//
// The LLVM Compiler Infrastructure
//
-// This file was developed by Nate Begeman and is distributed under the
-// University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass combines dag nodes to form fewer, simpler DAG nodes. It can be run
// both before and after the DAG is legalized.
-//
-// FIXME: Missing folds
-// sdiv, udiv, srem, urem (X, const) where X is an integer can be expanded into
-// a sequence of multiplies, shifts, and adds. This should be controlled by
-// some kind of hint from the target that int div is expensive.
-// various folds of mulh[s,u] by constants such as -1, powers of 2, etc.
-//
-// FIXME: select C, pow2, pow2 -> something smart
-// FIXME: trunc(select X, Y, Z) -> select X, trunc(Y), trunc(Z)
-// FIXME: Dead stores -> nuke
-// FIXME: shr X, (and Y,31) -> shr X, Y (TRICKY!)
-// FIXME: mul (x, const) -> shifts + adds
-// FIXME: undef values
-// FIXME: divide by zero is currently left unfolded. do we want to turn this
-// into an undef?
-// FIXME: select ne (select cc, 1, 0), 0, true, false -> select cc, true, false
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "dagcombine"
#include "llvm/CodeGen/SelectionDAG.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Target/TargetData.h"
+#include "llvm/Target/TargetFrameInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
AddToWorkList(*UI);
}
- /// removeFromWorkList - remove all instances of N from the worklist.
- ///
- void removeFromWorkList(SDNode *N) {
- WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
- WorkList.end());
- }
-
/// visit - call the node-specific routine that knows how to fold each
/// particular type of node.
SDOperand visit(SDNode *N);
WorkList.push_back(N);
}
- SDOperand CombineTo(SDNode *N, const SDOperand *To, unsigned NumTo,
- bool AddTo = true) {
- assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
- ++NodesCombined;
- DOUT << "\nReplacing.1 "; DEBUG(N->dump(&DAG));
- DOUT << "\nWith: "; DEBUG(To[0].Val->dump(&DAG));
- DOUT << " and " << NumTo-1 << " other values\n";
- std::vector<SDNode*> NowDead;
- DAG.ReplaceAllUsesWith(N, To, &NowDead);
-
- if (AddTo) {
- // Push the new nodes and any users onto the worklist
- for (unsigned i = 0, e = NumTo; i != e; ++i) {
- AddToWorkList(To[i].Val);
- AddUsersToWorkList(To[i].Val);
- }
- }
-
- // Nodes can be reintroduced into the worklist. Make sure we do not
- // process a node that has been replaced.
- removeFromWorkList(N);
- for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
- removeFromWorkList(NowDead[i]);
-
- // Finally, since the node is now dead, remove it from the graph.
- DAG.DeleteNode(N);
- return SDOperand(N, 0);
+ /// removeFromWorkList - remove all instances of N from the worklist.
+ ///
+ void removeFromWorkList(SDNode *N) {
+ WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
+ WorkList.end());
}
+ SDOperand CombineTo(SDNode *N, const SDOperand *To, unsigned NumTo,
+ bool AddTo = true);
+
SDOperand CombineTo(SDNode *N, SDOperand Res, bool AddTo = true) {
return CombineTo(N, &Res, 1, AddTo);
}
SDOperand To[] = { Res0, Res1 };
return CombineTo(N, To, 2, AddTo);
}
+
private:
/// SimplifyDemandedBits - Check the specified integer node value to see if
/// it can be simplified or if things it uses can be simplified by bit
/// propagation. If so, return true.
- bool SimplifyDemandedBits(SDOperand Op, uint64_t Demanded = ~0ULL) {
- TargetLowering::TargetLoweringOpt TLO(DAG);
- uint64_t KnownZero, KnownOne;
- Demanded &= MVT::getIntVTBitMask(Op.getValueType());
- if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
- return false;
-
- // Revisit the node.
- AddToWorkList(Op.Val);
-
- // Replace the old value with the new one.
- ++NodesCombined;
- DOUT << "\nReplacing.2 "; DEBUG(TLO.Old.Val->dump(&DAG));
- DOUT << "\nWith: "; DEBUG(TLO.New.Val->dump(&DAG));
- DOUT << '\n';
-
- std::vector<SDNode*> NowDead;
- DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &NowDead);
-
- // Push the new node and any (possibly new) users onto the worklist.
- AddToWorkList(TLO.New.Val);
- AddUsersToWorkList(TLO.New.Val);
-
- // Nodes can end up on the worklist more than once. Make sure we do
- // not process a node that has been replaced.
- for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
- removeFromWorkList(NowDead[i]);
-
- // Finally, if the node is now dead, remove it from the graph. The node
- // may not be dead if the replacement process recursively simplified to
- // something else needing this node.
- if (TLO.Old.Val->use_empty()) {
- removeFromWorkList(TLO.Old.Val);
-
- // If the operands of this node are only used by the node, they will now
- // be dead. Make sure to visit them first to delete dead nodes early.
- for (unsigned i = 0, e = TLO.Old.Val->getNumOperands(); i != e; ++i)
- if (TLO.Old.Val->getOperand(i).Val->hasOneUse())
- AddToWorkList(TLO.Old.Val->getOperand(i).Val);
-
- DAG.DeleteNode(TLO.Old.Val);
- }
- return true;
+ bool SimplifyDemandedBits(SDOperand Op) {
+ APInt Demanded = APInt::getAllOnesValue(Op.getValueSizeInBits());
+ return SimplifyDemandedBits(Op, Demanded);
}
+ bool SimplifyDemandedBits(SDOperand Op, const APInt &Demanded);
+
bool CombineToPreIndexedLoadStore(SDNode *N);
bool CombineToPostIndexedLoadStore(SDNode *N);
// otherwise - N should be replaced by the returned Operand.
//
SDOperand visitTokenFactor(SDNode *N);
+ SDOperand visitMERGE_VALUES(SDNode *N);
SDOperand visitADD(SDNode *N);
SDOperand visitSUB(SDNode *N);
SDOperand visitADDC(SDNode *N);
SDOperand XformToShuffleWithZero(SDNode *N);
SDOperand ReassociateOps(unsigned Opc, SDOperand LHS, SDOperand RHS);
+ SDOperand visitShiftByConstant(SDNode *N, unsigned Amt);
+
bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS);
SDOperand SimplifyBinOpWithSameOpcodeHands(SDNode *N);
SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2);
bool NotExtCompare = false);
SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
ISD::CondCode Cond, bool foldBooleans = true);
- bool SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, unsigned HiOp);
+ SDOperand SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
+ unsigned HiOp);
SDOperand ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *, MVT::ValueType);
SDOperand BuildSDIV(SDNode *N);
SDOperand BuildUDIV(SDNode *N);
SDNode *MatchRotate(SDOperand LHS, SDOperand RHS);
SDOperand ReduceLoadWidth(SDNode *N);
- SDOperand GetDemandedBits(SDOperand V, uint64_t Mask);
+ SDOperand GetDemandedBits(SDOperand V, const APInt &Mask);
/// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
/// looking for aliasing nodes and adding them to the Aliases vector.
};
}
+
+namespace {
+/// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
+/// nodes from the worklist.
+class VISIBILITY_HIDDEN WorkListRemover :
+ public SelectionDAG::DAGUpdateListener {
+ DAGCombiner &DC;
+public:
+ explicit WorkListRemover(DAGCombiner &dc) : DC(dc) {}
+
+ virtual void NodeDeleted(SDNode *N) {
+ DC.removeFromWorkList(N);
+ }
+
+ virtual void NodeUpdated(SDNode *N) {
+ // Ignore updates.
+ }
+};
+}
+
//===----------------------------------------------------------------------===//
// TargetLowering::DAGCombinerInfo implementation
//===----------------------------------------------------------------------===//
/// isNegatibleForFree - Return 1 if we can compute the negated form of the
/// specified expression for the same cost as the expression itself, or 2 if we
/// can compute the negated form more cheaply than the expression itself.
-static char isNegatibleForFree(SDOperand Op, unsigned Depth = 0) {
+static char isNegatibleForFree(SDOperand Op, bool AfterLegalize,
+ unsigned Depth = 0) {
// No compile time optimizations on this type.
if (Op.getValueType() == MVT::ppcf128)
return 0;
switch (Op.getOpcode()) {
default: return false;
case ISD::ConstantFP:
- return 1;
+ // Don't invert constant FP values after legalize. The negated constant
+ // isn't necessarily legal.
+ return AfterLegalize ? 0 : 1;
case ISD::FADD:
// FIXME: determine better conditions for this xform.
if (!UnsafeFPMath) return 0;
// -(A+B) -> -A - B
- if (char V = isNegatibleForFree(Op.getOperand(0), Depth+1))
+ if (char V = isNegatibleForFree(Op.getOperand(0), AfterLegalize, Depth+1))
return V;
// -(A+B) -> -B - A
- return isNegatibleForFree(Op.getOperand(1), Depth+1);
+ return isNegatibleForFree(Op.getOperand(1), AfterLegalize, Depth+1);
case ISD::FSUB:
// We can't turn -(A-B) into B-A when we honor signed zeros.
if (!UnsafeFPMath) return 0;
if (HonorSignDependentRoundingFPMath()) return 0;
// -(X*Y) -> (-X * Y) or (X*-Y)
- if (char V = isNegatibleForFree(Op.getOperand(0), Depth+1))
+ if (char V = isNegatibleForFree(Op.getOperand(0), AfterLegalize, Depth+1))
return V;
- return isNegatibleForFree(Op.getOperand(1), Depth+1);
+ return isNegatibleForFree(Op.getOperand(1), AfterLegalize, Depth+1);
case ISD::FP_EXTEND:
case ISD::FP_ROUND:
case ISD::FSIN:
- return isNegatibleForFree(Op.getOperand(0), Depth+1);
+ return isNegatibleForFree(Op.getOperand(0), AfterLegalize, Depth+1);
}
}
/// GetNegatedExpression - If isNegatibleForFree returns true, this function
/// returns the newly negated expression.
static SDOperand GetNegatedExpression(SDOperand Op, SelectionDAG &DAG,
- unsigned Depth = 0) {
+ bool AfterLegalize, unsigned Depth = 0) {
// fneg is removable even if it has multiple uses.
if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
assert(UnsafeFPMath);
// -(A+B) -> -A - B
- if (isNegatibleForFree(Op.getOperand(0), Depth+1))
+ if (isNegatibleForFree(Op.getOperand(0), AfterLegalize, Depth+1))
return DAG.getNode(ISD::FSUB, Op.getValueType(),
- GetNegatedExpression(Op.getOperand(0), DAG, Depth+1),
+ GetNegatedExpression(Op.getOperand(0), DAG,
+ AfterLegalize, Depth+1),
Op.getOperand(1));
// -(A+B) -> -B - A
return DAG.getNode(ISD::FSUB, Op.getValueType(),
- GetNegatedExpression(Op.getOperand(1), DAG, Depth+1),
+ GetNegatedExpression(Op.getOperand(1), DAG,
+ AfterLegalize, Depth+1),
Op.getOperand(0));
case ISD::FSUB:
// We can't turn -(A-B) into B-A when we honor signed zeros.
