//
// This pass combines dag nodes to form fewer, simpler DAG nodes. It can be run
// both before and after the DAG is legalized.
-//
+//
+// This pass is not a substitute for the LLVM IR instcombine pass. This pass is
+// primarily intended to handle simplification opportunities that are implicit
+// in the LLVM IR and exposed by the various codegen lowering phases.
+//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "dagcombine"
#include "llvm/CodeGen/SelectionDAG.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/LLVMContext.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetFrameInfo.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/Support/Compiler.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
+#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <set>
using namespace llvm;
STATISTIC(NodesCombined , "Number of dag nodes combined");
STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
+STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
namespace {
static cl::opt<bool>
//------------------------------ DAGCombiner ---------------------------------//
- class VISIBILITY_HIDDEN DAGCombiner {
+ class DAGCombiner {
SelectionDAG &DAG;
const TargetLowering &TLI;
CombineLevel Level;
+ CodeGenOpt::Level OptLevel;
bool LegalOperations;
bool LegalTypes;
- bool Fast;
// Worklist of all of the nodes that need to be simplified.
std::vector<SDNode*> WorkList;
WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
WorkList.end());
}
-
+
SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
- bool AddTo = true);
-
+ bool AddTo = true);
+
SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
return CombineTo(N, &Res, 1, AddTo);
}
-
+
SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
- bool AddTo = true) {
+ bool AddTo = true) {
SDValue To[] = { Res0, Res1 };
return CombineTo(N, To, 2, AddTo);
}
void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
-
- private:
-
+
+ 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 CombineToPreIndexedLoadStore(SDNode *N);
bool CombineToPostIndexedLoadStore(SDNode *N);
-
-
+
+
/// combine - call the node-specific routine that knows how to fold each
/// particular type of node. If that doesn't do anything, try the
/// target-specific DAG combines.
SDValue XformToShuffleWithZero(SDNode *N);
SDValue ReassociateOps(unsigned Opc, DebugLoc DL, SDValue LHS, SDValue RHS);
-
+
SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
SDValue SimplifySelect(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2);
- SDValue SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2,
- SDValue N3, ISD::CondCode CC,
+ SDValue SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2,
+ SDValue N3, ISD::CondCode CC,
bool NotExtCompare = false);
- SDValue SimplifySetCC(MVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
+ SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
DebugLoc DL, bool foldBooleans = true);
- SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
+ SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
unsigned HiOp);
- SDValue CombineConsecutiveLoads(SDNode *N, MVT VT);
- SDValue ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *, MVT);
+ SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
+ SDValue ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *, EVT);
SDValue BuildSDIV(SDNode *N);
SDValue BuildUDIV(SDNode *N);
SDNode *MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL);
SDValue ReduceLoadWidth(SDNode *N);
-
+ SDValue ReduceLoadOpStoreWidth(SDNode *N);
+
SDValue GetDemandedBits(SDValue V, const APInt &Mask);
-
+
/// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
/// looking for aliasing nodes and adding them to the Aliases vector.
void GatherAllAliases(SDNode *N, SDValue OriginalChain,
/// overlap.
bool isAlias(SDValue Ptr1, int64_t Size1,
const Value *SrcValue1, int SrcValueOffset1,
+ unsigned SrcValueAlign1,
SDValue Ptr2, int64_t Size2,
- const Value *SrcValue2, int SrcValueOffset2) const;
-
+ const Value *SrcValue2, int SrcValueOffset2,
+ unsigned SrcValueAlign2) const;
+
/// FindAliasInfo - Extracts the relevant alias information from the memory
/// node. Returns true if the operand was a load.
bool FindAliasInfo(SDNode *N,
SDValue &Ptr, int64_t &Size,
- const Value *&SrcValue, int &SrcValueOffset) const;
-
+ const Value *&SrcValue, int &SrcValueOffset,
+ unsigned &SrcValueAlignment) const;
+
/// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
/// looking for a better chain (aliasing node.)
SDValue FindBetterChain(SDNode *N, SDValue Chain);
/// getShiftAmountTy - Returns a type large enough to hold any valid
/// shift amount - before type legalization these can be huge.
- MVT getShiftAmountTy() {
+ EVT getShiftAmountTy() {
return LegalTypes ? TLI.getShiftAmountTy() : TLI.getPointerTy();
}
public:
- DAGCombiner(SelectionDAG &D, AliasAnalysis &A, bool fast)
+ DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
: DAG(D),
TLI(D.getTargetLoweringInfo()),
Level(Unrestricted),
+ OptLevel(OL),
LegalOperations(false),
LegalTypes(false),
- Fast(fast),
AA(A) {}
-
+
/// Run - runs the dag combiner on all nodes in the work list
void Run(CombineLevel AtLevel);
};
namespace {
/// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
/// nodes from the worklist.
-class VISIBILITY_HIDDEN WorkListRemover :
- public SelectionDAG::DAGUpdateListener {
+class WorkListRemover : public SelectionDAG::DAGUpdateListener {
DAGCombiner &DC;
public:
explicit WorkListRemover(DAGCombiner &dc) : DC(dc) {}
-
+
virtual void NodeDeleted(SDNode *N, SDNode *E) {
DC.removeFromWorkList(N);
}
-
+
virtual void NodeUpdated(SDNode *N) {
// Ignore updates.
}
}
SDValue TargetLowering::DAGCombinerInfo::
-CombineTo(SDNode *N, const std::vector<SDValue> &To) {
- return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size());
+CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
+ return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
}
SDValue TargetLowering::DAGCombinerInfo::
-CombineTo(SDNode *N, SDValue Res) {
- return ((DAGCombiner*)DC)->CombineTo(N, Res);
+CombineTo(SDNode *N, SDValue Res, bool AddTo) {
+ return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
}
SDValue TargetLowering::DAGCombinerInfo::
-CombineTo(SDNode *N, SDValue Res0, SDValue Res1) {
- return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1);
+CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
+ return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
}
void TargetLowering::DAGCombinerInfo::
// fneg is removable even if it has multiple uses.
if (Op.getOpcode() == ISD::FNEG) return 2;
-
+
// Don't allow anything with multiple uses.
if (!Op.hasOneUse()) return 0;
-
+
// Don't recurse exponentially.
if (Depth > 6) return 0;
-
+
switch (Op.getOpcode()) {
default: return false;
case ISD::ConstantFP:
case ISD::FADD:
// FIXME: determine better conditions for this xform.
if (!UnsafeFPMath) return 0;
-
+
// fold (fsub (fadd A, B)) -> (fsub (fneg A), B)
if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
return V;
// fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
return isNegatibleForFree(Op.getOperand(1), LegalOperations, Depth+1);
case ISD::FSUB:
- // We can't turn -(A-B) into B-A when we honor signed zeros.
+ // We can't turn -(A-B) into B-A when we honor signed zeros.
if (!UnsafeFPMath) return 0;
-
+
// fold (fneg (fsub A, B)) -> (fsub B, A)
return 1;
-
+
case ISD::FMUL:
case ISD::FDIV:
if (HonorSignDependentRoundingFPMath()) return 0;
-
+
// fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
return V;
-
+
return isNegatibleForFree(Op.getOperand(1), LegalOperations, Depth+1);
-
+
case ISD::FP_EXTEND:
case ISD::FP_ROUND:
case ISD::FSIN:
bool LegalOperations, unsigned Depth = 0) {
// fneg is removable even if it has multiple uses.
if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
-
+
// Don't allow anything with multiple uses.
assert(Op.hasOneUse() && "Unknown reuse!");
-
+
assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
switch (Op.getOpcode()) {
- default: assert(0 && "Unknown code");
+ default: llvm_unreachable("Unknown code");
case ISD::ConstantFP: {
APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
V.changeSign();
case ISD::FADD:
// FIXME: determine better conditions for this xform.
assert(UnsafeFPMath);
-
+
// fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
if (isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
- GetNegatedExpression(Op.getOperand(0), DAG,
+ GetNegatedExpression(Op.getOperand(0), DAG,
LegalOperations, Depth+1),
Op.getOperand(1));
// fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
- GetNegatedExpression(Op.getOperand(1), DAG,
+ GetNegatedExpression(Op.getOperand(1), DAG,
LegalOperations, Depth+1),
Op.getOperand(0));
case ISD::FSUB:
- // We can't turn -(A-B) into B-A when we honor signed zeros.
+ // We can't turn -(A-B) into B-A when we honor signed zeros.
assert(UnsafeFPMath);
// fold (fneg (fsub 0, B)) -> B
if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
if (N0CFP->getValueAPF().isZero())
return Op.getOperand(1);
-
+
// fold (fneg (fsub A, B)) -> (fsub B, A)
return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
Op.getOperand(1), Op.getOperand(0));
-
+
case ISD::FMUL:
case ISD::FDIV:
assert(!HonorSignDependentRoundingFPMath());
-
+
// fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
if (isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
- GetNegatedExpression(Op.getOperand(0), DAG,
+ GetNegatedExpression(Op.getOperand(0), DAG,
LegalOperations, Depth+1),
Op.getOperand(1));
-
+
// fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
Op.getOperand(0),
GetNegatedExpression(Op.getOperand(1), DAG,
LegalOperations, Depth+1));
-
+
case ISD::FP_EXTEND:
case ISD::FSIN:
return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
- GetNegatedExpression(Op.getOperand(0), DAG,
+ GetNegatedExpression(Op.getOperand(0), DAG,
LegalOperations, Depth+1));
case ISD::FP_ROUND:
return DAG.getNode(ISD::FP_ROUND, Op.getDebugLoc(), Op.getValueType(),
- GetNegatedExpression(Op.getOperand(0), DAG,
+ GetNegatedExpression(Op.getOperand(0), DAG,
LegalOperations, Depth+1),
Op.getOperand(1));
}
// isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
// that selects between the values 1 and 0, making it equivalent to a setcc.
-// Also, set the incoming LHS, RHS, and CC references to the appropriate
+// Also, set the incoming LHS, RHS, and CC references to the appropriate
// nodes based on the type of node we are checking. This simplifies life a
// bit for the callers.
static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
CC = N.getOperand(2);
return true;
}
- if (N.getOpcode() == ISD::SELECT_CC &&
+ if (N.getOpcode() == ISD::SELECT_CC &&
N.getOperand(2).getOpcode() == ISD::Constant &&
N.getOperand(3).getOpcode() == ISD::Constant &&
cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
SDValue DAGCombiner::ReassociateOps(unsigned Opc, DebugLoc DL,
SDValue N0, SDValue N1) {
- MVT VT = N0.getValueType();
+ EVT VT = N0.getValueType();
if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
if (isa<ConstantSDNode>(N1)) {
// reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
bool AddTo) {
assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
++NodesCombined;
- DOUT << "\nReplacing.1 "; DEBUG(N->dump(&DAG));
- DOUT << "\nWith: "; DEBUG(To[0].getNode()->dump(&DAG));
- DOUT << " and " << NumTo-1 << " other values\n";
- DEBUG(for (unsigned i = 0, e = NumTo; i != e; ++i)
+ DEBUG(errs() << "\nReplacing.1 ";
+ N->dump(&DAG);
+ errs() << "\nWith: ";
+ To[0].getNode()->dump(&DAG);
+ errs() << " and " << NumTo-1 << " other values\n";
+ for (unsigned i = 0, e = NumTo; i != e; ++i)
assert(N->getValueType(i) == To[i].getValueType() &&
"Cannot combine value to value of different type!"));
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].getNode());
- AddUsersToWorkList(To[i].getNode());
+ if (To[i].getNode()) {
+ AddToWorkList(To[i].getNode());
+ AddUsersToWorkList(To[i].getNode());
+ }
}
}
-
+
// 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.
// 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);
}
void
DAGCombiner::CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &
TLO) {
- // Replace all uses. If any nodes become isomorphic to other nodes and
+ // 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.getNode());
AddUsersToWorkList(TLO.New.getNode());
-
+
// 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.getNode()->use_empty()) {
removeFromWorkList(TLO.Old.getNode());
-
+
// 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.getNode()->getNumOperands(); i != e; ++i)
if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
-
+
DAG.DeleteNode(TLO.Old.getNode());
}
}
APInt KnownZero, KnownOne;
if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
return false;
-
+
// Revisit the node.
AddToWorkList(Op.getNode());
-
+
// Replace the old value with the new one.
++NodesCombined;
- DOUT << "\nReplacing.2 "; DEBUG(TLO.Old.getNode()->dump(&DAG));
- DOUT << "\nWith: "; DEBUG(TLO.New.getNode()->dump(&DAG));
- DOUT << '\n';
-
+ DEBUG(errs() << "\nReplacing.2 ";
+ TLO.Old.getNode()->dump(&DAG);
+ errs() << "\nWith: ";
+ TLO.New.getNode()->dump(&DAG);
+ errs() << '\n');
+
CommitTargetLoweringOpt(TLO);
return true;
}
// to the root node, preventing it from being deleted, and tracking any
// changes of the root.
HandleSDNode Dummy(DAG.getRoot());
-
+
// The root of the dag may dangle to deleted nodes until the dag combiner is
// done. Set it to null to avoid confusion.
DAG.setRoot(SDValue());
-
+
// while the worklist isn't empty, inspect the node on the end of it and
// try and combine it.
while (!WorkList.empty()) {
SDNode *N = WorkList.back();
WorkList.pop_back();
-
+
// If N has no uses, it is dead. Make sure to revisit all N's operands once
// N is deleted from the DAG, since they too may now be dead or may have a
// reduced number of uses, allowing other xforms.
if (N->use_empty() && N != &Dummy) {
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
AddToWorkList(N->getOperand(i).getNode());
-
+
DAG.DeleteNode(N);
continue;
}
-
+
SDValue RV = combine(N);
-
+
if (RV.getNode() == 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
+ // CombineTo was used. Since CombineTo takes care of the worklist
// mechanics for us, we have no work to do in this case.
if (RV.getNode() == N)
continue;
-
+
assert(N->getOpcode() != ISD::DELETED_NODE &&
RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
"Node was deleted but visit returned new node!");
- DOUT << "\nReplacing.3 "; DEBUG(N->dump(&DAG));
- DOUT << "\nWith: "; DEBUG(RV.getNode()->dump(&DAG));
- DOUT << '\n';
+ DEBUG(errs() << "\nReplacing.3 ";
+ N->dump(&DAG);
+ errs() << "\nWith: ";
+ RV.getNode()->dump(&DAG);
+ errs() << '\n');
WorkListRemover DeadNodes(*this);
if (N->getNumValues() == RV.getNode()->getNumValues())
DAG.ReplaceAllUsesWith(N, RV.getNode(), &DeadNodes);
SDValue OpV = RV;
DAG.ReplaceAllUsesWith(N, &OpV, &DeadNodes);
}
-
+
// Push the new node and any users onto the worklist
AddToWorkList(RV.getNode());
AddUsersToWorkList(RV.getNode());
-
+
// 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).getNode());
-
+
// 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.
// 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).
DAG.setRoot(Dummy.getValue());
}
TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
// Expose the DAG combiner to the target combiner impls.
