcode()) {
+ default: break;
+ case ISD::AND:
+ case ISD::OR:
+ case ISD::XOR:
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
+ if ((~Demanded & C->getValue()) != 0) {
+ MVT::ValueType VT = Op.getValueType();
+ SDOperand New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
+ DAG.getConstant(Demanded & C->getValue(),
+ VT));
+ return CombineTo(Op, New);
+ }
+ break;
+ }
+ return false;
+}
+
+/// SimplifyDemandedBits - Look at Op. At this point, we know that only the
+/// DemandedMask bits of the result of Op are ever used downstream. If we can
+/// use this information to simplify Op, create a new simplified DAG node and
+/// return true, returning the original and new nodes in Old and New. Otherwise,
+/// analyze the expression and return a mask of KnownOne and KnownZero bits for
+/// the expression (used to simplify the caller). The KnownZero/One bits may
+/// only be accurate for those bits in the DemandedMask.
+bool TargetLowering::SimplifyDemandedBits(SDOperand Op, uint64_t DemandedMask,
+ uint64_t &KnownZero,
+ uint64_t &KnownOne,
+ TargetLoweringOpt &TLO,
+ unsigned Depth) const {
+ KnownZero = KnownOne = 0; // Don't know anything.
+ // Other users may use these bits.
+ if (!Op.Val->hasOneUse()) {
+ if (Depth != 0) {
+ // If not at the root, Just compute the KnownZero/KnownOne bits to
+ // simplify things downstream.
+ ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
+ return false;
+ }
+ // If this is the root being simplified, allow it to have multiple uses,
+ // just set the DemandedMask to all bits.
+ DemandedMask = MVT::getIntVTBitMask(Op.getValueType());
+ } else if (DemandedMask == 0) {
+ // Not demanding any bits from Op.
+ if (Op.getOpcode() != ISD::UNDEF)
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
+ return false;
+ } else if (Depth == 6) { // Limit search depth.
+ return false;
+ }
+
+ uint64_t KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
switch (Op.getOpcode()) {
case ISD::Constant:
- return (cast<ConstantSDNode>(Op)->getValue() & Mask) == 0;
- case ISD::SETCC:
- return ((Mask & 1) == 0) &&
- getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult;
- case ISD::ZEXTLOAD:
- SrcBits = MVT::getSizeInBits(cast<VTSDNode>(Op.getOperand(3))->getVT());
- return (Mask & ((1ULL << SrcBits)-1)) == 0; // Returning only the zext bits.
- case ISD::ZERO_EXTEND:
- SrcBits = MVT::getSizeInBits(Op.getOperand(0).getValueType());
- return MaskedValueIsZero(Op.getOperand(0),Mask & (~0ULL >> (64-SrcBits)));
- case ISD::AssertZext:
- SrcBits = MVT::getSizeInBits(cast<VTSDNode>(Op.getOperand(1))->getVT());
- return (Mask & ((1ULL << SrcBits)-1)) == 0; // Returning only the zext bits.
+ // We know all of the bits for a constant!
+ KnownOne = cast<ConstantSDNode>(Op)->getValue() & DemandedMask;
+ KnownZero = ~KnownOne & DemandedMask;
+ return false; // Don't fall through, will infinitely loop.
case ISD::AND:
- // If either of the operands has zero bits, the result will too.
- if (MaskedValueIsZero(Op.getOperand(1), Mask) ||
- MaskedValueIsZero(Op.getOperand(0), Mask))
+ // If the RHS is a constant, check to see if the LHS would be zero without
+ // using the bits from the RHS. Below, we use knowledge about the RHS to
+ // simplify the LHS, here we're using information from the LHS to simplify
+ // the RHS.
+ if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ uint64_t LHSZero, LHSOne;
+ ComputeMaskedBits(Op.getOperand(0), DemandedMask,
+ LHSZero, LHSOne, Depth+1);
+ // If the LHS already has zeros where RHSC does, this and is dead.
+ if ((LHSZero & DemandedMask) == (~RHSC->getValue() & DemandedMask))
+ return TLO.CombineTo(Op, Op.getOperand(0));
+ // If any of the set bits in the RHS are known zero on the LHS, shrink
+ // the constant.
+ if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & DemandedMask))
+ return true;
+ }
+
+ if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
+ KnownOne, TLO, Depth+1))
return true;
- // (X & C1) & C2 == 0 iff C1 & C2 == 0.
- if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
- return MaskedValueIsZero(Op.getOperand(0),AndRHS->getValue() & Mask);
- return false;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & ~KnownZero,
+ KnownZero2, KnownOne2, TLO, Depth+1))
+ return true;
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // If all of the demanded bits are known one on one side, return the other.
+ // These bits cannot contribute to the result of the 'and'.
+ if ((DemandedMask & ~KnownZero2 & KnownOne)==(DemandedMask & ~KnownZero2))
+ return TLO.CombineTo(Op, Op.getOperand(0));
+ if ((DemandedMask & ~KnownZero & KnownOne2)==(DemandedMask & ~KnownZero))
+ return TLO.CombineTo(Op, Op.getOperand(1));
+ // If all of the demanded bits in the inputs are known zeros, return zero.
+ if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
+ return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
+ // If the RHS is a constant, see if we can simplify it.
+ if (TLO.ShrinkDemandedConstant(Op, DemandedMask & ~KnownZero2))
+ return true;
+
+ // Output known-1 bits are only known if set in both the LHS & RHS.
+ KnownOne &= KnownOne2;
+ // Output known-0 are known to be clear if zero in either the LHS | RHS.
+ KnownZero |= KnownZero2;
+ break;
case ISD::OR:
+ if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
+ KnownOne, TLO, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & ~KnownOne,
+ KnownZero2, KnownOne2, TLO, Depth+1))
+ return true;
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // If all of the demanded bits are known zero on one side, return the other.