assert(!HonorSignDependentRoundingFPMath());
// -(X*Y) -> -X * Y
- if (isNegatibleForFree(Op.getOperand(0), Depth+1))
+ if (isNegatibleForFree(Op.getOperand(0), AfterLegalize, Depth+1))
return DAG.getNode(Op.getOpcode(), Op.getValueType(),
- GetNegatedExpression(Op.getOperand(0), DAG, Depth+1),
+ GetNegatedExpression(Op.getOperand(0), DAG,
+ AfterLegalize, Depth+1),
Op.getOperand(1));
// -(X*Y) -> X * -Y
return DAG.getNode(Op.getOpcode(), Op.getValueType(),
Op.getOperand(0),
- GetNegatedExpression(Op.getOperand(1), DAG, Depth+1));
+ GetNegatedExpression(Op.getOperand(1), DAG,
+ AfterLegalize, Depth+1));
case ISD::FP_EXTEND:
- case ISD::FP_ROUND:
case ISD::FSIN:
return DAG.getNode(Op.getOpcode(), Op.getValueType(),
- GetNegatedExpression(Op.getOperand(0), DAG, Depth+1));
+ GetNegatedExpression(Op.getOperand(0), DAG,
+ AfterLegalize, Depth+1));
+ case ISD::FP_ROUND:
+ return DAG.getNode(ISD::FP_ROUND, Op.getValueType(),
+ GetNegatedExpression(Op.getOperand(0), DAG,
+ AfterLegalize, Depth+1),
+ Op.getOperand(1));
}
}
return SDOperand();
}
+SDOperand DAGCombiner::CombineTo(SDNode *N, const SDOperand *To, unsigned NumTo,
+ bool AddTo) {
+ assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
+ ++NodesCombined;
+ DOUT << "\nReplacing.1 "; DEBUG(N->dump(&DAG));
+ DOUT << "\nWith: "; DEBUG(To[0].Val->dump(&DAG));
+ DOUT << " and " << NumTo-1 << " other values\n";
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesWith(N, To, &DeadNodes);
+
+ if (AddTo) {
+ // Push the new nodes and any users onto the worklist
+ for (unsigned i = 0, e = NumTo; i != e; ++i) {
+ AddToWorkList(To[i].Val);
+ AddUsersToWorkList(To[i].Val);
+ }
+ }
+
+ // Nodes can be reintroduced into the worklist. Make sure we do not
+ // process a node that has been replaced.
+ removeFromWorkList(N);
+
+ // Finally, since the node is now dead, remove it from the graph.
+ DAG.DeleteNode(N);
+ return SDOperand(N, 0);
+}
+
+/// SimplifyDemandedBits - Check the specified integer node value to see if
+/// it can be simplified or if things it uses can be simplified by bit
+/// propagation. If so, return true.
+bool DAGCombiner::SimplifyDemandedBits(SDOperand Op, const APInt &Demanded) {
+ TargetLowering::TargetLoweringOpt TLO(DAG, AfterLegalize);
+ APInt KnownZero, KnownOne;
+ if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
+ return false;
+
+ // Revisit the node.
+ AddToWorkList(Op.Val);
+
+ // Replace the old value with the new one.
+ ++NodesCombined;
+ DOUT << "\nReplacing.2 "; DEBUG(TLO.Old.Val->dump(&DAG));
+ DOUT << "\nWith: "; DEBUG(TLO.New.Val->dump(&DAG));
+ DOUT << '\n';
+
+ // Replace all uses. If any nodes become isomorphic to other nodes and
+ // are deleted, make sure to remove them from our worklist.
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes);
+
+ // Push the new node and any (possibly new) users onto the worklist.
+ AddToWorkList(TLO.New.Val);
+ AddUsersToWorkList(TLO.New.Val);
+
+ // Finally, if the node is now dead, remove it from the graph. The node
+ // may not be dead if the replacement process recursively simplified to
+ // something else needing this node.
+ if (TLO.Old.Val->use_empty()) {
+ removeFromWorkList(TLO.Old.Val);
+
+ // If the operands of this node are only used by the node, they will now
+ // be dead. Make sure to visit them first to delete dead nodes early.
+ for (unsigned i = 0, e = TLO.Old.Val->getNumOperands(); i != e; ++i)
+ if (TLO.Old.Val->getOperand(i).Val->hasOneUse())
+ AddToWorkList(TLO.Old.Val->getOperand(i).Val);
+
+ DAG.DeleteNode(TLO.Old.Val);
+ }
+ return true;
+}
+
//===----------------------------------------------------------------------===//
// Main DAG Combiner implementation
//===----------------------------------------------------------------------===//
SDOperand RV = combine(N);
- if (RV.Val) {
- ++NodesCombined;
- // If we get back the same node we passed in, rather than a new node or
- // zero, we know that the node must have defined multiple values and
- // CombineTo was used. Since CombineTo takes care of the worklist
- // mechanics for us, we have no work to do in this case.
- if (RV.Val != N) {
- assert(N->getOpcode() != ISD::DELETED_NODE &&
- RV.Val->getOpcode() != ISD::DELETED_NODE &&
- "Node was deleted but visit returned new node!");
-
- DOUT << "\nReplacing.3 "; DEBUG(N->dump(&DAG));
- DOUT << "\nWith: "; DEBUG(RV.Val->dump(&DAG));
- DOUT << '\n';
- std::vector<SDNode*> NowDead;
- if (N->getNumValues() == RV.Val->getNumValues())
- DAG.ReplaceAllUsesWith(N, RV.Val, &NowDead);
- else {
- assert(N->getValueType(0) == RV.getValueType() && "Type mismatch");
- SDOperand OpV = RV;
- DAG.ReplaceAllUsesWith(N, &OpV, &NowDead);
- }
-
- // Push the new node and any users onto the worklist
- AddToWorkList(RV.Val);
- AddUsersToWorkList(RV.Val);
-
- // Nodes can be reintroduced into the worklist. Make sure we do not
- // process a node that has been replaced.
- removeFromWorkList(N);
- for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
- removeFromWorkList(NowDead[i]);
-
- // Finally, since the node is now dead, remove it from the graph.
- DAG.DeleteNode(N);
- }
+ if (RV.Val == 0)
+ continue;
+
+ ++NodesCombined;
+
+ // If we get back the same node we passed in, rather than a new node or
+ // zero, we know that the node must have defined multiple values and
+ // CombineTo was used. Since CombineTo takes care of the worklist
+ // mechanics for us, we have no work to do in this case.
+ if (RV.Val == N)
+ continue;
+
+ assert(N->getOpcode() != ISD::DELETED_NODE &&
+ RV.Val->getOpcode() != ISD::DELETED_NODE &&
+ "Node was deleted but visit returned new node!");
+
+ DOUT << "\nReplacing.3 "; DEBUG(N->dump(&DAG));
+ DOUT << "\nWith: "; DEBUG(RV.Val->dump(&DAG));
+ DOUT << '\n';
+ WorkListRemover DeadNodes(*this);
+ if (N->getNumValues() == RV.Val->getNumValues())
+ DAG.ReplaceAllUsesWith(N, RV.Val, &DeadNodes);
+ else {
+ assert(N->getValueType(0) == RV.getValueType() &&
+ N->getNumValues() == 1 && "Type mismatch");
+ SDOperand OpV = RV;
+ DAG.ReplaceAllUsesWith(N, &OpV, &DeadNodes);
}
+
+ // Push the new node and any users onto the worklist
+ AddToWorkList(RV.Val);
+ AddUsersToWorkList(RV.Val);
+
+ // Add any uses of the old node to the worklist in case this node is the
+ // last one that uses them. They may become dead after this node is
+ // deleted.
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
+ AddToWorkList(N->getOperand(i).Val);
+
+ // Nodes can be reintroduced into the worklist. Make sure we do not
+ // process a node that has been replaced.
+ removeFromWorkList(N);
+
+ // Finally, since the node is now dead, remove it from the graph.
+ DAG.DeleteNode(N);
}
// If the root changed (e.g. it was a dead load, update the root).
switch(N->getOpcode()) {
default: break;
case ISD::TokenFactor: return visitTokenFactor(N);
+ case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
case ISD::ADD: return visitADD(N);
case ISD::SUB: return visitSUB(N);
case ISD::ADDC: return visitADDC(N);
// If we've change things around then replace token factor.
if (Changed) {
- if (Ops.size() == 0) {
+ if (Ops.empty()) {
// The entry token is the only possible outcome.
Result = DAG.getEntryNode();
} else {
return Result;
}
+/// MERGE_VALUES can always be eliminated.
+SDOperand DAGCombiner::visitMERGE_VALUES(SDNode *N) {
+ WorkListRemover DeadNodes(*this);
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
+ DAG.ReplaceAllUsesOfValueWith(SDOperand(N, i), N->getOperand(i),
+ &DeadNodes);
+ removeFromWorkList(N);
+ DAG.DeleteNode(N);
+ return SDOperand(N, 0); // Return N so it doesn't get rechecked!
+}
+
+
static
SDOperand combineShlAddConstant(SDOperand N0, SDOperand N1, SelectionDAG &DAG) {
MVT::ValueType VT = N0.getValueType();
RHS.getOpcode() == ISD::Constant &&
cast<ConstantSDNode>(RHS)->isNullValue()) {
std::swap(LHS, RHS);
- bool isInt = MVT::isInteger(isSlctCC ? Slct.getOperand(0).getValueType()
- : Slct.getOperand(0).getOperand(0).getValueType());
+ SDOperand Op0 = Slct.getOperand(0);
+ bool isInt = MVT::isInteger(isSlctCC ? Op0.getValueType()
+ : Op0.getOperand(0).getValueType());
CC = ISD::getSetCCInverse(CC, isInt);
DoXform = true;
InvCC = true;
return N1;
// fold (add c1, c2) -> c1+c2
if (N0C && N1C)
- return DAG.getNode(ISD::ADD, VT, N0, N1);
+ return DAG.getConstant(N0C->getValue() + N1C->getValue(), VT);
// canonicalize constant to RHS
if (N0C && !N1C)
return DAG.getNode(ISD::ADD, VT, N1, N0);
// fold (a+b) -> (a|b) iff a and b share no bits.
if (MVT::isInteger(VT) && !MVT::isVector(VT)) {
- uint64_t LHSZero, LHSOne;
- uint64_t RHSZero, RHSOne;
- uint64_t Mask = MVT::getIntVTBitMask(VT);
+ APInt LHSZero, LHSOne;
+ APInt RHSZero, RHSOne;
+ APInt Mask = APInt::getAllOnesValue(MVT::getSizeInBits(VT));
DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
- if (LHSZero) {
+ if (LHSZero.getBoolValue()) {
DAG.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
// If all possibly-set bits on the LHS are clear on the RHS, return an OR.
return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, MVT::Flag));
// fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
- uint64_t LHSZero, LHSOne;
- uint64_t RHSZero, RHSOne;
- uint64_t Mask = MVT::getIntVTBitMask(VT);
+ APInt LHSZero, LHSOne;
+ APInt RHSZero, RHSOne;
+ APInt Mask = APInt::getAllOnesValue(MVT::getSizeInBits(VT));
DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
- if (LHSZero) {
+ if (LHSZero.getBoolValue()) {
DAG.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
// If all possibly-set bits on the LHS are clear on the RHS, return an OR.
return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
// If we know the sign bits of both operands are zero, strength reduce to a
// udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
- uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
- if (DAG.MaskedValueIsZero(N1, SignBit) &&
- DAG.MaskedValueIsZero(N0, SignBit))
- return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
+ if (!MVT::isVector(VT)) {
+ if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
+ return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
+ }
// fold (sdiv X, pow2) -> simple ops after legalize
if (N1C && N1C->getValue() && !TLI.isIntDivCheap() &&
(isPowerOf2_64(N1C->getSignExtended()) ||
return DAG.getNode(ISD::SREM, VT, N0, N1);
// If we know the sign bits of both operands are zero, strength reduce to a
// urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
- uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
- if (DAG.MaskedValueIsZero(N1, SignBit) &&
- DAG.MaskedValueIsZero(N0, SignBit))
- return DAG.getNode(ISD::UREM, VT, N0, N1);
+ if (!MVT::isVector(VT)) {
+ if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
+ return DAG.getNode(ISD::UREM, VT, N0, N1);
+ }
- // Unconditionally lower X%C -> X-X/C*C. This allows the X/C logic to hack on
- // the remainder operation.