- TargetLowering::DAGCombinerInfo
- DagCombineInfo(DAG, Level == Unrestricted, false, this);
+ TargetLowering::DAGCombinerInfo
+ DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
RV = TLI.PerformDAGCombine(N, DagCombineInfo);
}
}
- // If N is a commutative binary node, try commuting it to enable more
+ // If N is a commutative binary node, try commuting it to enable more
// sdisel CSE.
- if (RV.getNode() == 0 &&
+ if (RV.getNode() == 0 &&
SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
N->getNumValues() == 1) {
SDValue N0 = N->getOperand(0);
}
return RV;
-}
+}
/// getInputChainForNode - Given a node, return its input chain if it has one,
/// otherwise return a null sd operand.
if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
return N->getOperand(1);
}
-
+
SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
- SmallPtrSet<SDNode*, 16> SeenOps;
+ SmallPtrSet<SDNode*, 16> SeenOps;
bool Changed = false; // If we should replace this token factor.
-
+
// Start out with this token factor.
TFs.push_back(N);
-
+
// Iterate through token factors. The TFs grows when new token factors are
// encountered.
for (unsigned i = 0; i < TFs.size(); ++i) {
SDNode *TF = TFs[i];
-
+
// Check each of the operands.
for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
SDValue Op = TF->getOperand(i);
-
+
switch (Op.getOpcode()) {
case ISD::EntryToken:
// Entry tokens don't need to be added to the list. They are
// rededundant.
Changed = true;
break;
-
+
case ISD::TokenFactor:
- if ((CombinerAA || Op.hasOneUse()) &&
+ if (Op.hasOneUse() &&
std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
// Queue up for processing.
TFs.push_back(Op.getNode());
break;
}
// Fall thru
-
+
default:
// Only add if it isn't already in the list.
if (SeenOps.insert(Op.getNode()))
}
}
}
-
+
SDValue Result;
// If we've change things around then replace token factor.
// Don't add users to work list.
return CombineTo(N, Result, false);
}
-
+
return Result;
}
/// MERGE_VALUES can always be eliminated.
SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
WorkListRemover DeadNodes(*this);
- for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
- DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i),
- &DeadNodes);
+ // Replacing results may cause a different MERGE_VALUES to suddenly
+ // be CSE'd with N, and carry its uses with it. Iterate until no
+ // uses remain, to ensure that the node can be safely deleted.
+ do {
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
+ DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i),
+ &DeadNodes);
+ } while (!N->use_empty());
removeFromWorkList(N);
DAG.DeleteNode(N);
return SDValue(N, 0); // Return N so it doesn't get rechecked!
static
SDValue combineShlAddConstant(DebugLoc DL, SDValue N0, SDValue N1,
SelectionDAG &DAG) {
- MVT VT = N0.getValueType();
+ EVT VT = N0.getValueType();
SDValue N00 = N0.getOperand(0);
SDValue N01 = N0.getOperand(1);
ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
return SDValue();
}
-static
-SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
- SelectionDAG &DAG, const TargetLowering &TLI,
- bool LegalOperations) {
- MVT VT = N->getValueType(0);
- unsigned Opc = N->getOpcode();
- bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
- SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
- SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
- ISD::CondCode CC = ISD::SETCC_INVALID;
-
- if (isSlctCC) {
- CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
- } else {
- SDValue CCOp = Slct.getOperand(0);
- if (CCOp.getOpcode() == ISD::SETCC)
- CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
- }
-
- bool DoXform = false;
- bool InvCC = false;
- assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
- "Bad input!");
-
- if (LHS.getOpcode() == ISD::Constant &&
- cast<ConstantSDNode>(LHS)->isNullValue()) {
- DoXform = true;
- } else if (CC != ISD::SETCC_INVALID &&
- RHS.getOpcode() == ISD::Constant &&
- cast<ConstantSDNode>(RHS)->isNullValue()) {
- std::swap(LHS, RHS);
- SDValue Op0 = Slct.getOperand(0);
- MVT OpVT = isSlctCC ? Op0.getValueType() :
- Op0.getOperand(0).getValueType();
- bool isInt = OpVT.isInteger();
- CC = ISD::getSetCCInverse(CC, isInt);
-
- if (LegalOperations && !TLI.isCondCodeLegal(CC, OpVT))
- return SDValue(); // Inverse operator isn't legal.
-
- DoXform = true;
- InvCC = true;
- }
-
- if (DoXform) {
- SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
- if (isSlctCC)
- return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
- Slct.getOperand(0), Slct.getOperand(1), CC);
- SDValue CCOp = Slct.getOperand(0);
- if (InvCC)
- CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
- CCOp.getOperand(0), CCOp.getOperand(1), CC);
- return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
- CCOp, OtherOp, Result);
- }
- return SDValue();
-}
-
SDValue DAGCombiner::visitADD(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N0.getValueType();
+ EVT VT = N0.getValueType();
// fold vector ops
if (VT.isVector()) {
if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
return SDValue(N, 0);
-
+
// fold (a+b) -> (a|b) iff a and b share no bits.
if (VT.isInteger() && !VT.isVector()) {
APInt LHSZero, LHSOne;
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.
// If all possibly-set bits on the RHS are clear on the LHS, return an OR.
if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
if (Result.getNode()) return Result;
}
- // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
- if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
- SDValue Result = combineSelectAndUse(N, N0, N1, DAG, TLI, LegalOperations);
- if (Result.getNode()) return Result;
- }
- if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
- SDValue Result = combineSelectAndUse(N, N1, N0, DAG, TLI, LegalOperations);
- if (Result.getNode()) return Result;
- }
-
return SDValue();
}
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N0.getValueType();
-
+ EVT VT = N0.getValueType();
+
// If the flag result is dead, turn this into an ADD.
if (N->hasNUsesOfValue(0, 1))
return CombineTo(N, DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0),
DAG.getNode(ISD::CARRY_FALSE,
N->getDebugLoc(), MVT::Flag));
-
+
// canonicalize constant to RHS.
if (N0C && !N1C)
return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
-
+
// fold (addc x, 0) -> x + no carry out
if (N1C && N1C->isNullValue())
return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
N->getDebugLoc(), MVT::Flag));
-
+
// fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
APInt LHSZero, LHSOne;
APInt RHSZero, RHSOne;
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.
// If all possibly-set bits on the RHS are clear on the LHS, return an OR.
if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
DAG.getNode(ISD::CARRY_FALSE,
N->getDebugLoc(), MVT::Flag));
}
-
+
return SDValue();
}
SDValue CarryIn = N->getOperand(2);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
-
+
// canonicalize constant to RHS
if (N0C && !N1C)
return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(),
N1, N0, CarryIn);
-
+
// fold (adde x, y, false) -> (addc x, y)
if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
-
+
return SDValue();
}
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
- MVT VT = N0.getValueType();
-
+ EVT VT = N0.getValueType();
+
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
return N0.getOperand(1);
// fold (A+B)-B -> A
if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
- return N0.getOperand(0);
+ return N0.getOperand(0);
// fold ((A+(B+or-C))-B) -> A+or-C
if (N0.getOpcode() == ISD::ADD &&
(N0.getOperand(1).getOpcode() == ISD::SUB ||
N0.getOperand(1).getOperand(1) == N1)
return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
N0.getOperand(0), N0.getOperand(1).getOperand(0));
- // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
- if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
- SDValue Result = combineSelectAndUse(N, N1, N0, DAG, TLI, LegalOperations);
- if (Result.getNode()) return Result;
- }
// If either operand of a sub is undef, the result is undef
if (N0.getOpcode() == ISD::UNDEF)
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N0.getValueType();
-
+ EVT VT = N0.getValueType();
+
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
}
-
+
// fold (mul x, undef) -> 0
if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
return DAG.getConstant(0, VT);
DAG.getConstant(N1C->getAPIntValue().logBase2(),
getShiftAmountTy()));
// fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
- if (N1C && isPowerOf2_64(-N1C->getSExtValue()))
- // FIXME: If the input is something that is easily negated (e.g. a
+ if (N1C && (-N1C->getAPIntValue()).isPowerOf2()) {
+ unsigned Log2Val = (-N1C->getAPIntValue()).logBase2();
+ // FIXME: If the input is something that is easily negated (e.g. a
// single-use add), we should put the negate there.
return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
DAG.getConstant(0, VT),
DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
- DAG.getConstant(Log2_64(-N1C->getSExtValue()),
- getShiftAmountTy())));
+ DAG.getConstant(Log2Val, getShiftAmountTy())));
+ }
// (mul (shl X, c1), c2) -> (mul X, c2 << c1)
if (N1C && N0.getOpcode() == ISD::SHL &&
isa<ConstantSDNode>(N0.getOperand(1))) {
return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
N0.getOperand(0), C3);
}
-
+
// Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
// use.
{
if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
N0.getNode()->hasOneUse()) {
Sh = N0; Y = N1;
- } else if (N1.getOpcode() == ISD::SHL &&
+ } else if (N1.getOpcode() == ISD::SHL &&
isa<ConstantSDNode>(N1.getOperand(1)) &&
N1.getNode()->hasOneUse()) {
Sh = N1; Y = N0;
}
// fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
- if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
+ if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
isa<ConstantSDNode>(N0.getOperand(1)))
return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
DAG.getNode(ISD::MUL, N0.getDebugLoc(), VT,
N0.getOperand(0), N1),
DAG.getNode(ISD::MUL, N1.getDebugLoc(), VT,
N0.getOperand(1), N1));
-
+
// reassociate mul
SDValue RMUL = ReassociateOps(ISD::MUL, N->getDebugLoc(), N0, N1);
if (RMUL.getNode() != 0)
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
}
-
+
// fold (sdiv c1, c2) -> c1/c2
if (N0C && N1C && !N1C->isNullValue())
return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
}
// fold (sdiv X, pow2) -> simple ops after legalize
if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap() &&
- (isPowerOf2_64(N1C->getSExtValue()) ||
+ (isPowerOf2_64(N1C->getSExtValue()) ||
isPowerOf2_64(-N1C->getSExtValue()))) {
// If dividing by powers of two is cheap, then don't perform the following
// fold.
// if integer divide is expensive and we satisfy the requirements, emit an
// alternate sequence.
- if (N1C && (N1C->getSExtValue() < -1 || N1C->getSExtValue() > 1) &&
+ if (N1C && (N1C->getSExtValue() < -1 || N1C->getSExtValue() > 1) &&
!TLI.isIntDivCheap()) {
SDValue Op = BuildSDIV(N);
if (Op.getNode()) return Op;
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
}
-
+
// fold (udiv c1, c2) -> c1/c2
if (N0C && N1C && !N1C->isNullValue())
return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
// fold (udiv x, (1 << c)) -> x >>u c
if (N1C && N1C->getAPIntValue().isPowerOf2())
- return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
+ return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
DAG.getConstant(N1C->getAPIntValue().logBase2(),
getShiftAmountTy()));
// fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
if (N1.getOpcode() == ISD::SHL) {
if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
if (SHC->getAPIntValue().isPowerOf2()) {
- MVT ADDVT = N1.getOperand(1).getValueType();
+ EVT ADDVT = N1.getOperand(1).getValueType();
SDValue Add = DAG.getNode(ISD::ADD, N->getDebugLoc(), ADDVT,
N1.getOperand(1),
DAG.getConstant(SHC->getAPIntValue()
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (srem c1, c2) -> c1%c2
if (N0C && N1C && !N1C->isNullValue())
return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
return DAG.getNode(ISD::UREM, N->getDebugLoc(), VT, N0, N1);
}
-
+
// 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()) {
return Sub;
}
}
-
+
// undef % X -> 0
if (N0.getOpcode() == ISD::UNDEF)
return DAG.getConstant(0, VT);
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (urem c1, c2) -> c1%c2
if (N0C && N1C && !N1C->isNullValue())
return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
}
}
}
-
+
// 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()) {
return Sub;
}
}
-
+
// undef % X -> 0
if (N0.getOpcode() == ISD::UNDEF)
return DAG.getConstant(0, VT);
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (mulhs x, 0) -> 0
if (N1C && N1C->isNullValue())
return N1;
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (mulhu x, 0) -> 0
if (N1C && N1C->isNullValue())
return N1;
/// compute two values. LoOp and HiOp give the opcodes for the two computations
/// that are being performed. Return true if a simplification was made.
///
-SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
+SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
unsigned HiOp) {
// If the high half is not needed, just compute the low half.
bool HiExists = N->hasAnyUseOfValue(1);
SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
if (Res.getNode()) return Res;
-
+
return SDValue();
}
SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
if (Res.getNode()) return Res;
-
+
return SDValue();
}
/// two operands of the same opcode, try to simplify it.
SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
- MVT VT = N0.getValueType();
+ EVT VT = N0.getValueType();
assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
-
+
// For each of OP in AND/OR/XOR:
// fold (OP (zext x), (zext y)) -> (zext (OP x, y))
// fold (OP (sext x), (sext y)) -> (sext (OP x, y))
// fold (OP (aext x), (aext y)) -> (aext (OP x, y))
- // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y))
+ // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
if ((N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND||
- N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::TRUNCATE) &&
- N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
+ N0.getOpcode() == ISD::SIGN_EXTEND ||
+ (N0.getOpcode() == ISD::TRUNCATE &&
+ !TLI.isTruncateFree(N0.getOperand(0).getValueType(), VT))) &&
+ N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
+ (!LegalOperations ||
+ TLI.isOperationLegal(N->getOpcode(), N0.getOperand(0).getValueType()))) {
SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
N0.getOperand(0).getValueType(),
N0.getOperand(0), N1.getOperand(0));
AddToWorkList(ORNode.getNode());
return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, ORNode);
}
-
+
// For each of OP in SHL/SRL/SRA/AND...
// fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
// fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
ORNode, N0.getOperand(1));
}
-
+
return SDValue();
}
SDValue LL, LR, RL, RR, CC0, CC1;
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N1.getValueType();
+ EVT VT = N1.getValueType();
unsigned BitWidth = VT.getSizeInBits();
-
+
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
}
-
+
// fold (and x, undef) -> 0
if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
return DAG.getConstant(0, VT);
if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(),
N0.getValueType(), N0Op0);
-
+
// Replace uses of the AND with uses of the Zero extend node.
CombineTo(N, Zext);
-
+
// We actually want to replace all uses of the any_extend with the
// zero_extend, to avoid duplicating things. This will later cause this
// AND to be folded.
if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
-
+
if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
LL.getValueType().isInteger()) {
// fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
if (Tmp.getNode()) return Tmp;
}
-
+
// fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
// fold (and (sra)) -> (and (srl)) when possible.
if (!VT.isVector() &&
// fold (zext_inreg (extload x)) -> (zextload x)
if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT EVT = LN0->getMemoryVT();
+ EVT MemVT = 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.
unsigned BitWidth = N1.getValueSizeInBits();
if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
- BitWidth - EVT.getSizeInBits())) &&
+ BitWidth - MemVT.getSizeInBits())) &&
((!LegalOperations && !LN0->isVolatile()) ||
- TLI.isLoadExtLegal(ISD::ZEXTLOAD, EVT))) {
+ TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
LN0->getChain(), LN0->getBasePtr(),
LN0->getSrcValue(),
- LN0->getSrcValueOffset(), EVT,
+ LN0->getSrcValueOffset(), MemVT,
LN0->isVolatile(), LN0->getAlignment());
AddToWorkList(N);
CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT EVT = LN0->getMemoryVT();
+ EVT MemVT = 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.
unsigned BitWidth = N1.getValueSizeInBits();
if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
- BitWidth - EVT.getSizeInBits())) &&
+ BitWidth - MemVT.getSizeInBits())) &&
((!LegalOperations && !LN0->isVolatile()) ||
- TLI.isLoadExtLegal(ISD::ZEXTLOAD, EVT))) {
+ TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(), EVT,
+ LN0->getSrcValueOffset(), MemVT,
LN0->isVolatile(), LN0->getAlignment());
AddToWorkList(N);
CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
return SDValue(N, 0); // Return N so it doesn't get rechecked!