+ // These bits cannot contribute to the result of the 'or'.
+ if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
+ return TLO.CombineTo(Op, Op.getOperand(0));
+ if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
+ return TLO.CombineTo(Op, Op.getOperand(1));
+ // If all of the potentially set bits on one side are known to be set on
+ // the other side, just use the 'other' side.
+ if ((DemandedMask & (~KnownZero) & KnownOne2) ==
+ (DemandedMask & (~KnownZero)))
+ return TLO.CombineTo(Op, Op.getOperand(0));
+ if ((DemandedMask & (~KnownZero2) & KnownOne) ==
+ (DemandedMask & (~KnownZero2)))
+ return TLO.CombineTo(Op, Op.getOperand(1));
+ // If the RHS is a constant, see if we can simplify it.
+ if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
+ return true;
+
+ // Output known-0 bits are only known if clear in both the LHS & RHS.
+ KnownZero &= KnownZero2;
+ // Output known-1 are known to be set if set in either the LHS | RHS.
+ KnownOne |= KnownOne2;
+ break;
case ISD::XOR:
- return MaskedValueIsZero(Op.getOperand(0), Mask) &&
- MaskedValueIsZero(Op.getOperand(1), Mask);
+ if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
+ KnownOne, TLO, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask, KnownZero2,
+ KnownOne2, TLO, Depth+1))
+ return true;
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // If all of the demanded bits are known zero on one side, return the other.
+ // These bits cannot contribute to the result of the 'xor'.
+ if ((DemandedMask & KnownZero) == DemandedMask)
+ return TLO.CombineTo(Op, Op.getOperand(0));
+ if ((DemandedMask & KnownZero2) == DemandedMask)
+ return TLO.CombineTo(Op, Op.getOperand(1));
+
+ // Output known-0 bits are known if clear or set in both the LHS & RHS.
+ KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
+ // Output known-1 are known to be set if set in only one of the LHS, RHS.
+ KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
+
+ // If all of the unknown bits are known to be zero on one side or the other
+ // (but not both) turn this into an *inclusive* or.
+ // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
+ if (uint64_t UnknownBits = DemandedMask & ~(KnownZeroOut|KnownOneOut))
+ if ((UnknownBits & (KnownZero|KnownZero2)) == UnknownBits)
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
+ Op.getOperand(0),
+ Op.getOperand(1)));
+ // If all of the demanded bits on one side are known, and all of the set
+ // bits on that side are also known to be set on the other side, turn this
+ // into an AND, as we know the bits will be cleared.
+ // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
+ if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
+ if ((KnownOne & KnownOne2) == KnownOne) {
+ MVT::ValueType VT = Op.getValueType();
+ SDOperand ANDC = TLO.DAG.getConstant(~KnownOne & DemandedMask, VT);
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
+ ANDC));
+ }
+ }
+
+ // If the RHS is a constant, see if we can simplify it.
+ // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
+ if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
+ return true;
+
+ KnownZero = KnownZeroOut;
+ KnownOne = KnownOneOut;
+ break;
+ case ISD::SETCC:
+ // If we know the result of a setcc has the top bits zero, use this info.
+ if (getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult)
+ KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
+ break;
case ISD::SELECT:
- return MaskedValueIsZero(Op.getOperand(1), Mask) &&
- MaskedValueIsZero(Op.getOperand(2), Mask);
+ if (SimplifyDemandedBits(Op.getOperand(2), DemandedMask, KnownZero,
+ KnownOne, TLO, Depth+1))
+ return true;
+ if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero2,
+ KnownOne2, TLO, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // If the operands are constants, see if we can simplify them.
+ if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
+ return true;
+
+ // Only known if known in both the LHS and RHS.
+ KnownOne &= KnownOne2;
+ KnownZero &= KnownZero2;
+ break;
case ISD::SELECT_CC:
- return MaskedValueIsZero(Op.getOperand(2), Mask) &&
- MaskedValueIsZero(Op.getOperand(3), Mask);
- case ISD::SRL:
- // (ushr X, C1) & C2 == 0 iff X & (C2 << C1) == 0
- if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
- uint64_t NewVal = Mask << ShAmt->getValue();
- SrcBits = MVT::getSizeInBits(Op.getValueType());
- if (SrcBits != 64) NewVal &= (1ULL << SrcBits)-1;
- return MaskedValueIsZero(Op.getOperand(0), NewVal);
- }
- return false;
+ if (SimplifyDemandedBits(Op.getOperand(3), DemandedMask, KnownZero,
+ KnownOne, TLO, Depth+1))
+ return true;
+ if (SimplifyDemandedBits(Op.getOperand(2), DemandedMask, KnownZero2,
+ KnownOne2, TLO, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // If the operands are constants, see if we can simplify them.
+ if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
+ return true;
+
+ // Only known if known in both the LHS and RHS.
+ KnownOne &= KnownOne2;
+ KnownZero &= KnownZero2;
+ break;
case ISD::SHL:
- // (ushl X, C1) & C2 == 0 iff X & (C2 >> C1) == 0
- if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
- uint64_t NewVal = Mask >> ShAmt->getValue();
- return MaskedValueIsZero(Op.getOperand(0), NewVal);
+ if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask >> SA->getValue(),
+ KnownZero, KnownOne, TLO, Depth+1))
+ return true;
+ KnownZero <<= SA->getValue();
+ KnownOne <<= SA->getValue();
+ KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
}
- return false;
- case ISD::ADD:
- // (add X, Y) & C == 0 iff (X&C)|(Y&C) == 0 and all bits are low bits.