+ // If X/C can be simplified by the division-by-constant logic, lower
+ // X%C to the equivalent of X-X/C*C.
if (N1C && !N1C->isNullValue()) {
SDOperand Div = DAG.getNode(ISD::SDIV, VT, N0, N1);
- SDOperand Mul = DAG.getNode(ISD::MUL, VT, Div, N1);
- SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul);
AddToWorkList(Div.Val);
- AddToWorkList(Mul.Val);
- return Sub;
+ SDOperand OptimizedDiv = combine(Div.Val);
+ if (OptimizedDiv.Val && OptimizedDiv.Val != Div.Val) {
+ SDOperand Mul = DAG.getNode(ISD::MUL, VT, OptimizedDiv, N1);
+ SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul);
+ AddToWorkList(Mul.Val);
+ return Sub;
+ }
}
// undef % X -> 0
}
}
- // Unconditionally lower X%C -> X-X/C*C. This allows the X/C logic to hack on
- // the remainder operation.
+ // If X/C can be simplified by the division-by-constant logic, lower
+ // X%C to the equivalent of X-X/C*C.
if (N1C && !N1C->isNullValue()) {
SDOperand Div = DAG.getNode(ISD::UDIV, VT, N0, N1);
- SDOperand Mul = DAG.getNode(ISD::MUL, VT, Div, N1);
- SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul);
- AddToWorkList(Div.Val);
- AddToWorkList(Mul.Val);
- return Sub;
+ SDOperand OptimizedDiv = combine(Div.Val);
+ if (OptimizedDiv.Val && OptimizedDiv.Val != Div.Val) {
+ SDOperand Mul = DAG.getNode(ISD::MUL, VT, OptimizedDiv, N1);
+ SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul);
+ AddToWorkList(Mul.Val);
+ return Sub;
+ }
}
// undef % X -> 0
/// compute two values. LoOp and HiOp give the opcodes for the two computations
/// that are being performed. Return true if a simplification was made.
///
-bool DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N,
- unsigned LoOp, unsigned HiOp) {
+SDOperand DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
+ unsigned HiOp) {
// If the high half is not needed, just compute the low half.
- if (!N->hasAnyUseOfValue(1) &&
+ bool HiExists = N->hasAnyUseOfValue(1);
+ if (!HiExists &&
(!AfterLegalize ||
TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
- DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0),
- DAG.getNode(LoOp, N->getValueType(0),
- N->op_begin(),
- N->getNumOperands()));
- return true;
+ SDOperand Res = DAG.getNode(LoOp, N->getValueType(0), N->op_begin(),
+ N->getNumOperands());
+ return CombineTo(N, Res, Res);
}
// If the low half is not needed, just compute the high half.
- if (!N->hasAnyUseOfValue(0) &&
+ bool LoExists = N->hasAnyUseOfValue(0);
+ if (!LoExists &&
(!AfterLegalize ||
TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
- DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1),
- DAG.getNode(HiOp, N->getValueType(1),
- N->op_begin(),
- N->getNumOperands()));
- return true;
+ SDOperand Res = DAG.getNode(HiOp, N->getValueType(1), N->op_begin(),
+ N->getNumOperands());
+ return CombineTo(N, Res, Res);
}
- // If the two computed results can be siplified separately, separate them.
- SDOperand Lo = DAG.getNode(LoOp, N->getValueType(0),
- N->op_begin(), N->getNumOperands());
- SDOperand Hi = DAG.getNode(HiOp, N->getValueType(1),
- N->op_begin(), N->getNumOperands());
- unsigned LoExists = !Lo.use_empty();
- unsigned HiExists = !Hi.use_empty();
- SDOperand LoOpt = Lo;
- SDOperand HiOpt = Hi;
- if (!LoExists || !HiExists) {
- SDOperand Pair = DAG.getNode(ISD::BUILD_PAIR, MVT::Other, Lo, Hi);
- assert(Pair.use_empty() && "Pair with type MVT::Other already exists!");
- LoOpt = combine(Lo.Val);
- HiOpt = combine(Hi.Val);
- if (!LoOpt.Val)
- LoOpt = Pair.getOperand(0);
- if (!HiOpt.Val)
- HiOpt = Pair.getOperand(1);
- DAG.DeleteNode(Pair.Val);
- }
- if ((LoExists || LoOpt != Lo) &&
- (HiExists || HiOpt != Hi) &&
- TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType()) &&
- TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())) {
- DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), LoOpt);
- DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), HiOpt);
- return true;
- }
+ // If both halves are used, return as it is.
+ if (LoExists && HiExists)
+ return SDOperand();
- return false;
+ // If the two computed results can be simplified separately, separate them.
+ if (LoExists) {
+ SDOperand Lo = DAG.getNode(LoOp, N->getValueType(0),
+ N->op_begin(), N->getNumOperands());
+ AddToWorkList(Lo.Val);
+ SDOperand LoOpt = combine(Lo.Val);
+ if (LoOpt.Val && LoOpt.Val != Lo.Val &&
+ TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType()))
+ return CombineTo(N, LoOpt, LoOpt);
+ }
+
+ if (HiExists) {
+ SDOperand Hi = DAG.getNode(HiOp, N->getValueType(1),
+ N->op_begin(), N->getNumOperands());
+ AddToWorkList(Hi.Val);
+ SDOperand HiOpt = combine(Hi.Val);
+ if (HiOpt.Val && HiOpt != Hi &&
+ TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType()))
+ return CombineTo(N, HiOpt, HiOpt);
+ }
+ return SDOperand();
}
SDOperand DAGCombiner::visitSMUL_LOHI(SDNode *N) {
-
- if (SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS))
- return SDOperand();
+ SDOperand Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
+ if (Res.Val) return Res;
return SDOperand();
}
SDOperand DAGCombiner::visitUMUL_LOHI(SDNode *N) {
-
- if (SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU))
- return SDOperand();
+ SDOperand Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
+ if (Res.Val) return Res;
return SDOperand();
}
SDOperand DAGCombiner::visitSDIVREM(SDNode *N) {
-
- if (SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM))
- return SDOperand();
+ SDOperand Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
+ if (Res.Val) return Res;
return SDOperand();
}
SDOperand DAGCombiner::visitUDIVREM(SDNode *N) {
-
- if (SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM))
- return SDOperand();
+ SDOperand Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
+ if (Res.Val) return Res;
return SDOperand();
}
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
MVT::ValueType VT = N1.getValueType();
+ unsigned BitWidth = MVT::getSizeInBits(VT);
// fold vector ops
if (MVT::isVector(VT)) {
if (N1C && N1C->isAllOnesValue())
return N0;
// if (and x, c) is known to be zero, return 0
- if (N1C && DAG.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
+ if (N1C && DAG.MaskedValueIsZero(SDOperand(N, 0),
+ APInt::getAllOnesValue(BitWidth)))
return DAG.getConstant(0, VT);
// reassociate and
SDOperand RAND = ReassociateOps(ISD::AND, N0, N1);
return N1;
// fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
- unsigned InMask = MVT::getIntVTBitMask(N0.getOperand(0).getValueType());
- if (DAG.MaskedValueIsZero(N0.getOperand(0),
- ~N1C->getValue() & InMask)) {
+ SDOperand N0Op0 = N0.getOperand(0);
+ APInt Mask = ~N1C->getAPIntValue();
+ Mask.trunc(N0Op0.getValueSizeInBits());
+ if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
SDOperand Zext = DAG.getNode(ISD::ZERO_EXTEND, N0.getValueType(),
- N0.getOperand(0));
+ N0Op0);
// Replace uses of the AND with uses of the Zero extend node.
CombineTo(N, Zext);
// fold (zext_inreg (extload x)) -> (zextload x)
if (ISD::isEXTLoad(N0.Val) && ISD::isUNINDEXEDLoad(N0.Val)) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT::ValueType EVT = LN0->getLoadedVT();
+ MVT::ValueType EVT = LN0->getMemoryVT();
// If we zero all the possible extended bits, then we can turn this into
// a zextload if we are running before legalize or the operation is legal.
- if (DAG.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
+ unsigned BitWidth = N1.getValueSizeInBits();
+ if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
+ BitWidth - MVT::getSizeInBits(EVT))) &&
(!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
if (ISD::isSEXTLoad(N0.Val) && ISD::isUNINDEXEDLoad(N0.Val) &&
N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT::ValueType EVT = LN0->getLoadedVT();
+ MVT::ValueType EVT = LN0->getMemoryVT();
// If we zero all the possible extended bits, then we can turn this into
// a zextload if we are running before legalize or the operation is legal.
- if (DAG.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
+ unsigned BitWidth = N1.getValueSizeInBits();
+ if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
+ BitWidth - MVT::getSizeInBits(EVT))) &&
(!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
if (N1C && N0.getOpcode() == ISD::LOAD) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
if (LN0->getExtensionType() != ISD::SEXTLOAD &&
- LN0->getAddressingMode() == ISD::UNINDEXED &&
- N0.hasOneUse()) {
+ LN0->isUnindexed() && N0.hasOneUse()) {
MVT::ValueType EVT, LoadedVT;
if (N1C->getValue() == 255)
EVT = MVT::i8;
else
EVT = MVT::Other;
- LoadedVT = LN0->getLoadedVT();
+ LoadedVT = LN0->getMemoryVT();
if (EVT != MVT::Other && LoadedVT > EVT &&
(!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
MVT::ValueType PtrType = N0.getOperand(1).getValueType();
// For big endian targets, we need to add an offset to the pointer to
// load the correct bytes. For little endian systems, we merely need to
// read fewer bytes from the same pointer.
- unsigned PtrOff =
- (MVT::getSizeInBits(LoadedVT) - MVT::getSizeInBits(EVT)) / 8;
+ unsigned LVTStoreBytes = MVT::getStoreSizeInBits(LoadedVT)/8;
+ unsigned EVTStoreBytes = MVT::getStoreSizeInBits(EVT)/8;
+ unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
+ unsigned Alignment = LN0->getAlignment();
SDOperand NewPtr = LN0->getBasePtr();
- if (!TLI.isLittleEndian())
+ if (TLI.isBigEndian()) {
NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr,
DAG.getConstant(PtrOff, PtrType));
+ Alignment = MinAlign(Alignment, PtrOff);
+ }
AddToWorkList(NewPtr.Val);
SDOperand Load =
DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(), NewPtr,
LN0->getSrcValue(), LN0->getSrcValueOffset(), EVT,
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), Alignment);
AddToWorkList(N);
CombineTo(N0.Val, Load, Load.getValue(1));
return SDOperand(N, 0); // Return N so it doesn't get rechecked!