}
}
-
+
// fold (and (load x), 255) -> (zextload x, i8)
// fold (and (extload x, i16), 255) -> (zextload x, i8)
if (N1C && N0.getOpcode() == ISD::LOAD) {
LN0->isUnindexed() && N0.hasOneUse() &&
// Do not change the width of a volatile load.
!LN0->isVolatile()) {
- MVT EVT = MVT::Other;
+ EVT ExtVT = MVT::Other;
uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue()))
- EVT = MVT::getIntegerVT(ActiveBits);
+ ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
- MVT LoadedVT = LN0->getMemoryVT();
+ EVT LoadedVT = LN0->getMemoryVT();
// Do not generate loads of non-round integer types since these can
// be expensive (and would be wrong if the type is not byte sized).
- if (EVT != MVT::Other && LoadedVT.bitsGT(EVT) && EVT.isRound() &&
- (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, EVT))) {
- MVT PtrType = N0.getOperand(1).getValueType();
+ if (ExtVT != MVT::Other && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
+ (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
+ EVT 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 LVTStoreBytes = LoadedVT.getStoreSizeInBits()/8;
- unsigned EVTStoreBytes = EVT.getStoreSizeInBits()/8;
+ unsigned LVTStoreBytes = LoadedVT.getStoreSize();
+ unsigned EVTStoreBytes = ExtVT.getStoreSize();
unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
unsigned Alignment = LN0->getAlignment();
SDValue NewPtr = LN0->getBasePtr();
SDValue Load =
DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), VT, LN0->getChain(),
NewPtr, LN0->getSrcValue(), LN0->getSrcValueOffset(),
- EVT, LN0->isVolatile(), Alignment);
+ ExtVT, LN0->isVolatile(), Alignment);
AddToWorkList(N);
CombineTo(N0.getNode(), Load, Load.getValue(1));
return SDValue(N, 0); // Return N so it doesn't get rechecked!
}
}
}
-
+
return SDValue();
}
SDValue LL, LR, RL, RR, CC0, CC1;
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N1.getValueType();
-
+ EVT VT = N1.getValueType();
+
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
}
-
+
// fold (or x, undef) -> -1
if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
- return DAG.getConstant(~0ULL, VT);
+ return DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
// fold (or c1, c2) -> c1|c2
if (N0C && N1C)
return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
-
+
if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
LL.getValueType().isInteger()) {
// fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
// fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
- if (cast<ConstantSDNode>(LR)->isNullValue() &&
+ if (cast<ConstantSDNode>(LR)->isNullValue() &&
(Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
LR.getValueType(), LL, RL);
}
// fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
// fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
- if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
+ if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
(Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
LR.getValueType(), LL, RL);
LL, LR, Result);
}
}
-
+
// Simplify: (or (op x...), (op y...)) -> (op (or x, y))
if (N0.getOpcode() == N1.getOpcode()) {
SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
if (Tmp.getNode()) return Tmp;
}
-
+
// (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
if (N0.getOpcode() == ISD::AND &&
N1.getOpcode() == ISD::AND &&
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)) {
SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
DAG.getConstant(LHSMask | RHSMask, VT));
}
}
-
+
// See if this is some rotate idiom.
if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
return SDValue(Rot, 0);
return false;
}
}
-
+
if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
Shift = Op;
return true;
}
- return false;
+ return false;
}
// MatchRotate - Handle an 'or' of two operands. If this is one of the many
// a rot[lr].
SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
// Must be a legal type. Expanded 'n promoted things won't work with rotates.
- MVT VT = LHS.getValueType();
+ EVT VT = LHS.getValueType();
if (!TLI.isTypeLegal(VT)) return 0;
// The target must have at least one rotate flavor.
SDValue RHSMask; // AND value if any.
if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
return 0; // Not part of a rotate.
-
+
if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
return 0; // Not shifting the same value.
if (LHSShift.getOpcode() == RHSShift.getOpcode())
return 0; // Shifts must disagree.
-
+
// Canonicalize shl to left side in a shl/srl pair.
if (RHSShift.getOpcode() == ISD::SHL) {
std::swap(LHS, RHS);
Rot = DAG.getNode(ISD::ROTL, DL, VT, LHSShiftArg, LHSShiftAmt);
else
Rot = DAG.getNode(ISD::ROTR, DL, VT, LHSShiftArg, RHSShiftAmt);
-
+
// If there is an AND of either shifted operand, apply it to the result.
if (LHSMask.getNode() || RHSMask.getNode()) {
APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
-
+
if (LHSMask.getNode()) {
APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
}
-
+
Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
}
-
+
return Rot.getNode();
}
-
+
// If there is a mask here, and we have a variable shift, we can't be sure
// that we're masking out the right stuff.
if (LHSMask.getNode() || RHSMask.getNode())
return 0;
-
+
// fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
// fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
if (RHSShiftAmt.getOpcode() == ISD::SUB &&
LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
- if (ConstantSDNode *SUBC =
+ if (ConstantSDNode *SUBC =
dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
if (SUBC->getAPIntValue() == OpSizeInBits) {
if (HasROTL)
}
}
}
-
+
// fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
// fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
if (LHSShiftAmt.getOpcode() == ISD::SUB &&
RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
- if (ConstantSDNode *SUBC =
+ if (ConstantSDNode *SUBC =
dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
if (SUBC->getAPIntValue() == OpSizeInBits) {
if (HasROTR)
}
} else if (LExtOp0.getOpcode() == ISD::SUB &&
RExtOp0 == LExtOp0.getOperand(1)) {
- // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
+ // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
// (rotr x, y)
// fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
// (rotl x, (sub 32, y))
}
}
}
-
+
return 0;
}
SDValue LHS, RHS, CC;
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N0.getValueType();
-
+ EVT VT = N0.getValueType();
+
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
}
-
+
// fold (xor undef, undef) -> 0. This is a common idiom (misuse).
if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
return DAG.getConstant(0, VT);
if (!LegalOperations || TLI.isCondCodeLegal(NotCC, LHS.getValueType())) {
switch (N0.getOpcode()) {
default:
- assert(0 && "Unhandled SetCC Equivalent!");
- abort();
+ llvm_unreachable("Unhandled SetCC Equivalent!");
case ISD::SETCC:
return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
case ISD::SELECT_CC:
N0.getNode()->hasOneUse() &&
isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
SDValue V = N0.getOperand(0);
- V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
+ V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
DAG.getConstant(1, V.getValueType()));
AddToWorkList(V.getNode());
return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
}
-
+
// fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
(N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
}
}
// fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
- if (N1C && N1C->isAllOnesValue() &&
+ if (N1C && N1C->isAllOnesValue() &&
(N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
&Ops[0], Ops.size());
}
}
-
+
// Simplify: xor (op x...), (op y...) -> (op (xor x, y))
if (N0.getOpcode() == N1.getOpcode()) {
SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
if (Tmp.getNode()) return Tmp;
}
-
+
// Simplify the expression using non-local knowledge.
if (!VT.isVector() &&
SimplifyDemandedBits(SDValue(N, 0)))
return SDValue(N, 0);
-
+
return SDValue();
}
SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
SDNode *LHS = N->getOperand(0).getNode();
if (!LHS->hasOneUse()) return SDValue();
-
+
// 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 SDValue();
case ISD::OR:
HighBitSet = true; // We can only transform sra if the high bit is set.
break;
case ISD::ADD:
- if (N->getOpcode() != ISD::SHL)
+ if (N->getOpcode() != ISD::SHL)
return SDValue(); // 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 SDValue();
// 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).getNode();
- if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
+ if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
BinOpLHSVal->getOpcode() != ISD::SRA &&
BinOpLHSVal->getOpcode() != ISD::SRL) ||
!isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
return SDValue();
-
- MVT VT = N->getValueType(0);
-
+
+ EVT 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
if (BinOpRHSSignSet != HighBitSet)
return SDValue();
}
-
+
// Fold the constants, shifting the binop RHS by the shift amount.
SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
N->getValueType(0),
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N0.getValueType();
+ EVT VT = N0.getValueType();
unsigned OpSizeInBits = VT.getSizeInBits();
-
+
// fold (shl c1, c2) -> c1<<c2
if (N0C && N1C)
return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
return N0;
// fold (shl x, c >= size(x)) -> undef
if (N1C && N1C->getZExtValue() >= OpSizeInBits)
- return DAG.getNode(ISD::UNDEF, N->getDebugLoc(), VT);
+ return DAG.getUNDEF(VT);
// fold (shl x, 0) -> x
if (N1C && N1C->isNullValue())
return N0;
N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
SDValue N101 = N1.getOperand(0).getOperand(1);
if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
- MVT TruncVT = N1.getValueType();
+ EVT TruncVT = N1.getValueType();
SDValue N100 = N1.getOperand(0).getOperand(0);
APInt TruncC = N101C->getAPIntValue();
TruncC.trunc(TruncVT.getSizeInBits());
return SDValue(N, 0);
// fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
- if (N1C && N0.getOpcode() == ISD::SHL &&
+ if (N1C && N0.getOpcode() == ISD::SHL &&
N0.getOperand(1).getOpcode() == ISD::Constant) {
uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
uint64_t c2 = N1C->getZExtValue();
}
// fold (shl (srl x, c1), c2) -> (shl (and x, (shl -1, c1)), (sub c2, c1)) or
// (srl (and x, (shl -1, c1)), (sub c1, c2))
- if (N1C && N0.getOpcode() == ISD::SRL &&
+ if (N1C && N0.getOpcode() == ISD::SRL &&
N0.getOperand(1).getOpcode() == ISD::Constant) {
uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
- uint64_t c2 = N1C->getZExtValue();
- SDValue Mask = DAG.getNode(ISD::AND, N0.getDebugLoc(), VT, N0.getOperand(0),
- DAG.getConstant(~0ULL << c1, VT));
- if (c2 > c1)
- return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, Mask,
- DAG.getConstant(c2-c1, N1.getValueType()));
- else
- return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Mask,
- DAG.getConstant(c1-c2, N1.getValueType()));
+ if (c1 < VT.getSizeInBits()) {
+ uint64_t c2 = N1C->getZExtValue();
+ SDValue HiBitsMask =
+ DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
+ VT.getSizeInBits() - c1),
+ VT);
+ SDValue Mask = DAG.getNode(ISD::AND, N0.getDebugLoc(), VT,
+ N0.getOperand(0),
+ HiBitsMask);
+ if (c2 > c1)
+ return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, Mask,
+ DAG.getConstant(c2-c1, N1.getValueType()));
+ else
+ return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Mask,
+ DAG.getConstant(c1-c2, N1.getValueType()));
+ }
}
// fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
- if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1))
+ if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
+ SDValue HiBitsMask =
+ DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
+ VT.getSizeInBits() -
+ N1C->getZExtValue()),
+ VT);
return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
- DAG.getConstant(~0ULL << N1C->getZExtValue(), VT));
-
+ HiBitsMask);
+ }
+
return N1C ? visitShiftByConstant(N, N1C->getZExtValue()) : SDValue();
}
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N0.getValueType();
-
+ EVT VT = N0.getValueType();
+
// fold (sra c1, c2) -> (sra c1, c2)
if (N0C && N1C)
return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
return N0;
// fold (sra x, (setge c, size(x))) -> undef
if (N1C && N1C->getZExtValue() >= VT.getSizeInBits())
- return DAG.getNode(ISD::UNDEF, N->getDebugLoc(), VT);
+ return DAG.getUNDEF(VT);
// fold (sra x, 0) -> x
if (N1C && N1C->isNullValue())
return N0;
// sext_inreg.
if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
unsigned LowBits = VT.getSizeInBits() - (unsigned)N1C->getZExtValue();
- MVT EVT = MVT::getIntegerVT(LowBits);
+ EVT EVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
if ((!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT)))
return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
N0.getOperand(0), DAG.getValueType(EVT));
// fold (sra (shl X, m), (sub result_size, n))
// -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
- // result_size - n != m.
- // If truncate is free for the target sext(shl) is likely to result in better
+ // result_size - n != m.
+ // If truncate is free for the target sext(shl) is likely to result in better
// code.
if (N0.getOpcode() == ISD::SHL) {
// Get the two constanst of the shifts, CN0 = m, CN = n.
if (N01C && N1C) {
// Determine what the truncate's result bitsize and type would be.
unsigned VTValSize = VT.getSizeInBits();
- MVT TruncVT =
- MVT::getIntegerVT(VTValSize - N1C->getZExtValue());
+ EVT TruncVT =
+ EVT::getIntegerVT(*DAG.getContext(), VTValSize - N1C->getZExtValue());
// Determine the residual right-shift amount.
- unsigned ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
+ signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
- // If the shift is not a no-op (in which case this should be just a sign
- // extend already), the truncated to type is legal, sign_extend is legal
+ // If the shift is not a no-op (in which case this should be just a sign
+ // extend already), the truncated to type is legal, sign_extend is legal
// on that type, and the the truncate to that type is both legal and free,
// perform the transform.
- if (ShiftAmt &&
+ if ((ShiftAmt > 0) &&
TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
TLI.isTruncateFree(VT, TruncVT)) {
}
}
}
-
+
// fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
if (N1.getOpcode() == ISD::TRUNCATE &&
N1.getOperand(0).getOpcode() == ISD::AND &&
N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
SDValue N101 = N1.getOperand(0).getOperand(1);
if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
- MVT TruncVT = N1.getValueType();
+ EVT TruncVT = N1.getValueType();
SDValue N100 = N1.getOperand(0).getOperand(0);
APInt TruncC = N101C->getAPIntValue();
TruncC.trunc(TruncVT.getSizeInBits());
}
}
- // Simplify, based on bits shifted out of the LHS.