- if ((Mask&(Mask+1)) == 0) { // All low bits
- if (MaskedValueIsZero(Op.getOperand(0), Mask) &&
- MaskedValueIsZero(Op.getOperand(1), Mask))
+ break;
+ case ISD::SRL:
+ if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ MVT::ValueType VT = Op.getValueType();
+ unsigned ShAmt = SA->getValue();
+
+ // Compute the new bits that are at the top now.
+ uint64_t HighBits = (1ULL << ShAmt)-1;
+ HighBits <<= MVT::getSizeInBits(VT) - ShAmt;
+ uint64_t TypeMask = MVT::getIntVTBitMask(VT);
+
+ if (SimplifyDemandedBits(Op.getOperand(0),
+ (DemandedMask << ShAmt) & TypeMask,
+ KnownZero, KnownOne, TLO, Depth+1))
return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero &= TypeMask;
+ KnownOne &= TypeMask;
+ KnownZero >>= ShAmt;
+ KnownOne >>= ShAmt;
+ KnownZero |= HighBits; // high bits known zero.
}
break;
- case ISD::SUB:
- if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) {
- // We know that the top bits of C-X are clear if X contains less bits
- // than C (i.e. no wrap-around can happen). For example, 20-X is
- // positive if we can prove that X is >= 0 and < 16.
- unsigned Bits = MVT::getSizeInBits(CLHS->getValueType(0));
- if ((CLHS->getValue() & (1 << (Bits-1))) == 0) { // sign bit clear
- unsigned NLZ = CountLeadingZeros_64(CLHS->getValue()+1);
- uint64_t MaskV = (1ULL << (63-NLZ))-1;
- if (MaskedValueIsZero(Op.getOperand(1), ~MaskV)) {
- // High bits are clear this value is known to be >= C.
- unsigned NLZ2 = CountLeadingZeros_64(CLHS->getValue());
- if ((Mask & ((1ULL << (64-NLZ2))-1)) == 0)
- return true;
+ case ISD::SRA:
+ if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ MVT::ValueType VT = Op.getValueType();
+ unsigned ShAmt = SA->getValue();
+
+ // Compute the new bits that are at the top now.
+ uint64_t HighBits = (1ULL << ShAmt)-1;
+ HighBits <<= MVT::getSizeInBits(VT) - ShAmt;
+ uint64_t TypeMask = MVT::getIntVTBitMask(VT);
+
+ uint64_t InDemandedMask = (DemandedMask << ShAmt) & TypeMask;
+
+ // If any of the demanded bits are produced by the sign extension, we also
+ // demand the input sign bit.
+ if (HighBits & DemandedMask)
+ InDemandedMask |= MVT::getIntVTSignBit(VT);
+
+ if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
+ KnownZero, KnownOne, TLO, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero &= TypeMask;
+ KnownOne &= TypeMask;
+ KnownZero >>= SA->getValue();
+ KnownOne >>= SA->getValue();
+
+ // Handle the sign bits.
+ uint64_t SignBit = MVT::getIntVTSignBit(VT);
+ SignBit >>= SA->getValue(); // Adjust to where it is now in the mask.
+
+ // If the input sign bit is known to be zero, or if none of the top bits
+ // are demanded, turn this into an unsigned shift right.
+ if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
+ Op.getOperand(1)));
+ } else if (KnownOne & SignBit) { // New bits are known one.
+ KnownOne |= HighBits;
+ }
+ }
+ break;
+ case ISD::SIGN_EXTEND_INREG: {
+ MVT::ValueType VT = Op.getValueType();
+ MVT::ValueType EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
+
+ // Sign extension. Compute the demanded bits in the result that are not
+ // present in the input.
+ uint64_t NewBits = ~MVT::getIntVTBitMask(EVT) & DemandedMask;
+
+ // If none of the extended bits are demanded, eliminate the sextinreg.
+ if (NewBits == 0)
+ return TLO.CombineTo(Op, Op.getOperand(0));
+
+ uint64_t InSignBit = MVT::getIntVTSignBit(EVT);
+ int64_t InputDemandedBits = DemandedMask & MVT::getIntVTBitMask(EVT);
+
+ // Since the sign extended bits are demanded, we know that the sign
+ // bit is demanded.
+ InputDemandedBits |= InSignBit;
+
+ if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
+ KnownZero, KnownOne, TLO, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+
+ // If the sign bit of the input is known set or clear, then we know the
+ // top bits of the result.
+
+ // If the input sign bit is known zero, convert this into a zero extension.
+ if (KnownZero & InSignBit)
+ return TLO.CombineTo(Op,
+ TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
+
+ if (KnownOne & InSignBit) { // Input sign bit known set
+ KnownOne |= NewBits;
+ KnownZero &= ~NewBits;
+ } else { // Input sign bit unknown
+ KnownZero &= ~NewBits;
+ KnownOne &= ~NewBits;
+ }
+ break;
+ }
+ case ISD::CTTZ:
+ case ISD::CTLZ:
+ case ISD::CTPOP: {
+ MVT::ValueType VT = Op.getValueType();
+ unsigned LowBits = Log2_32(MVT::getSizeInBits(VT))+1;
+ KnownZero = ~((1ULL << LowBits)-1) & MVT::getIntVTBitMask(VT);
+ KnownOne = 0;
+ break;
+ }
+ case ISD::ZEXTLOAD: {
+ MVT::ValueType VT = cast<VTSDNode>(Op.getOperand(3))->getVT();
+ KnownZero |= ~MVT::getIntVTBitMask(VT) & DemandedMask;
+ break;
+ }
+ case ISD::ZERO_EXTEND: {
+ uint64_t InMask = MVT::getIntVTBitMask(Op.getOperand(0).getValueType());
+
+ // If none of the top bits are demanded, convert this into an any_extend.