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
MVT::ValueType VT = N1.getValueType();
- unsigned OpSizeInBits = MVT::getSizeInBits(VT);
// fold vector ops
if (MVT::isVector(VT)) {
if (N1C && N1C->isAllOnesValue())
return N1;
// fold (or x, c) -> c iff (x & ~c) == 0
- if (N1C &&
- DAG.MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits))))
+ if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
return N1;
// reassociate or
SDOperand ROR = ReassociateOps(ISD::OR, N0, N1);
(N0.Val->hasOneUse() || N1.Val->hasOneUse())) {
// We can only do this xform if we know that bits from X that are set in C2
// but not in C1 are already zero. Likewise for Y.
- uint64_t LHSMask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
- uint64_t RHSMask = cast<ConstantSDNode>(N1.getOperand(1))->getValue();
+ const APInt &LHSMask =
+ cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
+ const APInt &RHSMask =
+ cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
// If there is an AND of either shifted operand, apply it to the result.
if (LHSMask.Val || RHSMask.Val) {
- uint64_t Mask = MVT::getIntVTBitMask(VT);
+ APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
if (LHSMask.Val) {
- uint64_t RHSBits = (1ULL << LShVal)-1;
- Mask &= cast<ConstantSDNode>(LHSMask)->getValue() | RHSBits;
+ APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
+ Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
}
if (RHSMask.Val) {
- uint64_t LHSBits = ~((1ULL << (OpSizeInBits-RShVal))-1);
- Mask &= cast<ConstantSDNode>(RHSMask)->getValue() | LHSBits;
+ APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
+ Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
}
Rot = DAG.getNode(ISD::AND, VT, Rot, DAG.getConstant(Mask, VT));
LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
if (ConstantSDNode *SUBC =
dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
- if (SUBC->getValue() == OpSizeInBits)
+ if (SUBC->getValue() == OpSizeInBits) {
if (HasROTL)
return DAG.getNode(ISD::ROTL, VT, LHSShiftArg, LHSShiftAmt).Val;
else
return DAG.getNode(ISD::ROTR, VT, LHSShiftArg, RHSShiftAmt).Val;
+ }
}
}
RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
if (ConstantSDNode *SUBC =
dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
- if (SUBC->getValue() == OpSizeInBits)
+ if (SUBC->getValue() == OpSizeInBits) {
if (HasROTL)
return DAG.getNode(ISD::ROTL, VT, LHSShiftArg, LHSShiftAmt).Val;
else
return DAG.getNode(ISD::ROTR, VT, LHSShiftArg, RHSShiftAmt).Val;
+ }
}
}
return SDOperand();
}
+/// visitShiftByConstant - Handle transforms common to the three shifts, when
+/// the shift amount is a constant.
+SDOperand DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
+ SDNode *LHS = N->getOperand(0).Val;
+ if (!LHS->hasOneUse()) return SDOperand();
+
+ // We want to pull some binops through shifts, so that we have (and (shift))
+ // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
+ // thing happens with address calculations, so it's important to canonicalize
+ // it.
+ bool HighBitSet = false; // Can we transform this if the high bit is set?
+
+ switch (LHS->getOpcode()) {
+ default: return SDOperand();
+ case ISD::OR:
+ case ISD::XOR:
+ HighBitSet = false; // We can only transform sra if the high bit is clear.
+ break;
+ case ISD::AND:
+ HighBitSet = true; // We can only transform sra if the high bit is set.
+ break;
+ case ISD::ADD:
+ if (N->getOpcode() != ISD::SHL)
+ return SDOperand(); // only shl(add) not sr[al](add).
+ HighBitSet = false; // We can only transform sra if the high bit is clear.
+ break;
+ }
+
+ // We require the RHS of the binop to be a constant as well.
+ ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
+ if (!BinOpCst) return SDOperand();
+
+
+ // FIXME: disable this for unless the input to the binop is a shift by a
+ // constant. If it is not a shift, it pessimizes some common cases like:
+ //
+ //void foo(int *X, int i) { X[i & 1235] = 1; }
+ //int bar(int *X, int i) { return X[i & 255]; }
+ SDNode *BinOpLHSVal = LHS->getOperand(0).Val;
+ if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
+ BinOpLHSVal->getOpcode() != ISD::SRA &&
+ BinOpLHSVal->getOpcode() != ISD::SRL) ||
+ !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
+ return SDOperand();
+
+ MVT::ValueType VT = N->getValueType(0);
+
+ // If this is a signed shift right, and the high bit is modified
+ // by the logical operation, do not perform the transformation.
+ // The highBitSet boolean indicates the value of the high bit of
+ // the constant which would cause it to be modified for this
+ // operation.
+ if (N->getOpcode() == ISD::SRA) {
+ bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
+ if (BinOpRHSSignSet != HighBitSet)
+ return SDOperand();
+ }
+
+ // Fold the constants, shifting the binop RHS by the shift amount.
+ SDOperand NewRHS = DAG.getNode(N->getOpcode(), N->getValueType(0),
+ LHS->getOperand(1), N->getOperand(1));
+
+ // Create the new shift.
+ SDOperand NewShift = DAG.getNode(N->getOpcode(), VT, LHS->getOperand(0),
+ N->getOperand(1));
+
+ // Create the new binop.
+ return DAG.getNode(LHS->getOpcode(), VT, NewShift, NewRHS);
+}
+
+
SDOperand DAGCombiner::visitSHL(SDNode *N) {
SDOperand N0 = N->getOperand(0);
SDOperand N1 = N->getOperand(1);
if (N1C && N1C->isNullValue())
return N0;
// if (shl x, c) is known to be zero, return 0
- if (DAG.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
+ if (DAG.MaskedValueIsZero(SDOperand(N, 0),
+ APInt::getAllOnesValue(MVT::getSizeInBits(VT))))
return DAG.getConstant(0, VT);
if (N1C && SimplifyDemandedBits(SDOperand(N, 0)))
return SDOperand(N, 0);
if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1))
return DAG.getNode(ISD::AND, VT, N0.getOperand(0),
DAG.getConstant(~0ULL << N1C->getValue(), VT));
- return SDOperand();
+
+ return N1C ? visitShiftByConstant(N, N1C->getValue()) : SDOperand();
}
SDOperand DAGCombiner::visitSRA(SDNode *N) {
// If the sign bit is known to be zero, switch this to a SRL.
- if (DAG.MaskedValueIsZero(N0, MVT::getIntVTSignBit(VT)))
+ if (DAG.SignBitIsZero(N0))
return DAG.getNode(ISD::SRL, VT, N0, N1);
- return SDOperand();
+
+ return N1C ? visitShiftByConstant(N, N1C->getValue()) : SDOperand();
}
SDOperand DAGCombiner::visitSRL(SDNode *N) {
if (N1C && N1C->isNullValue())
return N0;
// if (srl x, c) is known to be zero, return 0
- if (N1C && DAG.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits)))
+ if (N1C && DAG.MaskedValueIsZero(SDOperand(N, 0),
+ APInt::getAllOnesValue(OpSizeInBits)))
return DAG.getConstant(0, VT);
// fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2)
// fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
if (N1C && N0.getOpcode() == ISD::CTLZ &&
N1C->getValue() == Log2_32(MVT::getSizeInBits(VT))) {
- uint64_t KnownZero, KnownOne, Mask = MVT::getIntVTBitMask(VT);
+ APInt KnownZero, KnownOne;
+ APInt Mask = APInt::getAllOnesValue(MVT::getSizeInBits(VT));
DAG.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne);
// If any of the input bits are KnownOne, then the input couldn't be all
// zeros, thus the result of the srl will always be zero.
- if (KnownOne) return DAG.getConstant(0, VT);
+ if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
// If all of the bits input the to ctlz node are known to be zero, then
// the result of the ctlz is "32" and the result of the shift is one.
- uint64_t UnknownBits = ~KnownZero & Mask;
+ APInt UnknownBits = ~KnownZero & Mask;
if (UnknownBits == 0) return DAG.getConstant(1, VT);
// Otherwise, check to see if there is exactly one bit input to the ctlz.
// could be set on input to the CTLZ node. If this bit is set, the SRL
// will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
// to an SRL,XOR pair, which is likely to simplify more.
- unsigned ShAmt = CountTrailingZeros_64(UnknownBits);
+ unsigned ShAmt = UnknownBits.countTrailingZeros();
SDOperand Op = N0.getOperand(0);
if (ShAmt) {
Op = DAG.getNode(ISD::SRL, VT, Op,
if (N1C && SimplifyDemandedBits(SDOperand(N, 0)))
return SDOperand(N, 0);
- return SDOperand();
+ return N1C ? visitShiftByConstant(N, N1C->getValue()) : SDOperand();
}
SDOperand DAGCombiner::visitCTLZ(SDNode *N) {
return DAG.getNode(ISD::TRUNCATE, VT, XORNode);
}
// fold select C, 0, X -> ~C & X
- if (VT == VT0 && N1C && N1C->isNullValue()) {
+ if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
AddToWorkList(XORNode.Val);
return DAG.getNode(ISD::AND, VT, XORNode, N2);
}
// fold select C, X, 1 -> ~C | X
- if (VT == VT0 && N2C && N2C->getValue() == 1) {
+ if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getValue() == 1) {
SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
AddToWorkList(XORNode.Val);
return DAG.getNode(ISD::OR, VT, XORNode, N1);
return SDOperand(N, 0); // Don't revisit N.
// fold selects based on a setcc into other things, such as min/max/abs
- if (N0.getOpcode() == ISD::SETCC)
+ if (N0.getOpcode() == ISD::SETCC) {
// FIXME:
// Check against MVT::Other for SELECT_CC, which is a workaround for targets
// having to say they don't support SELECT_CC on every type the DAG knows
N1, N2, N0.getOperand(2));
else
return SimplifySelect(N0, N1, N2);
+ }
return SDOperand();
}
cast<CondCodeSDNode>(N->getOperand(2))->get());
}
+// ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
+// "fold ({s|z}ext (load x)) -> ({s|z}ext (truncate ({s|z}extload x)))"
+// transformation. Returns true if extension are possible and the above
+// mentioned transformation is profitable.
+static bool ExtendUsesToFormExtLoad(SDNode *N, SDOperand N0,
+ unsigned ExtOpc,
+ SmallVector<SDNode*, 4> &ExtendNodes,
+ TargetLowering &TLI) {
+ bool HasCopyToRegUses = false;
+ bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
+ for (SDNode::use_iterator UI = N0.Val->use_begin(), UE = N0.Val->use_end();
+ UI != UE; ++UI) {
+ SDNode *User = *UI;
+ if (User == N)
+ continue;
+ // FIXME: Only extend SETCC N, N and SETCC N, c for now.
+ if (User->getOpcode() == ISD::SETCC) {
+ ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
+ if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
+ // Sign bits will be lost after a zext.