+ // Simplify, based on bits shifted out of the LHS.
if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
return SDValue(N, 0);
-
-
+
+
// If the sign bit is known to be zero, switch this to a SRL.
if (DAG.SignBitIsZero(N0))
return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- MVT VT = N0.getValueType();
+ EVT VT = N0.getValueType();
unsigned OpSizeInBits = VT.getSizeInBits();
-
+
// fold (srl c1, c2) -> c1 >>u c2
if (N0C && N1C)
return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
return N0;
// fold (srl x, c >= size(x)) -> undef
if (N1C && N1C->getZExtValue() >= OpSizeInBits)
- return DAG.getNode(ISD::UNDEF, N->getDebugLoc(), VT);
+ return DAG.getUNDEF(VT);
// fold (srl x, 0) -> x
if (N1C && N1C->isNullValue())
return N0;
if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
APInt::getAllOnesValue(OpSizeInBits)))
return DAG.getConstant(0, VT);
-
+
// fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
- if (N1C && N0.getOpcode() == ISD::SRL &&
+ if (N1C && N0.getOpcode() == ISD::SRL &&
N0.getOperand(1).getOpcode() == ISD::Constant) {
uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
uint64_t c2 = N1C->getZExtValue();
return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
DAG.getConstant(c1 + c2, N1.getValueType()));
}
-
+
// fold (srl (anyextend x), c) -> (anyextend (srl x, c))
if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
// Shifting in all undef bits?
- MVT SmallVT = N0.getOperand(0).getValueType();
+ EVT SmallVT = N0.getOperand(0).getValueType();
if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
- return DAG.getNode(ISD::UNDEF, N->getDebugLoc(), VT);
+ return DAG.getUNDEF(VT);
SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
N0.getOperand(0), N1);
AddToWorkList(SmallShift.getNode());
return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
}
-
+
// fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
// bit, which is unmodified by sra.
if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
if (N0.getOpcode() == ISD::SRA)
return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
}
-
+
// fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
- if (N1C && N0.getOpcode() == ISD::CTLZ &&
+ if (N1C && N0.getOpcode() == ISD::CTLZ &&
N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
APInt KnownZero, KnownOne;
APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits());
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.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.
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.
if ((UnknownBits & (UnknownBits - 1)) == 0) {
// Okay, we know that only that the single bit specified by UnknownBits
N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
SDValue N101 = N1.getOperand(0).getOperand(1);
if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
- MVT TruncVT = N1.getValueType();
+ EVT TruncVT = N1.getValueType();
SDValue N100 = N1.getOperand(0).getOperand(0);
APInt TruncC = N101C->getAPIntValue();
TruncC.trunc(TruncVT.getSizeInBits());
DAG.getConstant(TruncC, TruncVT)));
}
}
-
+
// fold operands of srl based on knowledge that the low bits are not
// demanded.
if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
return SDValue(N, 0);
-
+
return N1C ? visitShiftByConstant(N, N1C->getZExtValue()) : SDValue();
}
SDValue DAGCombiner::visitCTLZ(SDNode *N) {
SDValue N0 = N->getOperand(0);
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
// fold (ctlz c1) -> c2
if (isa<ConstantSDNode>(N0))
SDValue DAGCombiner::visitCTTZ(SDNode *N) {
SDValue N0 = N->getOperand(0);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (cttz c1) -> c2
if (isa<ConstantSDNode>(N0))
return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
SDValue DAGCombiner::visitCTPOP(SDNode *N) {
SDValue N0 = N->getOperand(0);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (ctpop c1) -> c2
if (isa<ConstantSDNode>(N0))
return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
- MVT VT = N->getValueType(0);
- MVT VT0 = N0.getValueType();
+ EVT VT = N->getValueType(0);
+ EVT VT0 = N0.getValueType();
// fold (select C, X, X) -> X
if (N1 == N2)
// fold (select X, Y, 0) -> (and X, Y)
if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
-
+
// If we can fold this based on the true/false value, do so.
if (SimplifySelectOps(N, N1, N2))
return SDValue(N, 0); // Don't revisit N.
// 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
// about, since there is no way to mark an opcode illegal at all value types
- if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other))
+ if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
+ TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
N0.getOperand(0), N0.getOperand(1),
N1, N2, N0.getOperand(2));
- else
- return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
+ return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
}
return SDValue();
SDValue N3 = N->getOperand(3);
SDValue N4 = N->getOperand(4);
ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
-
+
// fold select_cc lhs, rhs, x, x, cc -> x
if (N2 == N3)
return N2;
-
+
// Determine if the condition we're dealing with is constant
SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
N0, N1, CC, N->getDebugLoc(), false);
else
return N3; // cond always false -> false val
}
-
+
// Fold to a simpler select_cc
if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
- return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
- SCC.getOperand(0), SCC.getOperand(1), N2, N3,
+ return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
+ SCC.getOperand(0), SCC.getOperand(1), N2, N3,
SCC.getOperand(2));
-
+
// If we can fold this based on the true/false value, do so.
if (SimplifySelectOps(N, N2, N3))
return SDValue(N, 0); // Don't revisit N.
-
+
// fold select_cc into other things, such as min/max/abs
return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
}
}
// 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)))"
+// "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
// transformation. Returns true if extension are possible and the above
-// mentioned transformation is profitable.
+// mentioned transformation is profitable.
static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
unsigned ExtOpc,
SmallVector<SDNode*, 4> &ExtendNodes,
SDNode *User = *UI;
if (User == N)
continue;
+ if (UI.getUse().getResNo() != N0.getResNo())
+ continue;
// FIXME: Only extend SETCC N, N and SETCC N, c for now.
- if (User->getOpcode() == ISD::SETCC) {
+ if (ExtOpc != ISD::ANY_EXTEND && 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.
}
if (Add)
ExtendNodes.push_back(User);
- } else {
- for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
- SDValue 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;
- }
- }
+ continue;
}
+ // If truncates aren't free and there are users we can't
+ // extend, it isn't worthwhile.
+ if (!isTruncFree)
+ return false;
+ // Remember if this value is live-out.
+ if (User->getOpcode() == ISD::CopyToReg)
+ HasCopyToRegUses = true;
}
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) {
- SDValue UseOp = User->getOperand(i);
- if (UseOp.getNode() == N && UseOp.getResNo() == 0) {
- BothLiveOut = true;
- break;
- }
+ SDUse &Use = UI.getUse();
+ if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
+ BothLiveOut = true;
+ break;
}
}
if (BothLiveOut)
SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
SDValue N0 = N->getOperand(0);
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
// fold (sext c1) -> c1
if (isa<ConstantSDNode>(N0))
return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
-
+
// fold (sext (sext x)) -> (sext x)
// fold (sext (aext x)) -> (sext x)
if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
N0.getOperand(0));
-
+
if (N0.getOpcode() == ISD::TRUNCATE) {
// fold (sext (truncate (load x))) -> (sext (smaller load x))
// fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
if (NarrowLoad.getNode()) {
if (NarrowLoad.getNode() != N0.getNode())
CombineTo(N0.getNode(), NarrowLoad);
- return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, NarrowLoad);
+ return SDValue(N, 0); // Return N so it doesn't get rechecked!
}
// See if the value being truncated is already sign extended. If so, just
unsigned MidBits = N0.getValueType().getSizeInBits();
unsigned DestBits = VT.getSizeInBits();
unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
-
+
if (OpBits == DestBits) {
// Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
// bits, it is already ready.
if (NumSignBits > OpBits-MidBits)
return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
}
-
+
// fold (sext (truncate x)) -> (sextinreg x).
if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
N0.getValueType())) {
DAG.getValueType(N0.getValueType()));
}
}
-
+
// fold (sext (load x)) -> (sext (truncate (sextload x)))
if (ISD::isNON_EXTLoad(N0.getNode()) &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
if (DoXform) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(),
- VT, LN0->getChain(),
+ SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
+ LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(),
N0.getValueType(),
if (SOp == Trunc)
Ops.push_back(ExtLoad);
else
- Ops.push_back(DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
- VT, SOp));
+ Ops.push_back(DAG.getNode(ISD::SIGN_EXTEND,
+ N->getDebugLoc(), VT, SOp));
}
Ops.push_back(SetCC->getOperand(2));
if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT EVT = LN0->getMemoryVT();
+ EVT MemVT = LN0->getMemoryVT();
if ((!LegalOperations && !LN0->isVolatile()) ||
- TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT)) {
+ TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(), EVT,
+ LN0->getSrcValueOffset(), MemVT,
LN0->isVolatile(), LN0->getAlignment());
CombineTo(N, ExtLoad);
CombineTo(N0.getNode(),
return SDValue(N, 0); // Return N so it doesn't get rechecked!
}
}
-
- // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
+
if (N0.getOpcode() == ISD::SETCC) {
- SDValue SCC =
+ // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
+ if (VT.isVector() &&
+ // We know that the # elements of the results is the same as the
+ // # elements of the compare (and the # elements of the compare result
+ // for that matter). Check to see that they are the same size. If so,
+ // we know that the element size of the sext'd result matches the
+ // element size of the compare operands.
+ VT.getSizeInBits() == N0.getOperand(0).getValueType().getSizeInBits() &&
+
+ // Only do this before legalize for now.
+ !LegalOperations) {
+ return DAG.getVSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
+ N0.getOperand(1),
+ cast<CondCodeSDNode>(N0.getOperand(2))->get());
+ }
+
+ // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
+ SDValue NegOne =
+ DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
+ SDValue SCC =
SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
- DAG.getConstant(~0ULL, VT), DAG.getConstant(0, VT),
+ NegOne, DAG.getConstant(0, VT),
cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
if (SCC.getNode()) return SCC;
}
+
+
// fold (sext x) -> (zext x) if the sign bit is known zero.
if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
DAG.SignBitIsZero(N0))
return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
-
+
return SDValue();
}
SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
SDValue N0 = N->getOperand(0);
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
// fold (zext c1) -> c1
if (isa<ConstantSDNode>(N0))
}
return DAG.getZeroExtendInReg(Op, N->getDebugLoc(), N0.getValueType());
}
-
- // fold (zext (and (trunc x), cst)) -> (and x, cst).
+
+ // Fold (zext (and (trunc x), cst)) -> (and x, cst),
+ // if either of the casts is not free.
if (N0.getOpcode() == ISD::AND &&
N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
- N0.getOperand(1).getOpcode() == ISD::Constant) {
+ N0.getOperand(1).getOpcode() == ISD::Constant &&
+ (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
+ N0.getValueType()) ||
+ !TLI.isZExtFree(N0.getValueType(), VT))) {
SDValue X = N0.getOperand(0).getOperand(0);
if (X.getValueType().bitsLT(VT)) {
X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
X, DAG.getConstant(Mask, VT));
}
-
+
// fold (zext (load x)) -> (zext (truncate (zextload x)))
if (ISD::isNON_EXTLoad(N0.getNode()) &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT EVT = LN0->getMemoryVT();
+ EVT MemVT = LN0->getMemoryVT();
if ((!LegalOperations && !LN0->isVolatile()) ||
- TLI.isLoadExtLegal(ISD::ZEXTLOAD, EVT)) {
+ TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(), EVT,
+ LN0->getSrcValueOffset(), MemVT,
LN0->isVolatile(), LN0->getAlignment());
CombineTo(N, ExtLoad);
CombineTo(N0.getNode(),
return SDValue(N, 0); // Return N so it doesn't get rechecked!
}
}
-
+
// zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
if (N0.getOpcode() == ISD::SETCC) {
- SDValue SCC =
+ SDValue SCC =
SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
DAG.getConstant(1, VT), DAG.getConstant(0, VT),
cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
if (SCC.getNode()) return SCC;
}
-
+
return SDValue();
}
SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
SDValue N0 = N->getOperand(0);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (aext c1) -> c1
if (isa<ConstantSDNode>(N0))
return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
N0.getOpcode() == ISD::ZERO_EXTEND ||
N0.getOpcode() == ISD::SIGN_EXTEND)
return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
-
+
// fold (aext (truncate (load x))) -> (aext (smaller load x))
// fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
if (N0.getOpcode() == ISD::TRUNCATE) {
return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
}
-
- // fold (aext (and (trunc x), cst)) -> (and x, cst).
+
+ // Fold (aext (and (trunc x), cst)) -> (and x, cst)
+ // if the trunc is not free.
if (N0.getOpcode() == ISD::AND &&
N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
- N0.getOperand(1).getOpcode() == ISD::Constant) {
+ N0.getOperand(1).getOpcode() == ISD::Constant &&
+ !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
+ N0.getValueType())) {
SDValue X = N0.getOperand(0).getOperand(0);
if (X.getValueType().bitsLT(VT)) {
X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
X, DAG.getConstant(Mask, VT));
}
-
+
// fold (aext (load x)) -> (aext (truncate (extload x)))
- if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
+ if (ISD::isNON_EXTLoad(N0.getNode()) &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
- LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
- LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(),
- N0.getValueType(),
- LN0->isVolatile(), LN0->getAlignment());
- CombineTo(N, ExtLoad);
- // Redirect any chain users to the new load.
- DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1),
- SDValue(ExtLoad.getNode(), 1));
- // If any node needs the original loaded value, recompute it.
- if (!LN0->use_empty())
- CombineTo(LN0, DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
- N0.getValueType(), ExtLoad),
- ExtLoad.getValue(1));
- return SDValue(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::ANY_EXTEND, SetCCs, TLI);
+ if (DoXform) {
+ LoadSDNode *LN0 = cast<LoadSDNode>(N0);
+ SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
+ LN0->getChain(),
+ LN0->getBasePtr(), LN0->getSrcValue(),
+ LN0->getSrcValueOffset(),
+ N0.getValueType(),
+ LN0->isVolatile(), LN0->getAlignment());
+ CombineTo(N, ExtLoad);
+ SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
+ N0.getValueType(), ExtLoad);
+ CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
+
+ // Extend SetCC uses if necessary.
+ for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
+ SDNode *SetCC = SetCCs[i];
+ SmallVector<SDValue, 4> Ops;
+
+ for (unsigned j = 0; j != 2; ++j) {
+ SDValue SOp = SetCC->getOperand(j);
+ if (SOp == Trunc)
+ Ops.push_back(ExtLoad);
+ else
+ Ops.push_back(DAG.getNode(ISD::ANY_EXTEND,
+ N->getDebugLoc(), VT, SOp));
+ }
+
+ Ops.push_back(SetCC->getOperand(2));
+ CombineTo(SetCC, DAG.getNode(ISD::SETCC, N->getDebugLoc(),
+ SetCC->getValueType(0),
+ &Ops[0], Ops.size()));
+ }
+
+ return SDValue(N, 0); // Return N so it doesn't get rechecked!
+ }
}
-
+
// fold (aext (zextload x)) -> (aext (truncate (zextload x)))
// fold (aext (sextload x)) -> (aext (truncate (sextload x)))
// fold (aext ( extload x)) -> (aext (truncate (extload x)))
!ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT EVT = LN0->getMemoryVT();
+ EVT MemVT = LN0->getMemoryVT();
SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
VT, LN0->getChain(), LN0->getBasePtr(),
LN0->getSrcValue(),
- LN0->getSrcValueOffset(), EVT,
+ LN0->getSrcValueOffset(), MemVT,
LN0->isVolatile(), LN0->getAlignment());
CombineTo(N, ExtLoad);
CombineTo(N0.getNode(),
ExtLoad.getValue(1));
return SDValue(N, 0); // Return N so it doesn't get rechecked!