+ uint64_t NewBits = (~InMask) & DemandedMask;
+ if (NewBits == 0)
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
+ Op.getValueType(),
+ Op.getOperand(0)));
+
+ if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
+ KnownZero, KnownOne, TLO, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero |= NewBits;
+ break;
+ }
+ case ISD::SIGN_EXTEND: {
+ MVT::ValueType InVT = Op.getOperand(0).getValueType();
+ uint64_t InMask = MVT::getIntVTBitMask(InVT);
+ uint64_t InSignBit = MVT::getIntVTSignBit(InVT);
+ uint64_t NewBits = (~InMask) & DemandedMask;
+
+ // If none of the top bits are demanded, convert this into an any_extend.
+ if (NewBits == 0)
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
+ Op.getOperand(0)));
+
+ // Since some of the sign extended bits are demanded, we know that the sign
+ // bit is demanded.
+ uint64_t InDemandedBits = DemandedMask & InMask;
+ InDemandedBits |= InSignBit;
+
+ if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
+ KnownOne, TLO, Depth+1))
+ return true;
+
+ // If the sign bit is known zero, convert this to a zero extend.
+ if (KnownZero & InSignBit)
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
+ Op.getValueType(),
+ Op.getOperand(0)));
+
+ // If the sign bit is known one, the top bits match.
+ if (KnownOne & InSignBit) {
+ KnownOne |= NewBits;
+ KnownZero &= ~NewBits;
+ } else { // Otherwise, top bits aren't known.
+ KnownOne &= ~NewBits;
+ KnownZero &= ~NewBits;
+ }
+ break;
+ }
+ case ISD::ANY_EXTEND: {
+ uint64_t InMask = MVT::getIntVTBitMask(Op.getOperand(0).getValueType());
+ if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
+ KnownZero, KnownOne, TLO, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ break;
+ }
+ case ISD::TRUNCATE: {
+ // Simplify the input, using demanded bit information, and compute the known
+ // zero/one bits live out.
+ if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask,
+ KnownZero, KnownOne, TLO, Depth+1))
+ return true;
+
+ // If the input is only used by this truncate, see if we can shrink it based
+ // on the known demanded bits.
+ if (Op.getOperand(0).Val->hasOneUse()) {
+ SDOperand In = Op.getOperand(0);
+ switch (In.getOpcode()) {
+ default: break;
+ case ISD::SRL:
+ // Shrink SRL by a constant if none of the high bits shifted in are
+ // demanded.
+ if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
+ uint64_t HighBits = MVT::getIntVTBitMask(In.getValueType());
+ HighBits &= ~MVT::getIntVTBitMask(Op.getValueType());
+ HighBits >>= ShAmt->getValue();
+
+ if (ShAmt->getValue() < MVT::getSizeInBits(Op.getValueType()) &&
+ (DemandedMask & HighBits) == 0) {
+ // None of the shifted in bits are needed. Add a truncate of the
+ // shift input, then shift it.
+ SDOperand NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
+ Op.getValueType(),
+ In.getOperand(0));
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
+ NewTrunc, In.getOperand(1)));
+ }
}
+ break;
}
}
+
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ uint64_t OutMask = MVT::getIntVTBitMask(Op.getValueType());
+ KnownZero &= OutMask;
+ KnownOne &= OutMask;
+ break;
+ }
+ case ISD::AssertZext: {
+ MVT::ValueType VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
+ uint64_t InMask = MVT::getIntVTBitMask(VT);
+ if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
+ KnownZero, KnownOne, TLO, Depth+1))
+ return true;
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero |= ~InMask & DemandedMask;
break;
+ }
+ case ISD::ADD:
+ case ISD::SUB:
+ case ISD::INTRINSIC_WO_CHAIN:
+ case ISD::INTRINSIC_W_CHAIN:
+ case ISD::INTRINSIC_VOID:
+ // Just use ComputeMaskedBits to compute output bits.
+ ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
+ break;
+ }
+
+ // If we know the value of all of the demanded bits, return this as a
+ // constant.
+ if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
+ return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
+
+ return false;
+}
+
+/// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
+/// this predicate to simplify operations downstream. Mask is known to be zero
+/// for bits that V cannot have.
+bool TargetLowering::MaskedValueIsZero(SDOperand Op, uint64_t Mask,
+ unsigned Depth) const {
+ uint64_t KnownZero, KnownOne;
+ ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ return (KnownZero & Mask) == Mask;
+}
+
+/// ComputeMaskedBits - Determine which of the bits specified in Mask are
+/// known to be either zero or one and return them in the KnownZero/KnownOne
+/// bitsets. This code only analyzes bits in Mask, in order to short-circuit
+/// processing.
+void TargetLowering::ComputeMaskedBits(SDOperand Op, uint64_t Mask,
+ uint64_t &KnownZero, uint64_t &KnownOne,
+ unsigned Depth) const {
+ KnownZero = KnownOne = 0; // Don't know anything.
+ if (Depth == 6 || Mask == 0)
+ return; // Limit search depth.
+
+ uint64_t KnownZero2, KnownOne2;
+
+ switch (Op.getOpcode()) {
+ case ISD::Constant:
+ // We know all of the bits for a constant!
+ KnownOne = cast<ConstantSDNode>(Op)->getValue() & Mask;
+ KnownZero = ~KnownOne & Mask;
+ return;
+ case ISD::AND:
+ // If either the LHS or the RHS are Zero, the result is zero.
+ ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
+ Mask &= ~KnownZero;
+ ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // Output known-1 bits are only known if set in both the LHS & RHS.