+ return false;
+ bool Add = false;
+ for (unsigned i = 0; i != 2; ++i) {
+ SDOperand UseOp = User->getOperand(i);
+ if (UseOp == N0)
+ continue;
+ if (!isa<ConstantSDNode>(UseOp))
+ return false;
+ Add = true;
+ }
+ if (Add)
+ ExtendNodes.push_back(User);
+ } else {
+ for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
+ SDOperand UseOp = User->getOperand(i);
+ if (UseOp == N0) {
+ // If truncate from extended type to original load type is free
+ // on this target, then it's ok to extend a CopyToReg.
+ if (isTruncFree && User->getOpcode() == ISD::CopyToReg)
+ HasCopyToRegUses = true;
+ else
+ return false;
+ }
+ }
+ }
+ }
+
+ if (HasCopyToRegUses) {
+ bool BothLiveOut = false;
+ for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
+ UI != UE; ++UI) {
+ SDNode *User = *UI;
+ for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
+ SDOperand UseOp = User->getOperand(i);
+ if (UseOp.Val == N && UseOp.ResNo == 0) {
+ BothLiveOut = true;
+ break;
+ }
+ }
+ }
+ if (BothLiveOut)
+ // Both unextended and extended values are live out. There had better be
+ // good a reason for the transformation.
+ return ExtendNodes.size();
+ }
+ return true;
+}
+
SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
SDOperand N0 = N->getOperand(0);
MVT::ValueType VT = N->getValueType(0);
}
// fold (sext (load x)) -> (sext (truncate (sextload x)))
- if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
+ if (ISD::isNON_EXTLoad(N0.Val) &&
(!AfterLegalize||TLI.isLoadXLegal(ISD::SEXTLOAD, N0.getValueType()))){
- LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(),
- N0.getValueType(),
- LN0->isVolatile(),
- LN0->getAlignment());
- CombineTo(N, ExtLoad);
- CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
- ExtLoad.getValue(1));
- return SDOperand(N, 0); // Return N so it doesn't get rechecked!
+ bool DoXform = true;
+ SmallVector<SDNode*, 4> SetCCs;
+ if (!N0.hasOneUse())
+ DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
+ if (DoXform) {
+ LoadSDNode *LN0 = cast<LoadSDNode>(N0);
+ SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
+ LN0->getBasePtr(), LN0->getSrcValue(),
+ LN0->getSrcValueOffset(),
+ N0.getValueType(),
+ LN0->isVolatile(),
+ LN0->getAlignment());
+ CombineTo(N, ExtLoad);
+ SDOperand Trunc = DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad);
+ CombineTo(N0.Val, Trunc, ExtLoad.getValue(1));
+ // Extend SetCC uses if necessary.
+ for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
+ SDNode *SetCC = SetCCs[i];
+ SmallVector<SDOperand, 4> Ops;
+ for (unsigned j = 0; j != 2; ++j) {
+ SDOperand SOp = SetCC->getOperand(j);
+ if (SOp == Trunc)
+ Ops.push_back(ExtLoad);
+ else
+ Ops.push_back(DAG.getNode(ISD::SIGN_EXTEND, VT, SOp));
+ }
+ Ops.push_back(SetCC->getOperand(2));
+ CombineTo(SetCC, DAG.getNode(ISD::SETCC, SetCC->getValueType(0),
+ &Ops[0], Ops.size()));
+ }
+ return SDOperand(N, 0); // Return N so it doesn't get rechecked!
+ }
}
// fold (sext (sextload x)) -> (sext (truncate (sextload x)))
if ((ISD::isSEXTLoad(N0.Val) || ISD::isEXTLoad(N0.Val)) &&
ISD::isUNINDEXEDLoad(N0.Val) && N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT::ValueType EVT = LN0->getLoadedVT();
+ MVT::ValueType EVT = LN0->getMemoryVT();
if (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT)) {
SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
} else if (X.getValueType() > VT) {
X = DAG.getNode(ISD::TRUNCATE, VT, X);
}
- uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
+ APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
+ Mask.zext(MVT::getSizeInBits(VT));
return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT));
}
// fold (zext (load x)) -> (zext (truncate (zextload x)))
- if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
+ if (ISD::isNON_EXTLoad(N0.Val) &&
(!AfterLegalize||TLI.isLoadXLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
- LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(),
- N0.getValueType(),
- LN0->isVolatile(),
- LN0->getAlignment());
- CombineTo(N, ExtLoad);
- CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
- ExtLoad.getValue(1));
- return SDOperand(N, 0); // Return N so it doesn't get rechecked!
+ bool DoXform = true;
+ SmallVector<SDNode*, 4> SetCCs;
+ if (!N0.hasOneUse())
+ DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
+ if (DoXform) {
+ LoadSDNode *LN0 = cast<LoadSDNode>(N0);
+ SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
+ LN0->getBasePtr(), LN0->getSrcValue(),
+ LN0->getSrcValueOffset(),
+ N0.getValueType(),
+ LN0->isVolatile(),
+ LN0->getAlignment());
+ CombineTo(N, ExtLoad);
+ SDOperand Trunc = DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad);
+ CombineTo(N0.Val, Trunc, ExtLoad.getValue(1));
+ // Extend SetCC uses if necessary.
+ for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
+ SDNode *SetCC = SetCCs[i];
+ SmallVector<SDOperand, 4> Ops;
+ for (unsigned j = 0; j != 2; ++j) {
+ SDOperand SOp = SetCC->getOperand(j);
+ if (SOp == Trunc)
+ Ops.push_back(ExtLoad);
+ else
+ Ops.push_back(DAG.getNode(ISD::ZERO_EXTEND, VT, SOp));
+ }
+ Ops.push_back(SetCC->getOperand(2));
+ CombineTo(SetCC, DAG.getNode(ISD::SETCC, SetCC->getValueType(0),
+ &Ops[0], Ops.size()));
+ }
+ return SDOperand(N, 0); // Return N so it doesn't get rechecked!
+ }
}
// fold (zext (zextload x)) -> (zext (truncate (zextload x)))
if ((ISD::isZEXTLoad(N0.Val) || ISD::isEXTLoad(N0.Val)) &&
ISD::isUNINDEXEDLoad(N0.Val) && N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT::ValueType EVT = LN0->getLoadedVT();
+ MVT::ValueType EVT = LN0->getMemoryVT();
SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(), EVT,
} else if (X.getValueType() > VT) {
X = DAG.getNode(ISD::TRUNCATE, VT, X);
}
- uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
+ APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
+ Mask.zext(MVT::getSizeInBits(VT));
return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT));
}
!ISD::isNON_EXTLoad(N0.Val) && ISD::isUNINDEXEDLoad(N0.Val) &&
N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT::ValueType EVT = LN0->getLoadedVT();
+ MVT::ValueType EVT = LN0->getMemoryVT();
SDOperand ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), VT,
LN0->getChain(), LN0->getBasePtr(),
LN0->getSrcValue(),
/// GetDemandedBits - See if the specified operand can be simplified with the
/// knowledge that only the bits specified by Mask are used. If so, return the
/// simpler operand, otherwise return a null SDOperand.
-SDOperand DAGCombiner::GetDemandedBits(SDOperand V, uint64_t Mask) {
+SDOperand DAGCombiner::GetDemandedBits(SDOperand V, const APInt &Mask) {
switch (V.getOpcode()) {
default: break;
case ISD::OR:
if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
// See if we can recursively simplify the LHS.
unsigned Amt = RHSC->getValue();
- Mask = (Mask << Amt) & MVT::getIntVTBitMask(V.getValueType());
- SDOperand SimplifyLHS = GetDemandedBits(V.getOperand(0), Mask);
+ APInt NewMask = Mask << Amt;
+ SDOperand SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
if (SimplifyLHS.Val) {
return DAG.getNode(ISD::SRL, V.getValueType(),
SimplifyLHS, V.getOperand(1));
MVT::ValueType PtrType = N0.getOperand(1).getValueType();
// For big endian targets, we need to adjust the offset to the pointer to
// load the correct bytes.
- if (!TLI.isLittleEndian())
- ShAmt = MVT::getSizeInBits(N0.getValueType()) - ShAmt - EVTBits;
+ if (TLI.isBigEndian()) {
+ unsigned LVTStoreBits = MVT::getStoreSizeInBits(N0.getValueType());
+ unsigned EVTStoreBits = MVT::getStoreSizeInBits(EVT);
+ ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
+ }
uint64_t PtrOff = ShAmt / 8;
+ unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
SDOperand NewPtr = DAG.getNode(ISD::ADD, PtrType, LN0->getBasePtr(),
DAG.getConstant(PtrOff, PtrType));
AddToWorkList(NewPtr.Val);
SDOperand Load = (ExtType == ISD::NON_EXTLOAD)
? DAG.getLoad(VT, LN0->getChain(), NewPtr,
LN0->getSrcValue(), LN0->getSrcValueOffset(),
- LN0->isVolatile(), LN0->getAlignment())
+ LN0->isVolatile(), NewAlign)
: DAG.getExtLoad(ExtType, VT, LN0->getChain(), NewPtr,
LN0->getSrcValue(), LN0->getSrcValueOffset(), EVT,
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), NewAlign);
AddToWorkList(N);
if (CombineSRL) {
- DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1),
+ &DeadNodes);
CombineTo(N->getOperand(0).Val, Load);
} else
CombineTo(N0.Val, Load, Load.getValue(1));
SDOperand N1 = N->getOperand(1);
MVT::ValueType VT = N->getValueType(0);
MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT();
+ unsigned VTBits = MVT::getSizeInBits(VT);
unsigned EVTBits = MVT::getSizeInBits(EVT);
// fold (sext_in_reg c1) -> c1
}
// fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
- if (DAG.MaskedValueIsZero(N0, 1ULL << (EVTBits-1)))
+ if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
return DAG.getZeroExtendInReg(N0, EVT);
// fold operands of sext_in_reg based on knowledge that the top bits are not
// fold (sext_inreg (extload x)) -> (sextload x)
if (ISD::isEXTLoad(N0.Val) &&
ISD::isUNINDEXEDLoad(N0.Val) &&
- EVT == cast<LoadSDNode>(N0)->getLoadedVT() &&
+ EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
(!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
// fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
if (ISD::isZEXTLoad(N0.Val) && ISD::isUNINDEXEDLoad(N0.Val) &&
N0.hasOneUse() &&
- EVT == cast<LoadSDNode>(N0)->getLoadedVT() &&
+ EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
(!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
// See if we can simplify the input to this truncate through knowledge that
// only the low bits are being used. For example "trunc (or (shl x, 8), y)"
// -> trunc y
- SDOperand Shorter = GetDemandedBits(N0, MVT::getIntVTBitMask(VT));
+ SDOperand Shorter =
+ GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
+ MVT::getSizeInBits(VT)));
if (Shorter.Val)
return DAG.getNode(ISD::TRUNCATE, VT, Shorter);
}
}
+ // Fold bitconvert(fneg(x)) -> xor(bitconvert(x), signbit)
+ // Fold bitconvert(fabs(x)) -> and(bitconvert(x), ~signbit)
+ // This often reduces constant pool loads.