}
-
+
// aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
if (N0.getOpcode() == ISD::SETCC) {
- SDValue SCC =
+ SDValue SCC =
SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
DAG.getConstant(1, VT), DAG.getConstant(0, VT),
cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
if (SCC.getNode())
return SCC;
}
-
+
return SDValue();
}
APInt NewMask = Mask << Amt;
SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
if (SimplifyLHS.getNode())
- return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
+ return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
SimplifyLHS, V.getOperand(1));
}
}
unsigned Opc = N->getOpcode();
ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
SDValue N0 = N->getOperand(0);
- MVT VT = N->getValueType(0);
- MVT EVT = VT;
+ EVT VT = N->getValueType(0);
+ EVT ExtVT = VT;
// This transformation isn't valid for vector loads.
if (VT.isVector())
// extended to VT.
if (Opc == ISD::SIGN_EXTEND_INREG) {
ExtType = ISD::SEXTLOAD;
- EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
- if (LegalOperations && !TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))
+ ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
+ if (LegalOperations && !TLI.isLoadExtLegal(ISD::SEXTLOAD, ExtVT))
return SDValue();
}
- unsigned EVTBits = EVT.getSizeInBits();
+ unsigned EVTBits = ExtVT.getSizeInBits();
unsigned ShAmt = 0;
- if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
+ if (N0.getOpcode() == ISD::SRL && N0.hasOneUse() && ExtVT.isRound()) {
if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
ShAmt = N01->getZExtValue();
// Is the shift amount a multiple of size of VT?
if ((ShAmt & (EVTBits-1)) == 0) {
N0 = N0.getOperand(0);
- if (N0.getValueType().getSizeInBits() <= EVTBits)
+ // Is the load width a multiple of size of VT?
+ if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
return SDValue();
}
}
// Do not generate loads of non-round integer types since these can
// be expensive (and would be wrong if the type is not byte sized).
- if (isa<LoadSDNode>(N0) && N0.hasOneUse() && EVT.isRound() &&
+ if (isa<LoadSDNode>(N0) && N0.hasOneUse() && ExtVT.isRound() &&
cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() > EVTBits &&
// Do not change the width of a volatile load.
!cast<LoadSDNode>(N0)->isVolatile()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
- MVT PtrType = N0.getOperand(1).getValueType();
+ EVT 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.isBigEndian()) {
unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
- unsigned EVTStoreBits = EVT.getStoreSizeInBits();
+ unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
}
LN0->isVolatile(), NewAlign)
: DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(), NewPtr,
LN0->getSrcValue(), LN0->getSrcValueOffset() + PtrOff,
- EVT, LN0->isVolatile(), NewAlign);
+ ExtVT, LN0->isVolatile(), NewAlign);
// Replace the old load's chain with the new load's chain.
WorkListRemover DeadNodes(*this);
SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
- MVT VT = N->getValueType(0);
- MVT EVT = cast<VTSDNode>(N1)->getVT();
+ EVT VT = N->getValueType(0);
+ EVT EVT = cast<VTSDNode>(N1)->getVT();
unsigned VTBits = VT.getSizeInBits();
unsigned EVTBits = EVT.getSizeInBits();
-
+
// fold (sext_in_reg c1) -> c1
if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
-
+
// If the input is already sign extended, just drop the extension.
if (DAG.ComputeNumSignBits(N0) >= VT.getSizeInBits()-EVTBits+1)
return N0;
-
+
// fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
// fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
-
+
// fold operands of sext_in_reg based on knowledge that the top bits are not
// demanded.
if (SimplifyDemandedBits(SDValue(N, 0)))
return SDValue(N, 0);
-
+
// fold (sext_in_reg (load x)) -> (smaller sextload x)
// fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
SDValue NarrowLoad = ReduceLoadWidth(N);
}
// fold (sext_inreg (extload x)) -> (sextload x)
- if (ISD::isEXTLoad(N0.getNode()) &&
+ if (ISD::isEXTLoad(N0.getNode()) &&
ISD::isUNINDEXEDLoad(N0.getNode()) &&
EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
SDValue N0 = N->getOperand(0);
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
// noop truncate
if (N0.getValueType() == N->getValueType(0))
}
/// CombineConsecutiveLoads - build_pair (load, load) -> load
-/// if load locations are consecutive.
-SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, MVT VT) {
+/// if load locations are consecutive.
+SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
assert(N->getOpcode() == ISD::BUILD_PAIR);
- SDNode *LD1 = getBuildPairElt(N, 0);
- if (!ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse())
+ LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
+ LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
+ if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse())
return SDValue();
- MVT LD1VT = LD1->getValueType(0);
- SDNode *LD2 = getBuildPairElt(N, 1);
+ EVT LD1VT = LD1->getValueType(0);
const MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
if (ISD::isNON_EXTLoad(LD2) &&
LD2->hasOneUse() &&
// If both are volatile this would reduce the number of volatile loads.
// If one is volatile it might be ok, but play conservative and bail out.
- !cast<LoadSDNode>(LD1)->isVolatile() &&
- !cast<LoadSDNode>(LD2)->isVolatile() &&
+ !LD1->isVolatile() &&
+ !LD2->isVolatile() &&
TLI.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1, MFI)) {
- LoadSDNode *LD = cast<LoadSDNode>(LD1);
- unsigned Align = LD->getAlignment();
+ unsigned Align = LD1->getAlignment();
unsigned NewAlign = TLI.getTargetData()->
- getABITypeAlignment(VT.getTypeForMVT());
+ getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
if (NewAlign <= Align &&
(!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
- return DAG.getLoad(VT, N->getDebugLoc(), LD->getChain(), LD->getBasePtr(),
- LD->getSrcValue(), LD->getSrcValueOffset(),
- false, Align);
+ return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
+ LD1->getBasePtr(), LD1->getSrcValue(),
+ LD1->getSrcValueOffset(), false, Align);
}
return SDValue();
SDValue DAGCombiner::visitBIT_CONVERT(SDNode *N) {
SDValue N0 = N->getOperand(0);
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
// If the input is a BUILD_VECTOR with all constant elements, fold this now.
// Only do this before legalize, since afterward the target may be depending
if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
N0.getOperand(i).getOpcode() != ISD::Constant &&
N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
- isSimple = false;
+ isSimple = false;
break;
}
-
- MVT DestEltVT = N->getValueType(0).getVectorElementType();
+
+ EVT DestEltVT = N->getValueType(0).getVectorElementType();
assert(!DestEltVT.isVector() &&
"Element type of vector ValueType must not be vector!");
if (isSimple)
return ConstantFoldBIT_CONVERTofBUILD_VECTOR(N0.getNode(), DestEltVT);
}
-
+
// If the input is a constant, let getNode fold it.
if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
SDValue Res = DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, N0);
- if (Res.getNode() != N) return Res;
+ if (Res.getNode() != N) {
+ if (!LegalOperations ||
+ TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
+ return Res;
+
+ // Folding it resulted in an illegal node, and it's too late to
+ // do that. Clean up the old node and forego the transformation.
+ // Ideally this won't happen very often, because instcombine
+ // and the earlier dagcombine runs (where illegal nodes are
+ // permitted) should have folded most of them already.
+ DAG.DeleteNode(Res.getNode());
+ }
}
-
+
// (conv (conv x, t1), t2) -> (conv x, t2)
if (N0.getOpcode() == ISD::BIT_CONVERT)
return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT,
(!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
unsigned Align = TLI.getTargetData()->
- getABITypeAlignment(VT.getTypeForMVT());
+ getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
unsigned OrigAlign = LN0->getAlignment();
if (Align <= OrigAlign) {
SDValue NewConv = DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(), VT,
N0.getOperand(0));
AddToWorkList(NewConv.getNode());
-
+
APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
if (N0.getOpcode() == ISD::FNEG)
return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
NewConv, DAG.getConstant(~SignBit, VT));
}
-
+
// fold (bitconvert (fcopysign cst, x)) ->
// (or (and (bitconvert x), sign), (and cst, (not sign)))
// Note that we don't handle (copysign x, cst) because this can always be
isa<ConstantFPSDNode>(N0.getOperand(0)) &&
VT.isInteger() && !VT.isVector()) {
unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
- MVT IntXVT = MVT::getIntegerVT(OrigXWidth);
+ EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
if (TLI.isTypeLegal(IntXVT) || !LegalTypes) {
SDValue X = DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(),
IntXVT, N0.getOperand(1));
X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
AddToWorkList(X.getNode());
}
-
+
APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
X, DAG.getConstant(SignBit, VT));
}
}
- // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
+ // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
if (N0.getOpcode() == ISD::BUILD_PAIR) {
SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
if (CombineLD.getNode())
return CombineLD;
}
-
+
return SDValue();
}
SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
return CombineConsecutiveLoads(N, VT);
}
/// ConstantFoldBIT_CONVERTofBUILD_VECTOR - We know that BV is a build_vector
-/// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
+/// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
/// destination element value type.
SDValue DAGCombiner::
-ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *BV, MVT DstEltVT) {
- MVT SrcEltVT = BV->getOperand(0).getValueType();
-
+ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
+ EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
+
// If this is already the right type, we're done.
if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
-
+
unsigned SrcBitSize = SrcEltVT.getSizeInBits();
unsigned DstBitSize = DstEltVT.getSizeInBits();
-
+
// If this is a conversion of N elements of one type to N elements of another
// type, convert each element. This handles FP<->INT cases.
if (SrcBitSize == DstBitSize) {
SmallVector<SDValue, 8> Ops;
for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
+ SDValue Op = BV->getOperand(i);
+ // If the vector element type is not legal, the BUILD_VECTOR operands
+ // are promoted and implicitly truncated. Make that explicit here.
+ if (Op.getValueType() != SrcEltVT)
+ Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, BV->getDebugLoc(),
- DstEltVT, BV->getOperand(i)));
+ DstEltVT, Op));
AddToWorkList(Ops.back().getNode());
}
- MVT VT = MVT::getVectorVT(DstEltVT,
+ EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
BV->getValueType(0).getVectorNumElements());
return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
&Ops[0], Ops.size());
}
-
+
// Otherwise, we're growing or shrinking the elements. To avoid having to
// handle annoying details of growing/shrinking FP values, we convert them to
// int first.
// Convert the input float vector to a int vector where the elements are the
// same sizes.
assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
- MVT IntVT = MVT::getIntegerVT(SrcEltVT.getSizeInBits());
+ EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
BV = ConstantFoldBIT_CONVERTofBUILD_VECTOR(BV, IntVT).getNode();
SrcEltVT = IntVT;
}
-
+
// Now we know the input is an integer vector. If the output is a FP type,
// convert to integer first, then to FP of the right size.
if (DstEltVT.isFloatingPoint()) {
assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
- MVT TmpVT = MVT::getIntegerVT(DstEltVT.getSizeInBits());
+ EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
SDNode *Tmp = ConstantFoldBIT_CONVERTofBUILD_VECTOR(BV, TmpVT).getNode();
-
+
// Next, convert to FP elements of the same size.
return ConstantFoldBIT_CONVERTofBUILD_VECTOR(Tmp, DstEltVT);
}
-
+
// Okay, we know the src/dst types are both integers of differing types.
// Handling growing first.
assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
if (SrcBitSize < DstBitSize) {
unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
-
+
SmallVector<SDValue, 8> Ops;
for (unsigned i = 0, e = BV->getNumOperands(); i != e;
i += NumInputsPerOutput) {
SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
if (Op.getOpcode() == ISD::UNDEF) continue;
EltIsUndef = false;
-
- NewBits |=
- APInt(cast<ConstantSDNode>(Op)->getAPIntValue()).zext(DstBitSize);
+
+ NewBits |= (APInt(cast<ConstantSDNode>(Op)->getAPIntValue()).
+ zextOrTrunc(SrcBitSize).zext(DstBitSize));
}
-
+
if (EltIsUndef)
- Ops.push_back(DAG.getNode(ISD::UNDEF, BV->getDebugLoc(), DstEltVT));
+ Ops.push_back(DAG.getUNDEF(DstEltVT));
else
Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
}
- MVT VT = MVT::getVectorVT(DstEltVT, Ops.size());
+ EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), 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 VT = MVT::getVectorVT(DstEltVT, NumOutputsPerInput*BV->getNumOperands());
+ EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
+ NumOutputsPerInput*BV->getNumOperands());
SmallVector<SDValue, 8> Ops;
for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
for (unsigned j = 0; j != NumOutputsPerInput; ++j)
- Ops.push_back(DAG.getNode(ISD::UNDEF, BV->getDebugLoc(), DstEltVT));
+ Ops.push_back(DAG.getUNDEF(DstEltVT));
continue;
}
- APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->getAPIntValue();
+ APInt OpVal = APInt(cast<ConstantSDNode>(BV->getOperand(i))->
+ getAPIntValue()).zextOrTrunc(SrcBitSize);
for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
APInt ThisVal = APInt(OpVal).trunc(DstBitSize);
SDValue N1 = N->getOperand(1);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
}
-
+
// fold (fadd c1, c2) -> (fadd c1, c2)
if (N0CFP && N1CFP && VT != MVT::ppcf128)
return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
if (isNegatibleForFree(N0, LegalOperations) == 2)
return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
GetNegatedExpression(N0, DAG, LegalOperations));
-
+
// If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FADD &&
N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
N0.getOperand(1), N1));
-
+
return SDValue();
}
SDValue N1 = N->getOperand(1);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
}
-
+
// fold (fsub c1, c2) -> c1-c2
if (N0CFP && N1CFP && VT != MVT::ppcf128)
return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
if (isNegatibleForFree(N1, LegalOperations))
return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0,
GetNegatedExpression(N1, DAG, LegalOperations));
-
+
return SDValue();
}
SDValue N1 = N->getOperand(1);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
}
-
+
// fold (fmul c1, c2) -> c1*c2
if (N0CFP && N1CFP && VT != MVT::ppcf128)
return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
// fold (fmul A, 0) -> 0
if (UnsafeFPMath && N1CFP && N1CFP->getValueAPF().isZero())
return N1;
+ // fold (fmul A, 0) -> 0, vector edition.