+ KnownOne &= KnownOne2;
+ // Output known-0 are known to be clear if zero in either the LHS | RHS.
+ KnownZero |= KnownZero2;
+ return;
+ case ISD::OR:
+ ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
+ Mask &= ~KnownOne;
+ ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // Output known-0 bits are only known if clear in both the LHS & RHS.
+ KnownZero &= KnownZero2;
+ // Output known-1 are known to be set if set in either the LHS | RHS.
+ KnownOne |= KnownOne2;
+ return;
+ case ISD::XOR: {
+ ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
+ ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // Output known-0 bits are known if clear or set in both the LHS & RHS.
+ uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
+ // Output known-1 are known to be set if set in only one of the LHS, RHS.
+ KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
+ KnownZero = KnownZeroOut;
+ return;
+ }
+ case ISD::SELECT:
+ ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
+ ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // Only known if known in both the LHS and RHS.
+ KnownOne &= KnownOne2;
+ KnownZero &= KnownZero2;
+ return;
+ case ISD::SELECT_CC:
+ ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1);
+ ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // Only known if known in both the LHS and RHS.
+ KnownOne &= KnownOne2;
+ KnownZero &= KnownZero2;
+ return;
+ case ISD::SETCC:
+ // If we know the result of a setcc has the top bits zero, use this info.
+ if (getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult)
+ KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
+ return;
+ case ISD::SHL:
+ // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
+ if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ Mask >>= SA->getValue();
+ ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero <<= SA->getValue();
+ KnownOne <<= SA->getValue();
+ KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
+ }
+ return;
+ case ISD::SRL:
+ // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
+ if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ uint64_t HighBits = (1ULL << SA->getValue())-1;
+ HighBits <<= MVT::getSizeInBits(Op.getValueType())-SA->getValue();
+ Mask <<= SA->getValue();
+ ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero >>= SA->getValue();
+ KnownOne >>= SA->getValue();
+ KnownZero |= HighBits; // high bits known zero.
+ }
+ return;
+ case ISD::SRA:
+ if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ uint64_t HighBits = (1ULL << SA->getValue())-1;
+ HighBits <<= MVT::getSizeInBits(Op.getValueType())-SA->getValue();
+ Mask <<= SA->getValue();
+ ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
+ assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
+ KnownZero >>= SA->getValue();
+ KnownOne >>= SA->getValue();
+
+ // Handle the sign bits.
+ uint64_t SignBit = 1ULL << (MVT::getSizeInBits(Op.getValueType())-1);
+ SignBit >>= SA->getValue(); // Adjust to where it is now in the mask.
+
+ if (KnownZero & SignBit) { // New bits are known zero.
+ KnownZero |= HighBits;
+ } else if (KnownOne & SignBit) { // New bits are known one.
+ KnownOne |= HighBits;
+ }
+ }
+ return;
+ case ISD::SIGN_EXTEND_INREG: {
+ MVT::ValueType VT = Op.getValueType();
+ MVT::ValueType EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
+
+ // Sign extension. Compute the demanded bits in the result that are not
+ // present in the input.
+ uint64_t NewBits = ~MVT::getIntVTBitMask(EVT) & Mask;
+
+ uint64_t InSignBit = MVT::getIntVTSignBit(EVT);
+ int64_t InputDemandedBits = Mask & MVT::getIntVTBitMask(EVT);
+
+ // If the sign extended bits are demanded, we know that the sign
+ // bit is demanded.
+ if (NewBits)
+ InputDemandedBits |= InSignBit;
+
+ ComputeMaskedBits(Op.getOperand(0), InputDemandedBits,
+ KnownZero, KnownOne, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+
+ // If the sign bit of the input is known set or clear, then we know the
+ // top bits of the result.
+ if (KnownZero & InSignBit) { // Input sign bit known clear
+ KnownZero |= NewBits;
+ KnownOne &= ~NewBits;
+ } else if (KnownOne & InSignBit) { // Input sign bit known set
+ KnownOne |= NewBits;
+ KnownZero &= ~NewBits;
+ } else { // Input sign bit unknown
+ KnownZero &= ~NewBits;
+ KnownOne &= ~NewBits;
+ }
+ return;
+ }
case ISD::CTTZ:
case ISD::CTLZ:
- case ISD::CTPOP:
- // Bit counting instructions can not set the high bits of the result
- // register. The max number of bits sets depends on the input.
- return (Mask & (MVT::getSizeInBits(Op.getValueType())*2-1)) == 0;
+ case ISD::CTPOP: {
+ MVT::ValueType VT = Op.getValueType();
+ unsigned LowBits = Log2_32(MVT::getSizeInBits(VT))+1;
+ KnownZero = ~((1ULL << LowBits)-1) & MVT::getIntVTBitMask(VT);
+ KnownOne = 0;
+ return;
+ }
+ case ISD::ZEXTLOAD: {
+ MVT::ValueType VT = cast<VTSDNode>(Op.getOperand(3))->getVT();
+ KnownZero |= ~MVT::getIntVTBitMask(VT) & Mask;
+ return;
+ }
+ case ISD::ZERO_EXTEND: {
+ uint64_t InMask = MVT::getIntVTBitMask(Op.getOperand(0).getValueType());
+ uint64_t NewBits = (~InMask) & Mask;
+ ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero,
+ KnownOne, Depth+1);
+ KnownZero |= NewBits & Mask;
+ KnownOne &= ~NewBits;
+ return;
+ }
+ case ISD::SIGN_EXTEND: {
+ MVT::ValueType InVT = Op.getOperand(0).getValueType();
+ unsigned InBits = MVT::getSizeInBits(InVT);
+ uint64_t InMask = MVT::getIntVTBitMask(InVT);
+ uint64_t InSignBit = 1ULL << (InBits-1);
+ uint64_t NewBits = (~InMask) & Mask;
+ uint64_t InDemandedBits = Mask & InMask;
+
+ // If any of the sign extended bits are demanded, we know that the sign
+ // bit is demanded.