+ if ((N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FABS) &&
+ N0.Val->hasOneUse() && MVT::isInteger(VT) && !MVT::isVector(VT)) {
+ SDOperand NewConv = DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0));
+ AddToWorkList(NewConv.Val);
+
+ APInt SignBit = APInt::getSignBit(MVT::getSizeInBits(VT));
+ if (N0.getOpcode() == ISD::FNEG)
+ return DAG.getNode(ISD::XOR, VT, NewConv, DAG.getConstant(SignBit, VT));
+ assert(N0.getOpcode() == ISD::FABS);
+ return DAG.getNode(ISD::AND, VT, NewConv, DAG.getConstant(~SignBit, VT));
+ }
+
+ // Fold bitconvert(fcopysign(cst, x)) -> bitconvert(x)&sign | cst&~sign'
+ // Note that we don't handle copysign(x,cst) because this can always be folded
+ // to an fneg or fabs.
+ if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse() &&
+ isa<ConstantFPSDNode>(N0.getOperand(0)) &&
+ MVT::isInteger(VT) && !MVT::isVector(VT)) {
+ unsigned OrigXWidth = MVT::getSizeInBits(N0.getOperand(1).getValueType());
+ SDOperand X = DAG.getNode(ISD::BIT_CONVERT, MVT::getIntegerType(OrigXWidth),
+ N0.getOperand(1));
+ AddToWorkList(X.Val);
+
+ // If X has a different width than the result/lhs, sext it or truncate it.
+ unsigned VTWidth = MVT::getSizeInBits(VT);
+ if (OrigXWidth < VTWidth) {
+ X = DAG.getNode(ISD::SIGN_EXTEND, VT, X);
+ AddToWorkList(X.Val);
+ } else if (OrigXWidth > VTWidth) {
+ // To get the sign bit in the right place, we have to shift it right
+ // before truncating.
+ X = DAG.getNode(ISD::SRL, X.getValueType(), X,
+ DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
+ AddToWorkList(X.Val);
+ X = DAG.getNode(ISD::TRUNCATE, VT, X);
+ AddToWorkList(X.Val);
+ }
+
+ APInt SignBit = APInt::getSignBit(MVT::getSizeInBits(VT));
+ X = DAG.getNode(ISD::AND, VT, X, DAG.getConstant(SignBit, VT));
+ AddToWorkList(X.Val);
+
+ SDOperand Cst = DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0));
+ Cst = DAG.getNode(ISD::AND, VT, Cst, DAG.getConstant(~SignBit, VT));
+ AddToWorkList(Cst.Val);
+
+ return DAG.getNode(ISD::OR, VT, X, Cst);
+ }
+
return SDOperand();
}
for (unsigned i = 0, e = BV->getNumOperands(); i != e;
i += NumInputsPerOutput) {
bool isLE = TLI.isLittleEndian();
- uint64_t NewBits = 0;
+ APInt NewBits = APInt(DstBitSize, 0);
bool EltIsUndef = true;
for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
// Shift the previously computed bits over.
if (Op.getOpcode() == ISD::UNDEF) continue;
EltIsUndef = false;
- NewBits |= cast<ConstantSDNode>(Op)->getValue();
+ NewBits |=
+ APInt(cast<ConstantSDNode>(Op)->getAPIntValue()).zext(DstBitSize);
}
if (EltIsUndef)
Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
}
- MVT::ValueType VT = MVT::getVectorType(DstEltVT,
- Ops.size());
+ MVT::ValueType VT = MVT::getVectorType(DstEltVT, Ops.size());
return DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size());
}
// Finally, this must be the case where we are shrinking elements: each input
// turns into multiple outputs.
+ bool isS2V = ISD::isScalarToVector(BV);
unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
+ MVT::ValueType VT = MVT::getVectorType(DstEltVT,
+ NumOutputsPerInput * BV->getNumOperands());
SmallVector<SDOperand, 8> Ops;
for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
continue;
}
- uint64_t OpVal = cast<ConstantSDNode>(BV->getOperand(i))->getValue();
-
+ APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->getAPIntValue();
for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
- unsigned ThisVal = OpVal & ((1ULL << DstBitSize)-1);
- OpVal >>= DstBitSize;
+ APInt ThisVal = APInt(OpVal).trunc(DstBitSize);
Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
+ if (isS2V && i == 0 && j == 0 && APInt(ThisVal).zext(SrcBitSize) == OpVal)
+ // Simply turn this into a SCALAR_TO_VECTOR of the new type.
+ return DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Ops[0]);
+ OpVal = OpVal.lshr(DstBitSize);
}
// For big endian targets, swap the order of the pieces of each element.
- if (!TLI.isLittleEndian())
+ if (TLI.isBigEndian())
std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
}
- MVT::ValueType VT = MVT::getVectorType(DstEltVT, Ops.size());
return DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size());
}
if (N0CFP && !N1CFP)
return DAG.getNode(ISD::FADD, VT, N1, N0);
// fold (A + (-B)) -> A-B
- if (isNegatibleForFree(N1) == 2)
- return DAG.getNode(ISD::FSUB, VT, N0, GetNegatedExpression(N1, DAG));
+ if (isNegatibleForFree(N1, AfterLegalize) == 2)
+ return DAG.getNode(ISD::FSUB, VT, N0,
+ GetNegatedExpression(N1, DAG, AfterLegalize));
// fold ((-A) + B) -> B-A
- if (isNegatibleForFree(N0) == 2)
- return DAG.getNode(ISD::FSUB, VT, N1, GetNegatedExpression(N0, DAG));
+ if (isNegatibleForFree(N0, AfterLegalize) == 2)
+ return DAG.getNode(ISD::FSUB, VT, N1,
+ GetNegatedExpression(N0, DAG, AfterLegalize));
// If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FADD &&
return DAG.getNode(ISD::FSUB, VT, N0, N1);
// fold (0-B) -> -B
if (UnsafeFPMath && N0CFP && N0CFP->getValueAPF().isZero()) {
- if (isNegatibleForFree(N1))
- return GetNegatedExpression(N1, DAG);
+ if (isNegatibleForFree(N1, AfterLegalize))
+ return GetNegatedExpression(N1, DAG, AfterLegalize);
return DAG.getNode(ISD::FNEG, VT, N1);
}
// fold (A-(-B)) -> A+B
- if (isNegatibleForFree(N1))
- return DAG.getNode(ISD::FADD, VT, N0, GetNegatedExpression(N1, DAG));
+ if (isNegatibleForFree(N1, AfterLegalize))
+ return DAG.getNode(ISD::FADD, VT, N0,
+ GetNegatedExpression(N1, DAG, AfterLegalize));
return SDOperand();
}
return DAG.getNode(ISD::FNEG, VT, N0);
// -X * -Y -> X*Y
- if (char LHSNeg = isNegatibleForFree(N0)) {
- if (char RHSNeg = isNegatibleForFree(N1)) {
+ if (char LHSNeg = isNegatibleForFree(N0, AfterLegalize)) {
+ if (char RHSNeg = isNegatibleForFree(N1, AfterLegalize)) {
// Both can be negated for free, check to see if at least one is cheaper
// negated.
if (LHSNeg == 2 || RHSNeg == 2)
- return DAG.getNode(ISD::FMUL, VT, GetNegatedExpression(N0, DAG),
- GetNegatedExpression(N1, DAG));
+ return DAG.getNode(ISD::FMUL, VT,
+ GetNegatedExpression(N0, DAG, AfterLegalize),
+ GetNegatedExpression(N1, DAG, AfterLegalize));
}
}
// -X / -Y -> X*Y
- if (char LHSNeg = isNegatibleForFree(N0)) {
- if (char RHSNeg = isNegatibleForFree(N1)) {
+ if (char LHSNeg = isNegatibleForFree(N0, AfterLegalize)) {
+ if (char RHSNeg = isNegatibleForFree(N1, AfterLegalize)) {
// Both can be negated for free, check to see if at least one is cheaper
// negated.
if (LHSNeg == 2 || RHSNeg == 2)
- return DAG.getNode(ISD::FDIV, VT, GetNegatedExpression(N0, DAG),
- GetNegatedExpression(N1, DAG));
+ return DAG.getNode(ISD::FDIV, VT,
+ GetNegatedExpression(N0, DAG, AfterLegalize),
+ GetNegatedExpression(N1, DAG, AfterLegalize));
}
}
SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) {
SDOperand N0 = N->getOperand(0);
+ SDOperand N1 = N->getOperand(1);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
MVT::ValueType VT = N->getValueType(0);
// fold (fp_round c1fp) -> c1fp
if (N0CFP && N0.getValueType() != MVT::ppcf128)
- return DAG.getNode(ISD::FP_ROUND, VT, N0);
+ return DAG.getNode(ISD::FP_ROUND, VT, N0, N1);
// fold (fp_round (fp_extend x)) -> x
if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
return N0.getOperand(0);
+ // fold (fp_round (fp_round x)) -> (fp_round x)
+ if (N0.getOpcode() == ISD::FP_ROUND) {
+ // This is a value preserving truncation if both round's are.
+ bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
+ N0.Val->getConstantOperandVal(1) == 1;
+ return DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0),
+ DAG.getIntPtrConstant(IsTrunc));
+ }
+
// fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse()) {
- SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0));
+ SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0), N1);
AddToWorkList(Tmp.Val);
return DAG.getNode(ISD::FCOPYSIGN, VT, Tmp, N0.getOperand(1));
}
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
MVT::ValueType VT = N->getValueType(0);
+ // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
+ if (N->hasOneUse() && (*N->use_begin())->getOpcode() == ISD::FP_ROUND)
+ return SDOperand();
+
// fold (fp_extend c1fp) -> c1fp
if (N0CFP && VT != MVT::ppcf128)
return DAG.getNode(ISD::FP_EXTEND, VT, N0);
-
- // fold (fpext (load x)) -> (fpext (fpround (extload x)))
- if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() && VT != MVT::ppcf128 &&
+
+ // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
+ // value of X.
+ if (N0.getOpcode() == ISD::FP_ROUND && N0.Val->getConstantOperandVal(1) == 1){
+ SDOperand In = N0.getOperand(0);
+ if (In.getValueType() == VT) return In;
+ if (VT < In.getValueType())
+ return DAG.getNode(ISD::FP_ROUND, VT, In, N0.getOperand(1));
+ return DAG.getNode(ISD::FP_EXTEND, VT, In);
+ }
+
+ // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
+ if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
(!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(),
LN0->isVolatile(),
LN0->getAlignment());
CombineTo(N, ExtLoad);
- CombineTo(N0.Val, DAG.getNode(ISD::FP_ROUND, N0.getValueType(), ExtLoad),
+ CombineTo(N0.Val, DAG.getNode(ISD::FP_ROUND, N0.getValueType(), ExtLoad,
+ DAG.getIntPtrConstant(1)),
ExtLoad.getValue(1));
return SDOperand(N, 0); // Return N so it doesn't get rechecked!
}
SDOperand DAGCombiner::visitFNEG(SDNode *N) {
SDOperand N0 = N->getOperand(0);
- if (isNegatibleForFree(N0))
- return GetNegatedExpression(N0, DAG);
-
+ if (isNegatibleForFree(N0, AfterLegalize))
+ return GetNegatedExpression(N0, DAG, AfterLegalize);
+
+ // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
+ // constant pool values.