+ if (UnsafeFPMath && ISD::isBuildVectorAllZeros(N1.getNode()))
+ return N1;
// fold (fmul X, 2.0) -> (fadd X, X)
if (N1CFP && N1CFP->isExactlyValue(+2.0))
return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
- // fold (fmul X, (fneg 1.0)) -> (fneg X)
+ // fold (fmul X, -1.0) -> (fneg X)
if (N1CFP && N1CFP->isExactlyValue(-1.0))
if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
-
+
// fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
if (char LHSNeg = isNegatibleForFree(N0, LegalOperations)) {
if (char RHSNeg = isNegatibleForFree(N1, LegalOperations)) {
GetNegatedExpression(N1, DAG, LegalOperations));
}
}
-
+
// If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FMUL &&
N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
- DAG.getNode(ISD::FMUL, VT, N0.getOperand(1), N1));
-
+ DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
+ N0.getOperand(1), N1));
+
return SDValue();
}
SDValue N1 = N->getOperand(1);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
}
-
+
// fold (fdiv c1, c2) -> c1/c2
if (N0CFP && N1CFP && VT != MVT::ppcf128)
return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
-
-
+
+
// (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
if (char LHSNeg = isNegatibleForFree(N0, LegalOperations)) {
if (char RHSNeg = isNegatibleForFree(N1, LegalOperations)) {
// 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, N->getDebugLoc(), VT,
+ return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
GetNegatedExpression(N0, DAG, LegalOperations),
GetNegatedExpression(N1, DAG, LegalOperations));
}
}
-
+
return SDValue();
}
SDValue N1 = N->getOperand(1);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
// fold (frem c1, c2) -> fmod(c1,c2)
if (N0CFP && N1CFP && VT != MVT::ppcf128)
SDValue N1 = N->getOperand(1);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
if (N0CFP && N1CFP && VT != MVT::ppcf128) // Constant fold
return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
-
+
if (N1CFP) {
const APFloat& V = N1CFP->getValueAPF();
// copysign(x, c1) -> fabs(x) iff ispos(c1)
DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
}
}
-
+
// copysign(fabs(x), y) -> copysign(x, y)
// copysign(fneg(x), y) -> copysign(x, y)
// copysign(copysign(x,z), y) -> copysign(x, y)
// copysign(x, abs(y)) -> abs(x)
if (N1.getOpcode() == ISD::FABS)
return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
-
+
// copysign(x, copysign(y,z)) -> copysign(x, z)
if (N1.getOpcode() == ISD::FCOPYSIGN)
return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
N0, N1.getOperand(1));
-
+
// copysign(x, fp_extend(y)) -> copysign(x, y)
// copysign(x, fp_round(y)) -> copysign(x, y)
if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
N0, N1.getOperand(0));
-
+
return SDValue();
}
SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
SDValue N0 = N->getOperand(0);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
- MVT VT = N->getValueType(0);
- MVT OpVT = N0.getValueType();
+ EVT VT = N->getValueType(0);
+ EVT OpVT = N0.getValueType();
// fold (sint_to_fp c1) -> c1fp
if (N0C && OpVT != MVT::ppcf128)
return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
-
+
// If the input is a legal type, and SINT_TO_FP is not legal on this target,
// but UINT_TO_FP is legal on this target, try to convert.
if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
- // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
+ // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
if (DAG.SignBitIsZero(N0))
return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
}
SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
SDValue N0 = N->getOperand(0);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
- MVT VT = N->getValueType(0);
- MVT OpVT = N0.getValueType();
+ EVT VT = N->getValueType(0);
+ EVT OpVT = N0.getValueType();
// fold (uint_to_fp c1) -> c1fp
if (N0C && OpVT != MVT::ppcf128)
return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
-
+
// If the input is a legal type, and UINT_TO_FP is not legal on this target,
// but SINT_TO_FP is legal on this target, try to convert.
if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
- // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
+ // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
if (DAG.SignBitIsZero(N0))
return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
}
-
+
return SDValue();
}
SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
SDValue N0 = N->getOperand(0);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (fp_to_sint c1fp) -> c1
if (N0CFP)
return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
SDValue N0 = N->getOperand(0);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (fp_to_uint c1fp) -> c1
if (N0CFP && VT != MVT::ppcf128)
return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (fp_round c1fp) -> c1fp
if (N0CFP && N0.getValueType() != MVT::ppcf128)
return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), 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.
return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
DAG.getIntPtrConstant(IsTrunc));
}
-
+
// fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
Tmp, N0.getOperand(1));
}
-
+
return SDValue();
}
SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
SDValue N0 = N->getOperand(0);
- MVT VT = N->getValueType(0);
- MVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
+ EVT VT = N->getValueType(0);
+ EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
-
+
// fold (fp_round_inreg c1fp) -> c1fp
if (N0CFP && (TLI.isTypeLegal(EVT) || !LegalTypes)) {
SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
SDValue N0 = N->getOperand(0);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
- if (N->hasOneUse() &&
+ if (N->hasOneUse() &&
N->use_begin()->getOpcode() == ISD::FP_ROUND)
return SDValue();
In, N0.getOperand(1));
return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
}
-
+
// fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
SDValue DAGCombiner::visitFNEG(SDNode *N) {
SDValue N0 = N->getOperand(0);
+ EVT VT = N->getValueType(0);
if (isNegatibleForFree(N0, LegalOperations))
return GetNegatedExpression(N0, DAG, LegalOperations);
// Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
// constant pool values.
- if (N0.getOpcode() == ISD::BIT_CONVERT && N0.getNode()->hasOneUse() &&
- N0.getOperand(0).getValueType().isInteger() &&
- !N0.getOperand(0).getValueType().isVector()) {
+ if (N0.getOpcode() == ISD::BIT_CONVERT &&
+ !VT.isVector() &&
+ N0.getNode()->hasOneUse() &&
+ N0.getOperand(0).getValueType().isInteger()) {
SDValue Int = N0.getOperand(0);
- MVT IntVT = Int.getValueType();
+ EVT IntVT = Int.getValueType();
if (IntVT.isInteger() && !IntVT.isVector()) {
Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
AddToWorkList(Int.getNode());
return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
- N->getValueType(0), Int);
+ VT, Int);
}
}
-
+
return SDValue();
}
SDValue DAGCombiner::visitFABS(SDNode *N) {
SDValue N0 = N->getOperand(0);
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
- MVT VT = N->getValueType(0);
-
+ EVT VT = N->getValueType(0);
+
// fold (fabs c1) -> fabs(c1)
if (N0CFP && VT != MVT::ppcf128)
return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
// fold (fabs (fcopysign x, y)) -> (fabs x)
if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
-
+
// Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
// constant pool values.
if (N0.getOpcode() == ISD::BIT_CONVERT && N0.getNode()->hasOneUse() &&
N0.getOperand(0).getValueType().isInteger() &&
!N0.getOperand(0).getValueType().isVector()) {
SDValue Int = N0.getOperand(0);
- MVT IntVT = Int.getValueType();
+ EVT IntVT = Int.getValueType();
if (IntVT.isInteger() && !IntVT.isVector()) {
- Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
+ Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
AddToWorkList(Int.getNode());
return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
N->getValueType(0), Int);
}
}
-
+
return SDValue();
}
SDValue N1 = N->getOperand(1);
SDValue N2 = N->getOperand(2);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
-
+
// never taken branch, fold to chain
if (N1C && N1C->isNullValue())
return Chain;
return DAG.getNode(ISD::BR, N->getDebugLoc(), MVT::Other, Chain, N2);
// fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
// on the target.
- if (N1.getOpcode() == ISD::SETCC &&
+ if (N1.getOpcode() == ISD::SETCC &&
TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
Chain, N1.getOperand(2),
N1.getOperand(0), N1.getOperand(1), N2);
}
+ if (N1.hasOneUse() && N1.getOpcode() == ISD::SRL) {
+ // Match this pattern so that we can generate simpler code:
+ //
+ // %a = ...
+ // %b = and i32 %a, 2
+ // %c = srl i32 %b, 1
+ // brcond i32 %c ...
+ //
+ // into
+ //
+ // %a = ...
+ // %b = and %a, 2
+ // %c = setcc eq %b, 0
+ // brcond %c ...
+ //
+ // This applies only when the AND constant value has one bit set and the
+ // SRL constant is equal to the log2 of the AND constant. The back-end is
+ // smart enough to convert the result into a TEST/JMP sequence.
+ SDValue Op0 = N1.getOperand(0);
+ SDValue Op1 = N1.getOperand(1);
+
+ if (Op0.getOpcode() == ISD::AND &&
+ Op0.hasOneUse() &&
+ Op1.getOpcode() == ISD::Constant) {
+ SDValue AndOp1 = Op0.getOperand(1);
+
+ if (AndOp1.getOpcode() == ISD::Constant) {
+ const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
+
+ if (AndConst.isPowerOf2() &&
+ cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
+ SDValue SetCC =
+ DAG.getSetCC(N->getDebugLoc(),
+ TLI.getSetCCResultType(Op0.getValueType()),
+ Op0, DAG.getConstant(0, Op0.getValueType()),
+ ISD::SETNE);
+
+ // Replace the uses of SRL with SETCC
+ DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
+ removeFromWorkList(N1.getNode());
+ DAG.DeleteNode(N1.getNode());
+ return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
+ MVT::Other, Chain, SetCC, N2);
+ }
+ }
+ }
+ }
+
return SDValue();
}
SDValue DAGCombiner::visitBR_CC(SDNode *N) {
CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
-
+
// Use SimplifySetCC to simplify SETCC's.
SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
bool isLoad = true;
SDValue Ptr;
- MVT VT;
+ EVT VT;
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
if (LD->isIndexed())
return false;
if (isa<ConstantSDNode>(Offset) &&
cast<ConstantSDNode>(Offset)->isNullValue())
return false;
-
+
// Try turning it into a pre-indexed load / store except when:
// 1) The new base ptr is a frame index.
// 2) If N is a store and the new base ptr is either the same as or is a
// Check #1. Preinc'ing a frame index would require copying the stack pointer
// (plus the implicit offset) to a register to preinc anyway.
- if (isa<FrameIndexSDNode>(BasePtr))
+ if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
return false;
-
+
// Check #2.
if (!isLoad) {
SDValue Val = cast<StoreSDNode>(N)->getValue();
BasePtr, Offset, AM);
++PreIndexedNodes;
++NodesCombined;
- DOUT << "\nReplacing.4 "; DEBUG(N->dump(&DAG));
- DOUT << "\nWith: "; DEBUG(Result.getNode()->dump(&DAG));
- DOUT << '\n';
+ DEBUG(errs() << "\nReplacing.4 ";
+ N->dump(&DAG);
+ errs() << "\nWith: ";
+ Result.getNode()->dump(&DAG);
+ errs() << '\n');
WorkListRemover DeadNodes(*this);
if (isLoad) {
DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
bool isLoad = true;
SDValue Ptr;
- MVT VT;
+ EVT VT;
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
if (LD->isIndexed())
return false;
if (Ptr.getNode()->hasOneUse())
return false;
-
+
for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
E = Ptr.getNode()->use_end(); I != E; ++I) {
SDNode *Op = *I;
SDValue Offset;
ISD::MemIndexedMode AM = ISD::UNINDEXED;
if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
- if (Ptr == Offset)
+ if (Ptr == Offset && Op->getOpcode() == ISD::ADD)
std::swap(BasePtr, Offset);
if (Ptr != BasePtr)
continue;
// nor a successor of N. Otherwise, if Op is folded that would
// create a cycle.
+ if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
+ continue;
+
// Check for #1.
bool TryNext = false;
for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
BasePtr, Offset, AM);
++PostIndexedNodes;
++NodesCombined;
- DOUT << "\nReplacing.5 "; DEBUG(N->dump(&DAG));
- DOUT << "\nWith: "; DEBUG(Result.getNode()->dump(&DAG));
- DOUT << '\n';
+ DEBUG(errs() << "\nReplacing.5 ";
+ N->dump(&DAG);
+ errs() << "\nWith: ";
+ Result.getNode()->dump(&DAG);
+ errs() << '\n');
WorkListRemover DeadNodes(*this);
if (isLoad) {
DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
int64_t FrameOffset = 0;
if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) {
FrameIdx = FI->getIndex();
- } else if (Ptr.getOpcode() == ISD::ADD &&
+ } 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();
// 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),
+ unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx),
FrameOffset);
return std::max(Align, FIInfoAlign);
}
}
-
+
return 0;
}
LoadSDNode *LD = cast<LoadSDNode>(N);
SDValue Chain = LD->getChain();
SDValue Ptr = LD->getBasePtr();
-
+
// Try to infer better alignment information than the load already has.
- if (!Fast && LD->isUnindexed()) {
+ if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
if (unsigned Align = InferAlignment(Ptr, DAG)) {
if (Align > LD->getAlignment())
return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
// 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.getNode()->dump(&DAG));
- DOUT << "\n";
+ DEBUG(errs() << "\nReplacing.6 ";
+ N->dump(&DAG);
+ errs() << "\nWith chain: ";
+ Chain.getNode()->dump(&DAG);
+ errs() << "\n");
WorkListRemover DeadNodes(*this);
DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain, &DeadNodes);
// Indexed loads.
assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
if (N->hasNUsesOfValue(0, 0) && N->hasNUsesOfValue(0, 1)) {
- SDValue Undef = DAG.getNode(ISD::UNDEF, N->getDebugLoc(),
- N->getValueType(0));
- DOUT << "\nReplacing.6 "; DEBUG(N->dump(&DAG));
- DOUT << "\nWith: "; DEBUG(Undef.getNode()->dump(&DAG));
- DOUT << " and 2 other values\n";
+ SDValue Undef = DAG.getUNDEF(N->getValueType(0));
+ DEBUG(errs() << "\nReplacing.6 ";
+ N->dump(&DAG);
+ errs() << "\nWith: ";
+ Undef.getNode()->dump(&DAG);
+ errs() << " and 2 other values\n");
WorkListRemover DeadNodes(*this);
DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef, &DeadNodes);
DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
- DAG.getNode(ISD::UNDEF, N->getDebugLoc(),
- N->getValueType(1)),
+ DAG.getUNDEF(N->getValueType(1)),
&DeadNodes);
DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain, &DeadNodes);
removeFromWorkList(N);
}
}
}
-
+
// If this load is directly stored, replace the load value with the stored
// value.
// TODO: Handle store large -> read small portion.
return CombineTo(N, Chain.getOperand(1), Chain);
}
}
-
+
if (CombinerAA) {
// Walk up chain skipping non-aliasing memory nodes.
SDValue BetterChain = FindBetterChain(N, Chain);
-
+
// If there is a better chain.
if (Chain != BetterChain) {
SDValue ReplLoad;
BetterChain, Ptr, LD->getSrcValue(),
LD->getSrcValueOffset(),
LD->getMemoryVT(),
- LD->isVolatile(),
+ LD->isVolatile(),
LD->getAlignment());
}
SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
MVT::Other, Chain, ReplLoad.getValue(1));
+ // Make sure the new and old chains are cleaned up.
+ AddToWorkList(Token.getNode());
+
// Replace uses with load result and token factor. Don't add users
// to work list.
return CombineTo(N, ReplLoad.getValue(0), Token, false);
return SDValue();
}
+
+/// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
+/// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
+/// of the loaded bits, try narrowing the load and store if it would end up
+/// being a win for performance or code size.
+SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
+ StoreSDNode *ST = cast<StoreSDNode>(N);
+ if (ST->isVolatile())
+ return SDValue();
+
+ SDValue Chain = ST->getChain();
+ SDValue Value = ST->getValue();
+ SDValue Ptr = ST->getBasePtr();
+ EVT VT = Value.getValueType();
+
+ if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
+ return SDValue();
+
+ unsigned Opc = Value.getOpcode();
+ if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
+ Value.getOperand(1).getOpcode() != ISD::Constant)
+ return SDValue();
+
+ SDValue N0 = Value.getOperand(0);
+ if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse()) {
+ LoadSDNode *LD = cast<LoadSDNode>(N0);
+ if (LD->getBasePtr() != Ptr)
+ return SDValue();
+
+ // Find the type to narrow it the load / op / store to.