+ if (NewBits & Mask)
+ InDemandedBits |= InSignBit;
+
+ ComputeMaskedBits(Op.getOperand(0), InDemandedBits, KnownZero,
+ KnownOne, Depth+1);
+ // If the sign bit is known zero or one, the top bits match.
+ if (KnownZero & InSignBit) {
+ KnownZero |= NewBits;
+ KnownOne &= ~NewBits;
+ } else if (KnownOne & InSignBit) {
+ KnownOne |= NewBits;
+ KnownZero &= ~NewBits;
+ } else { // Otherwise, top bits aren't known.
+ KnownOne &= ~NewBits;
+ KnownZero &= ~NewBits;
+ }
+ return;
+ }
+ case ISD::ANY_EXTEND: {
+ MVT::ValueType VT = Op.getOperand(0).getValueType();
+ ComputeMaskedBits(Op.getOperand(0), Mask & MVT::getIntVTBitMask(VT),
+ KnownZero, KnownOne, Depth+1);
+ return;
+ }
+ case ISD::TRUNCATE: {
+ ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ uint64_t OutMask = MVT::getIntVTBitMask(Op.getValueType());
+ KnownZero &= OutMask;
+ KnownOne &= OutMask;
+ break;
+ }
+ case ISD::AssertZext: {
+ MVT::ValueType VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
+ uint64_t InMask = MVT::getIntVTBitMask(VT);
+ ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero,
+ KnownOne, Depth+1);
+ KnownZero |= (~InMask) & Mask;
+ return;
+ }
+ case ISD::ADD: {
+ // If either the LHS or the RHS are Zero, the result is zero.
+ ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
+ ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // Output known-0 bits are known if clear or set in both the low clear bits
+ // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
+ // low 3 bits clear.
+ uint64_t KnownZeroOut = std::min(CountTrailingZeros_64(~KnownZero),
+ CountTrailingZeros_64(~KnownZero2));
+
+ KnownZero = (1ULL << KnownZeroOut) - 1;
+ KnownOne = 0;
+ return;
+ }
+ case ISD::SUB: {
+ ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0));
+ if (!CLHS) return;
+
+ // We know that the top bits of C-X are clear if X contains less bits
+ // than C (i.e. no wrap-around can happen). For example, 20-X is
+ // positive if we can prove that X is >= 0 and < 16.
+ MVT::ValueType VT = CLHS->getValueType(0);
+ if ((CLHS->getValue() & MVT::getIntVTSignBit(VT)) == 0) { // sign bit clear
+ unsigned NLZ = CountLeadingZeros_64(CLHS->getValue()+1);
+ uint64_t MaskV = (1ULL << (63-NLZ))-1; // NLZ can't be 64 with no sign bit
+ MaskV = ~MaskV & MVT::getIntVTBitMask(VT);
+ ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero, KnownOne, Depth+1);
+
+ // If all of the MaskV bits are known to be zero, then we know the output
+ // top bits are zero, because we now know that the output is from [0-C].
+ if ((KnownZero & MaskV) == MaskV) {
+ unsigned NLZ2 = CountLeadingZeros_64(CLHS->getValue());
+ KnownZero = ~((1ULL << (64-NLZ2))-1) & Mask; // Top bits known zero.
+ KnownOne = 0; // No one bits known.
+ } else {
+ KnownOne = KnownOne = 0; // Otherwise, nothing known.
+ }
+ }
+ return;
+ }
default:
// Allow the target to implement this method for its nodes.
- if (Op.getOpcode() >= ISD::BUILTIN_OP_END)
- return isMaskedValueZeroForTargetNode(Op, Mask);
- break;
+ if (Op.getOpcode() >= ISD::BUILTIN_OP_END) {
+ case ISD::INTRINSIC_WO_CHAIN:
+ case ISD::INTRINSIC_W_CHAIN:
+ case ISD::INTRINSIC_VOID:
+ computeMaskedBitsForTargetNode(Op, Mask, KnownZero, KnownOne);
+ }
+ return;
}
- return false;
}
-bool TargetLowering::isMaskedValueZeroForTargetNode(const SDOperand &Op,
- uint64_t Mask) const {
- assert(Op.getOpcode() >= ISD::BUILTIN_OP_END &&
+/// computeMaskedBitsForTargetNode - Determine which of the bits specified
+/// in Mask are known to be either zero or one and return them in the
+/// KnownZero/KnownOne bitsets.
+void TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
+ uint64_t Mask,
+ uint64_t &KnownZero,
+ uint64_t &KnownOne,
+ unsigned Depth) const {
+ assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
+ Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
+ Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
+ Op.getOpcode() == ISD::INTRINSIC_VOID) &&
"Should use MaskedValueIsZero if you don't know whether Op"
" is a target node!");
- return false;
+ KnownZero = 0;
+ KnownOne = 0;
}
-std::vector<unsigned> TargetLowering::
-getRegForInlineAsmConstraint(const std::string &Constraint) const {
- // Scan to see if this constraint is a register name.
- const MRegisterInfo *RI = TM.getRegisterInfo();
- for (unsigned i = 1, e = RI->getNumRegs(); i != e; ++i) {
- if (const char *Name = RI->get(i).Name)
- if (StringsEqualNoCase(Constraint, Name))
- return std::vector<unsigned>(1, i);
+/// ComputeNumSignBits - Return the number of times the sign bit of the
+/// register is replicated into the other bits. We know that at least 1 bit
+/// is always equal to the sign bit (itself), but other cases can give us
+/// information. For example, immediately after an "SRA X, 2", we know that
+/// the top 3 bits are all equal to each other, so we return 3.