+ if (N0.getOpcode() == ISD::BIT_CONVERT && N0.Val->hasOneUse() &&
+ MVT::isInteger(N0.getOperand(0).getValueType()) &&
+ !MVT::isVector(N0.getOperand(0).getValueType())) {
+ SDOperand Int = N0.getOperand(0);
+ MVT::ValueType IntVT = Int.getValueType();
+ if (MVT::isInteger(IntVT) && !MVT::isVector(IntVT)) {
+ Int = DAG.getNode(ISD::XOR, IntVT, Int,
+ DAG.getConstant(MVT::getIntVTSignBit(IntVT), IntVT));
+ AddToWorkList(Int.Val);
+ return DAG.getNode(ISD::BIT_CONVERT, N->getValueType(0), Int);
+ }
+ }
+
return SDOperand();
}
if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
return DAG.getNode(ISD::FABS, VT, N0.getOperand(0));
+ // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
+ // constant pool values.
+ if (N0.getOpcode() == ISD::BIT_CONVERT && N0.Val->hasOneUse() &&
+ MVT::isInteger(N0.getOperand(0).getValueType()) &&
+ !MVT::isVector(N0.getOperand(0).getValueType())) {
+ SDOperand Int = N0.getOperand(0);
+ MVT::ValueType IntVT = Int.getValueType();
+ if (MVT::isInteger(IntVT) && !MVT::isVector(IntVT)) {
+ Int = DAG.getNode(ISD::AND, IntVT, Int,
+ DAG.getConstant(~MVT::getIntVTSignBit(IntVT), IntVT));
+ AddToWorkList(Int.Val);
+ return DAG.getNode(ISD::BIT_CONVERT, N->getValueType(0), Int);
+ }
+ }
+
return SDOperand();
}
SDOperand Ptr;
MVT::ValueType VT;
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
- if (LD->getAddressingMode() != ISD::UNINDEXED)
+ if (LD->isIndexed())
return false;
- VT = LD->getLoadedVT();
+ VT = LD->getMemoryVT();
if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
!TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
return false;
Ptr = LD->getBasePtr();
} else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
- if (ST->getAddressingMode() != ISD::UNINDEXED)
+ if (ST->isIndexed())
return false;
- VT = ST->getStoredVT();
+ VT = ST->getMemoryVT();
if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
!TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
return false;
// Check #2.
if (!isLoad) {
SDOperand Val = cast<StoreSDNode>(N)->getValue();
- if (Val == BasePtr || BasePtr.Val->isPredecessor(Val.Val))
+ if (Val == BasePtr || BasePtr.Val->isPredecessorOf(Val.Val))
return false;
}
SDNode *Use = *I;
if (Use == N)
continue;
- if (Use->isPredecessor(N))
+ if (Use->isPredecessorOf(N))
return false;
if (!((Use->getOpcode() == ISD::LOAD &&
cast<LoadSDNode>(Use)->getBasePtr() == Ptr) ||
- (Use->getOpcode() == ISD::STORE) &&
- cast<StoreSDNode>(Use)->getBasePtr() == Ptr))
+ (Use->getOpcode() == ISD::STORE &&
+ cast<StoreSDNode>(Use)->getBasePtr() == Ptr)))
RealUse = true;
}
if (!RealUse)
DOUT << "\nReplacing.4 "; DEBUG(N->dump(&DAG));
DOUT << "\nWith: "; DEBUG(Result.Val->dump(&DAG));
DOUT << '\n';
- std::vector<SDNode*> NowDead;
+ WorkListRemover DeadNodes(*this);
if (isLoad) {
DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(0),
- &NowDead);
+ &DeadNodes);
DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), Result.getValue(2),
- &NowDead);
+ &DeadNodes);
} else {
DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(1),
- &NowDead);
+ &DeadNodes);
}
- // Nodes can end up on the worklist more than once. Make sure we do
- // not process a node that has been replaced.
- for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
- removeFromWorkList(NowDead[i]);
// Finally, since the node is now dead, remove it from the graph.
DAG.DeleteNode(N);
// Replace the uses of Ptr with uses of the updated base value.
DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0),
- &NowDead);
+ &DeadNodes);
removeFromWorkList(Ptr.Val);
- for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
- removeFromWorkList(NowDead[i]);
DAG.DeleteNode(Ptr.Val);
return true;
SDOperand Ptr;
MVT::ValueType VT;
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
- if (LD->getAddressingMode() != ISD::UNINDEXED)
+ if (LD->isIndexed())
return false;
- VT = LD->getLoadedVT();
+ VT = LD->getMemoryVT();
if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
!TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
return false;
Ptr = LD->getBasePtr();
} else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
- if (ST->getAddressingMode() != ISD::UNINDEXED)
+ if (ST->isIndexed())
return false;
- VT = ST->getStoredVT();
+ VT = ST->getMemoryVT();
if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
!TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
return false;
SDNode *UseUse = *III;
if (!((UseUse->getOpcode() == ISD::LOAD &&
cast<LoadSDNode>(UseUse)->getBasePtr().Val == Use) ||
- (UseUse->getOpcode() == ISD::STORE) &&
- cast<StoreSDNode>(UseUse)->getBasePtr().Val == Use))
+ (UseUse->getOpcode() == ISD::STORE &&
+ cast<StoreSDNode>(UseUse)->getBasePtr().Val == Use)))
RealUse = true;
}
continue;
// Check for #2
- if (!Op->isPredecessor(N) && !N->isPredecessor(Op)) {
+ if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
SDOperand Result = isLoad
? DAG.getIndexedLoad(SDOperand(N,0), BasePtr, Offset, AM)
: DAG.getIndexedStore(SDOperand(N,0), BasePtr, Offset, AM);
DOUT << "\nReplacing.5 "; DEBUG(N->dump(&DAG));
DOUT << "\nWith: "; DEBUG(Result.Val->dump(&DAG));
DOUT << '\n';
- std::vector<SDNode*> NowDead;
+ WorkListRemover DeadNodes(*this);
if (isLoad) {
DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(0),
- &NowDead);
+ &DeadNodes);
DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), Result.getValue(2),
- &NowDead);
+ &DeadNodes);
} else {
DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(1),
- &NowDead);
+ &DeadNodes);
}
- // Nodes can end up on the worklist more than once. Make sure we do
- // not process a node that has been replaced.
- for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
- removeFromWorkList(NowDead[i]);
// Finally, since the node is now dead, remove it from the graph.
DAG.DeleteNode(N);
// Replace the uses of Use with uses of the updated base value.
DAG.ReplaceAllUsesOfValueWith(SDOperand(Op, 0),
Result.getValue(isLoad ? 1 : 0),
- &NowDead);
+ &DeadNodes);
removeFromWorkList(Op);
- for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
- removeFromWorkList(NowDead[i]);
DAG.DeleteNode(Op);
-
return true;
}
}
return false;
}
+/// InferAlignment - If we can infer some alignment information from this
+/// pointer, return it.
+static unsigned InferAlignment(SDOperand Ptr, SelectionDAG &DAG) {
+ // If this is a direct reference to a stack slot, use information about the
+ // stack slot's alignment.
+ int FrameIdx = 1 << 31;
+ int64_t FrameOffset = 0;
+ if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) {
+ FrameIdx = FI->getIndex();
+ } else if (Ptr.getOpcode() == ISD::ADD &&
+ isa<ConstantSDNode>(Ptr.getOperand(1)) &&
+ isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
+ FrameIdx = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
+ FrameOffset = Ptr.getConstantOperandVal(1);
+ }
+
+ if (FrameIdx != (1 << 31)) {
+ // FIXME: Handle FI+CST.
+ const MachineFrameInfo &MFI = *DAG.getMachineFunction().getFrameInfo();
+ if (MFI.isFixedObjectIndex(FrameIdx)) {
+ int64_t ObjectOffset = MFI.getObjectOffset(FrameIdx);
+
+ // The alignment of the frame index can be determined from its offset from
+ // the incoming frame position. If the frame object is at offset 32 and
+ // the stack is guaranteed to be 16-byte aligned, then we know that the
+ // object is 16-byte aligned.
+ unsigned StackAlign = DAG.getTarget().getFrameInfo()->getStackAlignment();
+ unsigned Align = MinAlign(ObjectOffset, StackAlign);
+
+ // Finally, the frame object itself may have a known alignment. Factor
+ // the alignment + offset into a new alignment. For example, if we know
+ // the FI is 8 byte aligned, but the pointer is 4 off, we really have a
+ // 4-byte alignment of the resultant pointer. Likewise align 4 + 4-byte
+ // offset = 4-byte alignment, align 4 + 1-byte offset = align 1, etc.
+ unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx),
+ FrameOffset);
+ return std::max(Align, FIInfoAlign);
+ }
+ }
+
+ return 0;
+}
SDOperand DAGCombiner::visitLOAD(SDNode *N) {
LoadSDNode *LD = cast<LoadSDNode>(N);
SDOperand Chain = LD->getChain();
SDOperand Ptr = LD->getBasePtr();
+
+ // Try to infer better alignment information than the load already has.
+ if (LD->isUnindexed()) {
+ if (unsigned Align = InferAlignment(Ptr, DAG)) {
+ if (Align > LD->getAlignment())
+ return DAG.getExtLoad(LD->getExtensionType(), LD->getValueType(0),
+ Chain, Ptr, LD->getSrcValue(),
+ LD->getSrcValueOffset(), LD->getMemoryVT(),
+ LD->isVolatile(), Align);
+ }
+ }
+
// If load is not volatile and there are no uses of the loaded value (and
// the updated indexed value in case of indexed loads), change uses of the
if (!LD->isVolatile()) {
if (N->getValueType(1) == MVT::Other) {
// Unindexed loads.
- if (N->hasNUsesOfValue(0, 0))
- return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain);
+ if (N->hasNUsesOfValue(0, 0)) {
+ // It's not safe to use the two value CombineTo variant here. e.g.
+ // v1, chain2 = load chain1, loc
+ // v2, chain3 = load chain2, loc
+ // v3 = add v2, c
+ // Now we replace use of chain2 with chain1. This makes the second load
+ // isomorphic to the one we are deleting, and thus makes this load live.
+ DOUT << "\nReplacing.6 "; DEBUG(N->dump(&DAG));
+ DOUT << "\nWith chain: "; DEBUG(Chain.Val->dump(&DAG));
+ DOUT << "\n";
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), Chain, &DeadNodes);
+ if (N->use_empty()) {
+ removeFromWorkList(N);
+ DAG.DeleteNode(N);
+ }
+ return SDOperand(N, 0); // Return N so it doesn't get rechecked!
+ }
} else {
// Indexed loads.
assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
if (N->hasNUsesOfValue(0, 0) && N->hasNUsesOfValue(0, 1)) {
- SDOperand Undef0 = DAG.getNode(ISD::UNDEF, N->getValueType(0));
- SDOperand Undef1 = DAG.getNode(ISD::UNDEF, N->getValueType(1));
- SDOperand To[] = { Undef0, Undef1, Chain };
- return CombineTo(N, To, 3);
+ SDOperand Undef = DAG.getNode(ISD::UNDEF, N->getValueType(0));
+ DOUT << "\nReplacing.6 "; DEBUG(N->dump(&DAG));
+ DOUT << "\nWith: "; DEBUG(Undef.Val->dump(&DAG));
+ DOUT << " and 2 other values\n";
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Undef, &DeadNodes);
+ DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1),
+ DAG.getNode(ISD::UNDEF, N->getValueType(1)),
+ &DeadNodes);
+ DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 2), Chain, &DeadNodes);
+ removeFromWorkList(N);
+ DAG.DeleteNode(N);
+ return SDOperand(N, 0); // Return N so it doesn't get rechecked!