+ SDValue N1 = Value.getOperand(1);
+ unsigned BitWidth = N1.getValueSizeInBits();
+ APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
+ if (Opc == ISD::AND)
+ Imm ^= APInt::getAllOnesValue(BitWidth);
+ if (Imm == 0 || Imm.isAllOnesValue())
+ return SDValue();
+ unsigned ShAmt = Imm.countTrailingZeros();
+ unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
+ unsigned NewBW = NextPowerOf2(MSB - ShAmt);
+ EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
+ while (NewBW < BitWidth &&
+ !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
+ TLI.isNarrowingProfitable(VT, NewVT))) {
+ NewBW = NextPowerOf2(NewBW);
+ NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
+ }
+ if (NewBW >= BitWidth)
+ return SDValue();
+
+ // If the lsb changed does not start at the type bitwidth boundary,
+ // start at the previous one.
+ if (ShAmt % NewBW)
+ ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
+ APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, ShAmt + NewBW);
+ if ((Imm & Mask) == Imm) {
+ APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
+ if (Opc == ISD::AND)
+ NewImm ^= APInt::getAllOnesValue(NewBW);
+ uint64_t PtrOff = ShAmt / 8;
+ // For big endian targets, we need to adjust the offset to the pointer to
+ // load the correct bytes.
+ if (TLI.isBigEndian())
+ PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
+
+ unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
+ if (NewAlign <
+ TLI.getTargetData()->getABITypeAlignment(NewVT.getTypeForEVT(*DAG.getContext())))
+ return SDValue();
+
+ SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
+ Ptr.getValueType(), Ptr,
+ DAG.getConstant(PtrOff, Ptr.getValueType()));
+ SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
+ LD->getChain(), NewPtr,
+ LD->getSrcValue(), LD->getSrcValueOffset(),
+ LD->isVolatile(), NewAlign);
+ SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
+ DAG.getConstant(NewImm, NewVT));
+ SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
+ NewVal, NewPtr,
+ ST->getSrcValue(), ST->getSrcValueOffset(),
+ false, NewAlign);
+
+ AddToWorkList(NewPtr.getNode());
+ AddToWorkList(NewLD.getNode());
+ AddToWorkList(NewVal.getNode());
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1),
+ &DeadNodes);
+ ++OpsNarrowed;
+ return NewST;
+ }
+ }
+
+ return SDValue();
+}
+
SDValue DAGCombiner::visitSTORE(SDNode *N) {
StoreSDNode *ST = cast<StoreSDNode>(N);
SDValue Chain = ST->getChain();
SDValue Value = ST->getValue();
SDValue Ptr = ST->getBasePtr();
-
+
// Try to infer better alignment information than the store already has.
- if (!Fast && ST->isUnindexed()) {
+ if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
if (unsigned Align = InferAlignment(Ptr, DAG)) {
if (Align > ST->getAlignment())
return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
// resultant store does not need a higher alignment than the original.
if (Value.getOpcode() == ISD::BIT_CONVERT && !ST->isTruncatingStore() &&
ST->isUnindexed()) {
- unsigned Align = ST->getAlignment();
- MVT SVT = Value.getOperand(0).getValueType();
- unsigned OrigAlign = TLI.getTargetData()->
- getABITypeAlignment(SVT.getTypeForMVT());
+ unsigned OrigAlign = ST->getAlignment();
+ EVT SVT = Value.getOperand(0).getValueType();
+ unsigned Align = TLI.getTargetData()->
+ getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
if (Align <= OrigAlign &&
((!LegalOperations && !ST->isVolatile()) ||
TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
// transform should not be done in this case.
if (Value.getOpcode() != ISD::TargetConstantFP) {
SDValue Tmp;
- switch (CFP->getValueType(0).getSimpleVT()) {
- default: assert(0 && "Unknown FP type");
+ switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
+ default: llvm_unreachable("Unknown FP type");
case MVT::f80: // We don't do this for these yet.
case MVT::f128:
case MVT::ppcf128:
}
}
- if (CombinerAA) {
+ if (CombinerAA) {
// Walk up chain skipping non-aliasing memory nodes.
SDValue BetterChain = FindBetterChain(N, Chain);
-
+
// If there is a better chain.
if (Chain != BetterChain) {
- // Replace the chain to avoid dependency.
SDValue ReplStore;
+
+ // Replace the chain to avoid dependency.
if (ST->isTruncatingStore()) {
ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
ST->getSrcValue(),ST->getSrcValueOffset(),
ST->getSrcValue(), ST->getSrcValueOffset(),
ST->isVolatile(), ST->getAlignment());
}
-
+
// Create token to keep both nodes around.
SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
MVT::Other, Chain, ReplStore);
+ // Make sure the new and old chains are cleaned up.
+ AddToWorkList(Token.getNode());
+
// Don't add users to work list.
return CombineTo(N, Token, false);
}
}
-
+
// Try transforming N to an indexed store.
if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
return SDValue(N, 0);
// 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"
- SDValue Shorter =
+ SDValue Shorter =
GetDemandedBits(Value,
APInt::getLowBitsSet(Value.getValueSizeInBits(),
ST->getMemoryVT().getSizeInBits()));
Ptr, ST->getSrcValue(),
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,
ST->getMemoryVT().getSizeInBits())))
return SDValue(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)) {
ST->isVolatile(), ST->getAlignment());
}
- return SDValue();
+ return ReduceLoadOpStoreWidth(N);
}
SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
SDValue InVec = N->getOperand(0);
SDValue InVal = N->getOperand(1);
SDValue EltNo = N->getOperand(2);
-
+
// If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new
// vector with the inserted element.
if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
InVec.getValueType(), &Ops[0], Ops.size());
}
-
+ // If the invec is an UNDEF and if EltNo is a constant, create a new
+ // BUILD_VECTOR with undef elements and the inserted element.
+ if (!LegalOperations && InVec.getOpcode() == ISD::UNDEF &&
+ isa<ConstantSDNode>(EltNo)) {
+ EVT VT = InVec.getValueType();
+ EVT EltVT = VT.getVectorElementType();
+ unsigned NElts = VT.getVectorNumElements();
+ SmallVector<SDValue, 8> Ops(NElts, DAG.getUNDEF(EltVT));
+
+ unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
+ if (Elt < Ops.size())
+ Ops[Elt] = InVal;
+ return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
+ InVec.getValueType(), &Ops[0], Ops.size());
+ }
return SDValue();
}
// (vextract (scalar_to_vector val, 0) -> val
SDValue InVec = N->getOperand(0);
- if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR)
- return InVec.getOperand(0);
+ if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
+ // If the operand is wider than the vector element type then it is implicitly
+ // truncated. Make that explicit here.
+ EVT EltVT = InVec.getValueType().getVectorElementType();
+ SDValue InOp = InVec.getOperand(0);
+ if (InOp.getValueType() != EltVT)
+ return DAG.getNode(ISD::TRUNCATE, InVec.getDebugLoc(), EltVT, InOp);
+ return InOp;
+ }
// Perform only after legalization to ensure build_vector / vector_shuffle
// optimizations have already been done.
unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
bool NewLoad = false;
bool BCNumEltsChanged = false;
- MVT VT = InVec.getValueType();
- MVT EVT = VT.getVectorElementType();
- MVT LVT = EVT;
+ EVT VT = InVec.getValueType();
+ EVT ExtVT = VT.getVectorElementType();
+ EVT LVT = ExtVT;
if (InVec.getOpcode() == ISD::BIT_CONVERT) {
- MVT BCVT = InVec.getOperand(0).getValueType();
- if (!BCVT.isVector() || EVT.bitsGT(BCVT.getVectorElementType()))
+ EVT BCVT = InVec.getOperand(0).getValueType();
+ if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
return SDValue();
if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
BCNumEltsChanged = true;
InVec = InVec.getOperand(0);
- EVT = BCVT.getVectorElementType();
+ ExtVT = BCVT.getVectorElementType();
NewLoad = true;
}
LoadSDNode *LN0 = NULL;
+ const ShuffleVectorSDNode *SVN = NULL;
if (ISD::isNormalLoad(InVec.getNode())) {
LN0 = cast<LoadSDNode>(InVec);
} else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
- InVec.getOperand(0).getValueType() == EVT &&
+ InVec.getOperand(0).getValueType() == ExtVT &&
ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
LN0 = cast<LoadSDNode>(InVec.getOperand(0));
- } else if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE) {
+ } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
// (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
// =>
// (load $addr+1*size)
-
+
// If the bit convert changed the number of elements, it is unsafe
// to examine the mask.
if (BCNumEltsChanged)
return SDValue();
- unsigned Idx = cast<ConstantSDNode>(InVec.getOperand(2).
- getOperand(Elt))->getZExtValue();
- unsigned NumElems = InVec.getOperand(2).getNumOperands();
- InVec = (Idx < NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
+
+ // Select the input vector, guarding against out of range extract vector.
+ unsigned NumElems = VT.getVectorNumElements();
+ int Idx = (Elt > NumElems) ? -1 : SVN->getMaskElt(Elt);
+ InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
+
if (InVec.getOpcode() == ISD::BIT_CONVERT)
InVec = InVec.getOperand(0);
if (ISD::isNormalLoad(InVec.getNode())) {
LN0 = cast<LoadSDNode>(InVec);
- Elt = (Idx < NumElems) ? Idx : Idx - NumElems;
+ Elt = (Idx < (int)NumElems) ? Idx : Idx - NumElems;
}
}
// Check the resultant load doesn't need a higher alignment than the
// original load.
unsigned NewAlign =
- TLI.getTargetData()->getABITypeAlignment(LVT.getTypeForMVT());
+ TLI.getTargetData()->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
return SDValue();
SDValue NewPtr = LN0->getBasePtr();
if (Elt) {
unsigned PtrOff = LVT.getSizeInBits() * Elt / 8;
- MVT PtrType = NewPtr.getValueType();
+ EVT PtrType = NewPtr.getValueType();
if (TLI.isBigEndian())
PtrOff = VT.getSizeInBits() / 8 - PtrOff;
NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
unsigned NumInScalars = N->getNumOperands();
- MVT VT = N->getValueType(0);
- unsigned NumElts = VT.getVectorNumElements();
- MVT EltType = VT.getVectorElementType();
+ EVT VT = N->getValueType(0);
// Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
// operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
for (unsigned i = 0; i != NumInScalars; ++i) {
// Ignore undef inputs.
if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
-
+
// If this input is something other than a EXTRACT_VECTOR_ELT with a
// constant index, bail out.
if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
VecIn1 = VecIn2 = SDValue(0, 0);
break;
}
-
+
// If the input vector type disagrees with the result of the build_vector,
// we can't make a shuffle.
SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
VecIn1 = VecIn2 = SDValue(0, 0);
break;
}
-
+
// Otherwise, remember this. We allow up to two distinct input vectors.
if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
continue;
-
+
if (VecIn1.getNode() == 0) {
VecIn1 = ExtractedFromVec;
} else if (VecIn2.getNode() == 0) {
break;
}
}
-
+
// If everything is good, we can make a shuffle operation.
if (VecIn1.getNode()) {
- SmallVector<SDValue, 8> BuildVecIndices;
+ SmallVector<int, 8> Mask;
for (unsigned i = 0; i != NumInScalars; ++i) {
if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
- BuildVecIndices.push_back(DAG.getNode(ISD::UNDEF,
- N->getDebugLoc(),
- TLI.getPointerTy()));
+ Mask.push_back(-1);
continue;
}
-
- SDValue Extract = N->getOperand(i);
-
+
// If extracting from the first vector, just use the index directly.
+ SDValue Extract = N->getOperand(i);
+ SDValue ExtVal = Extract.getOperand(1);
if (Extract.getOperand(0) == VecIn1) {
- BuildVecIndices.push_back(Extract.getOperand(1));
+ unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
+ if (ExtIndex > VT.getVectorNumElements())
+ return SDValue();
+
+ Mask.push_back(ExtIndex);
continue;
}
// Otherwise, use InIdx + VecSize
- unsigned Idx =
- cast<ConstantSDNode>(Extract.getOperand(1))->getZExtValue();
- BuildVecIndices.push_back(DAG.getIntPtrConstant(Idx+NumInScalars));
+ unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
+ Mask.push_back(Idx+NumInScalars);
}
-
+
// Add count and size info.
- MVT BuildVecVT = MVT::getVectorVT(TLI.getPointerTy(), NumElts);
- if (!TLI.isTypeLegal(BuildVecVT) && LegalTypes)
+ if (!TLI.isTypeLegal(VT) && LegalTypes)
return SDValue();
// Return the new VECTOR_SHUFFLE node.
- SDValue Ops[5];
+ SDValue Ops[2];
Ops[0] = VecIn1;
- if (VecIn2.getNode()) {
- Ops[1] = VecIn2;
- } else {
- // Use an undef build_vector as input for the second operand.
- std::vector<SDValue> UnOps(NumInScalars,
- DAG.getNode(ISD::UNDEF, N->getDebugLoc(),
- EltType));
- Ops[1] = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
- &UnOps[0], UnOps.size());
- AddToWorkList(Ops[1].getNode());
- }
-
- Ops[2] = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), BuildVecVT,
- &BuildVecIndices[0], BuildVecIndices.size());
- return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getDebugLoc(), VT, Ops, 3);
+ Ops[1] = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
+ return DAG.getVectorShuffle(VT, N->getDebugLoc(), Ops[0], Ops[1], &Mask[0]);
}
-
+
return SDValue();
}
}
SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
- SDValue ShufMask = N->getOperand(2);
- unsigned NumElts = ShufMask.getNumOperands();
+ return SDValue();
+
+ EVT VT = N->getValueType(0);
+ unsigned NumElts = VT.getVectorNumElements();
SDValue N0 = N->getOperand(0);
- SDValue N1 = N->getOperand(1);
assert(N0.getValueType().getVectorNumElements() == NumElts &&
"Vector shuffle must be normalized in DAG");
- // If the shuffle mask is an identity operation on the LHS, return the LHS.
- bool isIdentity = true;
- for (unsigned i = 0; i != NumElts; ++i) {
- if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
- cast<ConstantSDNode>(ShufMask.getOperand(i))->getZExtValue() != i) {
- isIdentity = false;
- break;
- }
- }
- if (isIdentity) return N->getOperand(0);
-
- // If the shuffle mask is an identity operation on the RHS, return the RHS.
- isIdentity = true;
- for (unsigned i = 0; i != NumElts; ++i) {
- if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
- cast<ConstantSDNode>(ShufMask.getOperand(i))->getZExtValue() !=
- i+NumElts) {
- isIdentity = false;
- break;
- }
- }
- if (isIdentity) return N->getOperand(1);
-
- // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
- // needed at all.
- bool isUnary = true;
- bool isSplat = true;
- int VecNum = -1;
- unsigned BaseIdx = 0;
- for (unsigned i = 0; i != NumElts; ++i)
- if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
- unsigned Idx=cast<ConstantSDNode>(ShufMask.getOperand(i))->getZExtValue();
- int V = (Idx < NumElts) ? 0 : 1;
- if (VecNum == -1) {
- VecNum = V;
- BaseIdx = Idx;
- } else {
- if (BaseIdx != Idx)
- isSplat = false;
- if (VecNum != V) {
- isUnary = false;
- break;
- }
- }
- }
-
- // Normalize unary shuffle so the RHS is undef.