+unsigned TargetLowering::ComputeNumSignBits(SDOperand Op, unsigned Depth) const{
+ MVT::ValueType VT = Op.getValueType();
+ assert(MVT::isInteger(VT) && "Invalid VT!");
+ unsigned VTBits = MVT::getSizeInBits(VT);
+ unsigned Tmp, Tmp2;
+
+ if (Depth == 6)
+ return 1; // Limit search depth.
+
+ switch (Op.getOpcode()) {
+ default: break;
+ case ISD::AssertSext:
+ Tmp = MVT::getSizeInBits(cast<VTSDNode>(Op.getOperand(1))->getVT());
+ return VTBits-Tmp+1;
+ case ISD::AssertZext:
+ Tmp = MVT::getSizeInBits(cast<VTSDNode>(Op.getOperand(1))->getVT());
+ return VTBits-Tmp;
+
+ case ISD::SEXTLOAD: // '17' bits known
+ Tmp = MVT::getSizeInBits(cast<VTSDNode>(Op.getOperand(3))->getVT());
+ return VTBits-Tmp+1;
+ case ISD::ZEXTLOAD: // '16' bits known
+ Tmp = MVT::getSizeInBits(cast<VTSDNode>(Op.getOperand(3))->getVT());
+ return VTBits-Tmp;
+
+ case ISD::Constant: {
+ uint64_t Val = cast<ConstantSDNode>(Op)->getValue();
+ // If negative, invert the bits, then look at it.
+ if (Val & MVT::getIntVTSignBit(VT))
+ Val = ~Val;
+
+ // Shift the bits so they are the leading bits in the int64_t.
+ Val <<= 64-VTBits;
+
+ // Return # leading zeros. We use 'min' here in case Val was zero before
+ // shifting. We don't want to return '64' as for an i32 "0".
+ return std::min(VTBits, CountLeadingZeros_64(Val));
}
+
+ case ISD::SIGN_EXTEND:
+ Tmp = VTBits-MVT::getSizeInBits(Op.getOperand(0).getValueType());
+ return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp;
+
+ case ISD::SIGN_EXTEND_INREG:
+ // Max of the input and what this extends.
+ Tmp = MVT::getSizeInBits(cast<VTSDNode>(Op.getOperand(1))->getVT());
+ Tmp = VTBits-Tmp+1;
+
+ Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1);
+ return std::max(Tmp, Tmp2);
+
+ case ISD::SRA:
+ Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
+ // SRA X, C -> adds C sign bits.
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ Tmp += C->getValue();
+ if (Tmp > VTBits) Tmp = VTBits;
+ }
+ return Tmp;
+ case ISD::SHL:
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ // shl destroys sign bits.
+ Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
+ if (C->getValue() >= VTBits || // Bad shift.
+ C->getValue() >= Tmp) break; // Shifted all sign bits out.
+ return Tmp - C->getValue();
+ }
+ break;
+ case ISD::AND:
+ case ISD::OR:
+ case ISD::XOR: // NOT is handled here.
+ // Logical binary ops preserve the number of sign bits.
+ Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
+ if (Tmp == 1) return 1; // Early out.
+ Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
+ return std::min(Tmp, Tmp2);
+
+ case ISD::SELECT:
+ Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
+ if (Tmp == 1) return 1; // Early out.
+ Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
+ return std::min(Tmp, Tmp2);
+
+ case ISD::SETCC:
+ // If setcc returns 0/-1, all bits are sign bits.
+ if (getSetCCResultContents() == ZeroOrNegativeOneSetCCResult)
+ return VTBits;
+ break;
+ case ISD::ROTL:
+ case ISD::ROTR:
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ unsigned RotAmt = C->getValue() & (VTBits-1);
+
+ // Handle rotate right by N like a rotate left by 32-N.
+ if (Op.getOpcode() == ISD::ROTR)
+ RotAmt = (VTBits-RotAmt) & (VTBits-1);
- // Not a physreg, must not be a register reference or something.
+ // If we aren't rotating out all of the known-in sign bits, return the
+ // number that are left. This handles rotl(sext(x), 1) for example.
+ Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
+ if (Tmp > RotAmt+1) return Tmp-RotAmt;
+ }
+ break;
+ case ISD::ADD:
+ // Add can have at most one carry bit. Thus we know that the output
+ // is, at worst, one more bit than the inputs.
+ Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
+ if (Tmp == 1) return 1; // Early out.
+
+ // Special case decrementing a value (ADD X, -1):
+ if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(0)))
+ if (CRHS->isAllOnesValue()) {
+ uint64_t KnownZero, KnownOne;
+ uint64_t Mask = MVT::getIntVTBitMask(VT);
+ ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
+
+ // If the input is known to be 0 or 1, the output is 0/-1, which is all
+ // sign bits set.
+ if ((KnownZero|1) == Mask)
+ return VTBits;
+
+ // If we are subtracting one from a positive number, there is no carry
+ // out of the result.
+ if (KnownZero & MVT::getIntVTSignBit(VT))
+ return Tmp;
+ }
+
+ Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
+ if (Tmp2 == 1) return 1;
+ return std::min(Tmp, Tmp2)-1;
+ break;
+
+ case ISD::SUB:
+ Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
+ if (Tmp2 == 1) return 1;
+
+ // Handle NEG.
+ if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0)))
+ if (CLHS->getValue() == 0) {
+ uint64_t KnownZero, KnownOne;
+ uint64_t Mask = MVT::getIntVTBitMask(VT);
+ ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
+ // If the input is known to be 0 or 1, the output is 0/-1, which is all
+ // sign bits set.