}
}
}
// Replace the chain to void dependency.
if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
ReplLoad = DAG.getLoad(N->getValueType(0), BetterChain, Ptr,
- LD->getSrcValue(), LD->getSrcValueOffset(),
- LD->isVolatile(), LD->getAlignment());
+ LD->getSrcValue(), LD->getSrcValueOffset(),
+ LD->isVolatile(), LD->getAlignment());
} else {
ReplLoad = DAG.getExtLoad(LD->getExtensionType(),
LD->getValueType(0),
BetterChain, Ptr, LD->getSrcValue(),
LD->getSrcValueOffset(),
- LD->getLoadedVT(),
+ LD->getMemoryVT(),
LD->isVolatile(),
LD->getAlignment());
}
return SDOperand();
}
+
SDOperand DAGCombiner::visitSTORE(SDNode *N) {
StoreSDNode *ST = cast<StoreSDNode>(N);
SDOperand Chain = ST->getChain();
SDOperand Value = ST->getValue();
SDOperand Ptr = ST->getBasePtr();
+ // Try to infer better alignment information than the store already has.
+ if (ST->isUnindexed()) {
+ if (unsigned Align = InferAlignment(Ptr, DAG)) {
+ if (Align > ST->getAlignment())
+ return DAG.getTruncStore(Chain, Value, Ptr, ST->getSrcValue(),
+ ST->getSrcValueOffset(), ST->getMemoryVT(),
+ ST->isVolatile(), Align);
+ }
+ }
+
// If this is a store of a bit convert, store the input value if the
// resultant store does not need a higher alignment than the original.
if (Value.getOpcode() == ISD::BIT_CONVERT && !ST->isTruncatingStore() &&
- ST->getAddressingMode() == ISD::UNINDEXED) {
+ ST->isUnindexed()) {
unsigned Align = ST->getAlignment();
MVT::ValueType SVT = Value.getOperand(0).getValueType();
unsigned OrigAlign = TLI.getTargetMachine().getTargetData()->
ST->getSrcValueOffset(), ST->isVolatile(),
ST->getAlignment());
} else if (TLI.isTypeLegal(MVT::i32)) {
- // Many FP stores are not make apparent until after legalize, e.g. for
+ // Many FP stores are not made apparent until after legalize, e.g. for
// argument passing. Since this is so common, custom legalize the
// 64-bit integer store into two 32-bit stores.
uint64_t Val = CFP->getValueAPF().convertToAPInt().getZExtValue();
SDOperand Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
SDOperand Hi = DAG.getConstant(Val >> 32, MVT::i32);
- if (!TLI.isLittleEndian()) std::swap(Lo, Hi);
+ if (TLI.isBigEndian()) std::swap(Lo, Hi);
int SVOffset = ST->getSrcValueOffset();
unsigned Alignment = ST->getAlignment();
Ptr = DAG.getNode(ISD::ADD, Ptr.getValueType(), Ptr,
DAG.getConstant(4, Ptr.getValueType()));
SVOffset += 4;
- if (Alignment > 4)
- Alignment = 4;
+ Alignment = MinAlign(Alignment, 4U);
SDOperand St1 = DAG.getStore(Chain, Hi, Ptr, ST->getSrcValue(),
SVOffset, isVolatile, Alignment);
return DAG.getNode(ISD::TokenFactor, MVT::Other, St0, St1);
SDOperand ReplStore;
if (ST->isTruncatingStore()) {
ReplStore = DAG.getTruncStore(BetterChain, Value, Ptr,
- ST->getSrcValue(), ST->getSrcValueOffset(), ST->getStoredVT(),
- ST->isVolatile(), ST->getAlignment());
+ ST->getSrcValue(),ST->getSrcValueOffset(),
+ ST->getMemoryVT(),
+ ST->isVolatile(), ST->getAlignment());
} else {
ReplStore = DAG.getStore(BetterChain, Value, Ptr,
- ST->getSrcValue(), ST->getSrcValueOffset(),
- ST->isVolatile(), ST->getAlignment());
+ ST->getSrcValue(), ST->getSrcValueOffset(),
+ ST->isVolatile(), ST->getAlignment());
}
// Create token to keep both nodes around.
if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
return SDOperand(N, 0);
- // FIXME: is there such a think as a truncating indexed store?
- if (ST->isTruncatingStore() && ST->getAddressingMode() == ISD::UNINDEXED &&
+ // FIXME: is there such a thing as a truncating indexed store?
+ if (ST->isTruncatingStore() && ST->isUnindexed() &&
MVT::isInteger(Value.getValueType())) {
// See if we can simplify the input to this truncstore with knowledge that
// only the low bits are being used. For example:
// "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
SDOperand Shorter =
- GetDemandedBits(Value, MVT::getIntVTBitMask(ST->getStoredVT()));
+ GetDemandedBits(Value,
+ APInt::getLowBitsSet(Value.getValueSizeInBits(),
+ MVT::getSizeInBits(ST->getMemoryVT())));
AddToWorkList(Value.Val);
if (Shorter.Val)
return DAG.getTruncStore(Chain, Shorter, Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(), ST->getStoredVT(),
+ ST->getSrcValueOffset(), ST->getMemoryVT(),
ST->isVolatile(), ST->getAlignment());
// Otherwise, see if we can simplify the operation with
// SimplifyDemandedBits, which only works if the value has a single use.
- if (SimplifyDemandedBits(Value, MVT::getIntVTBitMask(ST->getStoredVT())))
+ if (SimplifyDemandedBits(Value,
+ APInt::getLowBitsSet(
+ Value.getValueSizeInBits(),
+ MVT::getSizeInBits(ST->getMemoryVT()))))
return SDOperand(N, 0);
}
+ // If this is a load followed by a store to the same location, then the store
+ // is dead/noop.
+ if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
+ if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
+ ST->isUnindexed() && !ST->isVolatile() &&
+ // There can't be any side effects between the load and store, such as
+ // a call or store.
+ Chain.reachesChainWithoutSideEffects(SDOperand(Ld, 1))) {
+ // The store is dead, remove it.
+ return Chain;
+ }
+ }
+
+ // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
+ // truncating store. We can do this even if this is already a truncstore.
+ if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
+ && TLI.isTypeLegal(Value.getOperand(0).getValueType()) &&
+ Value.Val->hasOneUse() && ST->isUnindexed() &&
+ TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
+ ST->getMemoryVT())) {
+ return DAG.getTruncStore(Chain, Value.getOperand(0), Ptr, ST->getSrcValue(),
+ ST->getSrcValueOffset(), ST->getMemoryVT(),
+ ST->isVolatile(), ST->getAlignment());
+ }
+
return SDOperand();
}
unsigned NumElts = MVT::getVectorNumElements(VT);
if (InVec.getOpcode() == ISD::BIT_CONVERT) {
MVT::ValueType BCVT = InVec.getOperand(0).getValueType();
- if (NumElts != MVT::getVectorNumElements(BCVT))
+ if (!MVT::isVector(BCVT) ||
+ NumElts != MVT::getVectorNumElements(BCVT))
return SDOperand();
InVec = InVec.getOperand(0);
EVT = MVT::getVectorElementType(BCVT);
// Otherwise, use InIdx + VecSize
unsigned Idx = cast<ConstantSDNode>(Extract.getOperand(1))->getValue();
- BuildVecIndices.push_back(DAG.getConstant(Idx+NumInScalars,
- TLI.getPointerTy()));
+ BuildVecIndices.push_back(DAG.getIntPtrConstant(Idx+NumInScalars));
}
// Add count and size info.
- MVT::ValueType BuildVecVT =
- MVT::getVectorType(TLI.getPointerTy(), NumElts);
+ MVT::ValueType BuildVecVT = MVT::getVectorType(TLI.getPointerTy(), NumElts);
// Return the new VECTOR_SHUFFLE node.
SDOperand Ops[5];
LoadSDNode *RLD = cast<LoadSDNode>(RHS);
// If this is an EXTLOAD, the VT's must match.
- if (LLD->getLoadedVT() == RLD->getLoadedVT()) {
+ if (LLD->getMemoryVT() == RLD->getMemoryVT()) {
// FIXME: this conflates two src values, discarding one. This is not
// the right thing to do, but nothing uses srcvalues now. When they do,
// turn SrcValue into a list of locations.
if (TheSelect->getOpcode() == ISD::SELECT) {
// Check that the condition doesn't reach either load. If so, folding
// this will induce a cycle into the DAG.
- if (!LLD->isPredecessor(TheSelect->getOperand(0).Val) &&
- !RLD->isPredecessor(TheSelect->getOperand(0).Val)) {
+ if (!LLD->isPredecessorOf(TheSelect->getOperand(0).Val) &&
+ !RLD->isPredecessorOf(TheSelect->getOperand(0).Val)) {
Addr = DAG.getNode(ISD::SELECT, LLD->getBasePtr().getValueType(),
TheSelect->getOperand(0), LLD->getBasePtr(),
RLD->getBasePtr());
} else {
// Check that the condition doesn't reach either load. If so, folding
// this will induce a cycle into the DAG.
- if (!LLD->isPredecessor(TheSelect->getOperand(0).Val) &&
- !RLD->isPredecessor(TheSelect->getOperand(0).Val) &&
- !LLD->isPredecessor(TheSelect->getOperand(1).Val) &&
- !RLD->isPredecessor(TheSelect->getOperand(1).Val)) {
+ if (!LLD->isPredecessorOf(TheSelect->getOperand(0).Val) &&
+ !RLD->isPredecessorOf(TheSelect->getOperand(0).Val) &&
+ !LLD->isPredecessorOf(TheSelect->getOperand(1).Val) &&
+ !RLD->isPredecessorOf(TheSelect->getOperand(1).Val)) {
Addr = DAG.getNode(ISD::SELECT_CC, LLD->getBasePtr().getValueType(),
TheSelect->getOperand(0),
TheSelect->getOperand(1),
TheSelect->getValueType(0),
LLD->getChain(), Addr, LLD->getSrcValue(),
LLD->getSrcValueOffset(),
- LLD->getLoadedVT(),
+ LLD->getMemoryVT(),
LLD->isVolatile(),
LLD->getAlignment());
}
const Value *&SrcValue, int &SrcValueOffset) {
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
Ptr = LD->getBasePtr();
- Size = MVT::getSizeInBits(LD->getLoadedVT()) >> 3;
+ Size = MVT::getSizeInBits(LD->getMemoryVT()) >> 3;
SrcValue = LD->getSrcValue();
SrcValueOffset = LD->getSrcValueOffset();
return true;
} else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
Ptr = ST->getBasePtr();
- Size = MVT::getSizeInBits(ST->getStoredVT()) >> 3;
+ Size = MVT::getSizeInBits(ST->getMemoryVT()) >> 3;
SrcValue = ST->getSrcValue();
SrcValueOffset = ST->getSrcValueOffset();
} else {
projects / oota-llvm.git / blobdiff