+ if ((KnownZero|1) == Mask)
+ return VTBits;
+
+ // If the input is known to be positive (the sign bit is known clear),
+ // the output of the NEG has the same number of sign bits as the input.
+ if (KnownZero & MVT::getIntVTSignBit(VT))
+ return Tmp2;
+
+ // Otherwise, we treat this like a SUB.
+ }
+
+ // Sub can have at most one carry bit. Thus we know that the output
+ // is, at worst, one more bit than the inputs.
+ Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
+ if (Tmp == 1) return 1; // Early out.
+ return std::min(Tmp, Tmp2)-1;
+ break;
+ case ISD::TRUNCATE:
+ // FIXME: it's tricky to do anything useful for this, but it is an important
+ // case for targets like X86.
+ break;
+ }
+
+ // Allow the target to implement this method for its nodes.
+ if (Op.getOpcode() >= ISD::BUILTIN_OP_END ||
+ Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
+ Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
+ Op.getOpcode() == ISD::INTRINSIC_VOID) {
+ unsigned NumBits = ComputeNumSignBitsForTargetNode(Op, Depth);
+ if (NumBits > 1) return NumBits;
+ }
+
+ // Finally, if we can prove that the top bits of the result are 0's or 1's,
+ // use this information.
+ uint64_t KnownZero, KnownOne;
+ uint64_t Mask = MVT::getIntVTBitMask(VT);
+ ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth);
+
+ uint64_t SignBit = MVT::getIntVTSignBit(VT);
+ if (KnownZero & SignBit) { // SignBit is 0
+ Mask = KnownZero;
+ } else if (KnownOne & SignBit) { // SignBit is 1;
+ Mask = KnownOne;
+ } else {
+ // Nothing known.
+ return 1;
+ }
+
+ // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
+ // the number of identical bits in the top of the input value.
+ Mask ^= ~0ULL;
+ Mask <<= 64-VTBits;
+ // Return # leading zeros. We use 'min' here in case Val was zero before
+ // shifting. We don't want to return '64' as for an i32 "0".
+ return std::min(VTBits, CountLeadingZeros_64(Mask));
+}
+
+
+
+/// ComputeNumSignBitsForTargetNode - This method can be implemented by
+/// targets that want to expose additional information about sign bits to the
+/// DAG Combiner.
+unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDOperand Op,
+ unsigned Depth) const {
+ assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
+ Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
+ Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
+ Op.getOpcode() == ISD::INTRINSIC_VOID) &&
+ "Should use ComputeNumSignBits if you don't know whether Op"
+ " is a target node!");
+ return 1;
+}
+
+
+SDOperand TargetLowering::
+PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
+ // Default implementation: no optimization.
+ return SDOperand();
+}
+
+//===----------------------------------------------------------------------===//
+// Inline Assembler Implementation Methods
+//===----------------------------------------------------------------------===//
+
+TargetLowering::ConstraintType
+TargetLowering::getConstraintType(char ConstraintLetter) const {
+ // FIXME: lots more standard ones to handle.
+ switch (ConstraintLetter) {
+ default: return C_Unknown;
+ case 'r': return C_RegisterClass;
+ case 'm': // memory
+ case 'o': // offsetable
+ case 'V': // not offsetable
+ return C_Memory;
+ case 'i': // Simple Integer or Relocatable Constant
+ case 'n': // Simple Integer
+ case 's': // Relocatable Constant
+ case 'I': // Target registers.
+ case 'J':
+ case 'K':
+ case 'L':
+ case 'M':
+ case 'N':
+ case 'O':
+ case 'P':
+ return C_Other;
+ }
+}
+
+bool TargetLowering::isOperandValidForConstraint(SDOperand Op,
+ char ConstraintLetter) {
+ switch (ConstraintLetter) {
+ default: return false;
+ case 'i': // Simple Integer or Relocatable Constant
+ case 'n': // Simple Integer
+ case 's': // Relocatable Constant
+ return true; // FIXME: not right.
+ }
+}
+
+
+std::vector<unsigned> TargetLowering::
+getRegClassForInlineAsmConstraint(const std::string &Constraint,
+ MVT::ValueType VT) const {
return std::vector<unsigned>();
}
+
+std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
+getRegForInlineAsmConstraint(const std::string &Constraint,
+ MVT::ValueType VT) const {
+ if (Constraint[0] != '{')
+ return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
+ assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
+
+ // Remove the braces from around the name.
+ std::string RegName(Constraint.begin()+1, Constraint.end()-1);
+
+ // Figure out which register class contains this reg.
+ const MRegisterInfo *RI = TM.getRegisterInfo();
+ for (MRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
+ E = RI->regclass_end(); RCI != E; ++RCI) {
+ const TargetRegisterClass *RC = *RCI;
+
+ // If none of the the value types for this register class are valid, we
+ // can't use it. For example, 64-bit reg classes on 32-bit targets.
+ bool isLegal = false;
+ for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
+ I != E; ++I) {
+ if (isTypeLegal(*I)) {
+ isLegal = true;
+ break;
+ }
+ }
+
+ if (!isLegal) continue;
+
+ for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
+ I != E; ++I) {
+ if (StringsEqualNoCase(RegName, RI->get(*I).Name))
+ return std::make_pair(*I, RC);
+ }
+ }
+
+ return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
+}
+
+//===----------------------------------------------------------------------===//
+// Loop Strength Reduction hooks
+//===----------------------------------------------------------------------===//
+
+/// isLegalAddressImmediate - Return true if the integer value or
+/// GlobalValue can be used as the offset of the target addressing mode.
+bool TargetLowering::isLegalAddressImmediate(int64_t V) const {
+ return false;
+}
+bool TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
+ return false;
+}