- if (isUnary && VecNum == 1)
- std::swap(N0, N1);
+ // FIXME: implement canonicalizations from DAG.getVectorShuffle()
// If it is a splat, check if the argument vector is a build_vector with
// all scalar elements the same.
- if (isSplat) {
+ if (cast<ShuffleVectorSDNode>(N)->isSplat()) {
SDNode *V = N0.getNode();
+
// If this is a bit convert that changes the element type of the vector but
// not the number of vector elements, look through it. Be careful not to
if (V->getOpcode() == ISD::BUILD_VECTOR) {
unsigned NumElems = V->getNumOperands();
+ unsigned BaseIdx = cast<ShuffleVectorSDNode>(N)->getSplatIndex();
if (NumElems > BaseIdx) {
SDValue Base;
bool AllSame = true;
}
}
}
-
- // If it is a unary or the LHS and the RHS are the same node, turn the RHS
- // into an undef.
- if (isUnary || N0 == N1) {
- // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
- // first operand.
- SmallVector<SDValue, 8> MappedOps;
-
- for (unsigned i = 0; i != NumElts; ++i) {
- if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
- cast<ConstantSDNode>(ShufMask.getOperand(i))->getZExtValue() <
- NumElts) {
- MappedOps.push_back(ShufMask.getOperand(i));
- } else {
- unsigned NewIdx =
- cast<ConstantSDNode>(ShufMask.getOperand(i))->getZExtValue() -
- NumElts;
- MappedOps.push_back(DAG.getConstant(NewIdx,
- ShufMask.getOperand(i).getValueType()));
- }
- }
-
- ShufMask = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
- ShufMask.getValueType(),
- &MappedOps[0], MappedOps.size());
- AddToWorkList(ShufMask.getNode());
- return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getDebugLoc(),
- N->getValueType(0), N0,
- DAG.getNode(ISD::UNDEF, N->getDebugLoc(),
- N->getValueType(0)),
- ShufMask);
- }
-
return SDValue();
}
/// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
/// vector_shuffle V, Zero, <0, 4, 2, 4>
SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
+ EVT VT = N->getValueType(0);
+ DebugLoc dl = N->getDebugLoc();
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
if (N->getOpcode() == ISD::AND) {
if (RHS.getOpcode() == ISD::BIT_CONVERT)
RHS = RHS.getOperand(0);
if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
- std::vector<SDValue> IdxOps;
- unsigned NumOps = RHS.getNumOperands();
- unsigned NumElts = NumOps;
+ SmallVector<int, 8> Indices;
+ unsigned NumElts = RHS.getNumOperands();
for (unsigned i = 0; i != NumElts; ++i) {
SDValue Elt = RHS.getOperand(i);
if (!isa<ConstantSDNode>(Elt))
return SDValue();
else if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
- IdxOps.push_back(DAG.getIntPtrConstant(i));
+ Indices.push_back(i);
else if (cast<ConstantSDNode>(Elt)->isNullValue())
- IdxOps.push_back(DAG.getIntPtrConstant(NumElts));
+ Indices.push_back(NumElts);
else
return SDValue();
}
// Let's see if the target supports this vector_shuffle.
- if (!TLI.isVectorClearMaskLegal(IdxOps, TLI.getPointerTy(), DAG))
+ EVT RVT = RHS.getValueType();
+ if (!TLI.isVectorClearMaskLegal(Indices, RVT))
return SDValue();
// Return the new VECTOR_SHUFFLE node.
- MVT EVT = RHS.getValueType().getVectorElementType();
- MVT VT = MVT::getVectorVT(EVT, NumElts);
- MVT MaskVT = MVT::getVectorVT(TLI.getPointerTy(), NumElts);
- std::vector<SDValue> Ops;
- LHS = DAG.getNode(ISD::BIT_CONVERT, LHS.getDebugLoc(), VT, LHS);
- Ops.push_back(LHS);
- AddToWorkList(LHS.getNode());
- std::vector<SDValue> ZeroOps(NumElts, DAG.getConstant(0, EVT));
- Ops.push_back(DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
- VT, &ZeroOps[0], ZeroOps.size()));
- Ops.push_back(DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
- MaskVT, &IdxOps[0], IdxOps.size()));
- SDValue Result = DAG.getNode(ISD::VECTOR_SHUFFLE, N->getDebugLoc(),
- VT, &Ops[0], Ops.size());
-
- if (VT != N->getValueType(0))
- Result = DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
- N->getValueType(0), Result);
-
- return Result;
+ EVT EltVT = RVT.getVectorElementType();
+ SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
+ DAG.getConstant(0, EltVT));
+ SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
+ RVT, &ZeroOps[0], ZeroOps.size());
+ LHS = DAG.getNode(ISD::BIT_CONVERT, dl, RVT, LHS);
+ SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuf);
}
}
// things. Simplifying them may result in a loss of legality.
if (LegalOperations) return SDValue();
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
assert(VT.isVector() && "SimplifyVBinOp only works on vectors!");
- MVT EltType = VT.getVectorElementType();
+ EVT EltType = VT.getVectorElementType();
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
SDValue Shuffle = XformToShuffleWithZero(N);
// If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
// this operation.
- if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
+ if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
RHS.getOpcode() == ISD::BUILD_VECTOR) {
SmallVector<SDValue, 8> Ops;
for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
Ops.back().getOpcode() == ISD::ConstantFP) &&
"Scalar binop didn't fold!");
}
-
+
if (Ops.size() == LHS.getNumOperands()) {
- MVT VT = LHS.getValueType();
+ EVT VT = LHS.getValueType();
return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
&Ops[0], Ops.size());
}
}
-
+
return SDValue();
}
SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
SDValue N1, SDValue N2){
assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
-
+
SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
cast<CondCodeSDNode>(N0.getOperand(2))->get());
if (SCC.getOpcode() == ISD::SELECT_CC) {
SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
N0.getValueType(),
- SCC.getOperand(0), SCC.getOperand(1),
+ SCC.getOperand(0), SCC.getOperand(1),
SCC.getOperand(4));
AddToWorkList(SETCC.getNode());
return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
/// returns true. As such, they should return the appropriate thing (e.g. the
/// node) back to the top-level of the DAG combiner loop to avoid it being
/// looked at.
-bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
+bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
SDValue RHS) {
-
+
// If this is a select from two identical things, try to pull the operation
// through the select.
if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){
Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
LLD->getBasePtr().getValueType(),
TheSelect->getOperand(0),
- TheSelect->getOperand(1),
+ TheSelect->getOperand(1),
LLD->getBasePtr(), RLD->getBasePtr(),
TheSelect->getOperand(4));
}
}
-
+
if (Addr.getNode()) {
SDValue Load;
if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
Load = DAG.getLoad(TheSelect->getValueType(0),
TheSelect->getDebugLoc(),
LLD->getChain(),
- Addr,LLD->getSrcValue(),
+ Addr,LLD->getSrcValue(),
LLD->getSrcValueOffset(),
- LLD->isVolatile(),
+ LLD->isVolatile(),
LLD->getAlignment());
} else {
Load = DAG.getExtLoad(LLD->getExtensionType(),
LLD->getChain(), Addr, LLD->getSrcValue(),
LLD->getSrcValueOffset(),
LLD->getMemoryVT(),
- LLD->isVolatile(),
+ LLD->isVolatile(),
LLD->getAlignment());
}
// Users of the select now use the result of the load.
CombineTo(TheSelect, Load);
-
+
// Users of the old loads now use the new load's chain. We know the
// old-load value is dead now.
CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
}
}
}
-
+
return false;
}
-SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
+/// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
+/// where 'cond' is the comparison specified by CC.
+SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
SDValue N2, SDValue N3,
ISD::CondCode CC, bool NotExtCompare) {
- MVT VT = N2.getValueType();
+ // (x ? y : y) -> y.
+ if (N2 == N3) return N2;
+
+ EVT VT = N2.getValueType();
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
// fold select_cc false, x, y -> y
if (SCCC && SCCC->isNullValue())
return N3;
-
+
// Check to see if we can simplify the select into an fabs node
if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
// Allow either -0.0 or 0.0
N0 == N2 && N3.getOpcode() == ISD::FNEG &&
N2 == N3.getOperand(0))
return DAG.getNode(ISD::FABS, DL, VT, N0);
-
+
// select (setl[te] X, +/-0.0), fneg(X), X -> fabs
if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
N0 == N3 && N2.getOpcode() == ISD::FNEG &&
}
}
+ // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
+ // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
+ // in it. This is a win when the constant is not otherwise available because
+ // it replaces two constant pool loads with one. We only do this if the FP
+ // type is known to be legal, because if it isn't, then we are before legalize
+ // types an we want the other legalization to happen first (e.g. to avoid
+ // messing with soft float) and if the ConstantFP is not legal, because if
+ // it is legal, we may not need to store the FP constant in a constant pool.
+ if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
+ if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
+ if (TLI.isTypeLegal(N2.getValueType()) &&
+ (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
+ TargetLowering::Legal) &&
+ // If both constants have multiple uses, then we won't need to do an
+ // extra load, they are likely around in registers for other users.
+ (TV->hasOneUse() || FV->hasOneUse())) {
+ Constant *Elts[] = {
+ const_cast<ConstantFP*>(FV->getConstantFPValue()),
+ const_cast<ConstantFP*>(TV->getConstantFPValue())
+ };
+ const Type *FPTy = Elts[0]->getType();
+ const TargetData &TD = *TLI.getTargetData();
+
+ // Create a ConstantArray of the two constants.
+ Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts, 2);
+ SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
+ TD.getPrefTypeAlignment(FPTy));
+ unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
+
+ // Get the offsets to the 0 and 1 element of the array so that we can
+ // select between them.
+ SDValue Zero = DAG.getIntPtrConstant(0);
+ unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
+ SDValue One = DAG.getIntPtrConstant(EltSize);
+
+ SDValue Cond = DAG.getSetCC(DL,
+ TLI.getSetCCResultType(N0.getValueType()),
+ N0, N1, CC);
+ SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
+ Cond, One, Zero);
+ CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
+ CstOffset);
+ return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
+ PseudoSourceValue::getConstantPool(), 0, false,
+ Alignment);
+
+ }
+ }
+
// Check to see if we can perform the "gzip trick", transforming
// (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
N2.getValueType().isInteger() &&
(N1C->isNullValue() || // (a < 0) ? b : 0
(N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
- MVT XType = N0.getValueType();
- MVT AType = N2.getValueType();
+ EVT XType = N0.getValueType();
+ EVT AType = N2.getValueType();
if (XType.bitsGE(AType)) {
// and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
// single-bit constant.
return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
}
}
-
+
// fold select C, 16, 0 -> shl C, 4
if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent) {
-
+
// If the caller doesn't want us to simplify this into a zext of a compare,
// don't do it.
if (NotExtCompare && N2C->getAPIntValue() == 1)
return SDValue();
-
+
// Get a SetCC of the condition
// FIXME: Should probably make sure that setcc is legal if we ever have a
// target where it isn't.
AddToWorkList(SCC.getNode());
AddToWorkList(Temp.getNode());
-
+
if (N2C->getAPIntValue() == 1)
return Temp;
DAG.getConstant(N2C->getAPIntValue().logBase2(),
getShiftAmountTy()));
}
-
+
// Check to see if this is the equivalent of setcc
// FIXME: Turn all of these into setcc if setcc if setcc is legal
// otherwise, go ahead with the folds.
if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
- MVT XType = N0.getValueType();
+ EVT XType = N0.getValueType();
if (!LegalOperations ||
TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
return Res;
}
-
+
// fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
- if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
+ if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
(!LegalOperations ||
TLI.isOperationLegal(ISD::CTLZ, XType))) {
SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
- return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
+ return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
DAG.getConstant(Log2_32(XType.getSizeInBits()),
getShiftAmountTy()));
}
// fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
- if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
+ if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
XType, DAG.getConstant(0, XType), N0);
SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
}
}
-
+
// Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
// Y = sra (X, size(X)-1); xor (add (X, Y), Y)
if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) &&
N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1) &&
N2.getOperand(0) == N1 && N0.getValueType().isInteger()) {
- MVT XType = N0.getValueType();
+ EVT XType = N0.getValueType();
SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType, N0,
DAG.getConstant(XType.getSizeInBits()-1,
getShiftAmountTy()));
if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT &&
N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1)) {
if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0))) {
- MVT XType = N0.getValueType();
+ EVT XType = N0.getValueType();
if (SubC->isNullValue() && XType.isInteger()) {
SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
N0,
}
}
}
-
+
return SDValue();
}
/// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
-SDValue DAGCombiner::SimplifySetCC(MVT VT, SDValue N0,
+SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
SDValue N1, ISD::CondCode Cond,
DebugLoc DL, bool foldBooleans) {
- TargetLowering::DAGCombinerInfo
- DagCombineInfo(DAG, Level == Unrestricted, false, this);
+ TargetLowering::DAGCombinerInfo
+ DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
}
return S;
}
-/// FindBaseOffset - Return true if base is known not to alias with anything
-/// but itself. Provides base object and offset as results.
-static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset) {
+/// FindBaseOffset - Return true if base is a frame index, which is known not
+// to alias with anything but itself. Provides base object and offset as results.
+static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
+ GlobalValue *&GV, void *&CV) {
// Assume it is a primitive operation.
- Base = Ptr; Offset = 0;
-
+ Base = Ptr; Offset = 0; GV = 0; CV = 0;
+
// If it's an adding a simple constant then integrate the offset.
if (Base.getOpcode() == ISD::ADD) {
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
}
}
+ // Return the underlying GlobalValue, and update the Offset. Return false
+ // for GlobalAddressSDNode since the same GlobalAddress may be represented
+ // by multiple nodes with different offsets.
+ if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
+ GV = G->getGlobal();
+ Offset += G->getOffset();
+ return false;
+ }
+
+ // Return the underlying Constant value, and update the Offset. Return false
+ // for ConstantSDNodes since the same constant pool entry may be represented
+ // by multiple nodes with different offsets.
+ if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
+ CV = C->isMachineConstantPoolEntry() ? (void *)C->getMachineCPVal()
+ : (void *)C->getConstVal();
+ Offset += C->getOffset();
+ return false;
+ }
// If it's any of the following then it can't alias with anything but itself.
- return isa<FrameIndexSDNode>(Base) ||
- isa<ConstantPoolSDNode>(Base) ||
- isa<GlobalAddressSDNode>(Base);
+ return isa<FrameIndexSDNode>(Base);
}
/// isAlias - Return true if there is any possibility that the two addresses
/// overlap.
bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
const Value *SrcValue1, int SrcValueOffset1,
+ unsigned SrcValueAlign1,
SDValue Ptr2, int64_t Size2,
- const Value *SrcValue2, int SrcValueOffset2) const {
+ const Value *SrcValue2, int SrcValueOffset2,
+ unsigned SrcValueAlign2) const {
// If they are the same then they must be aliases.
if (Ptr1 == Ptr2) return true;
-
+
// Gather base node and offset information.
SDValue Base1, Base2;
int64_t Offset1, Offset2;
- bool KnownBase1 = FindBaseOffset(Ptr1, Base1, Offset1);
- bool KnownBase2 = FindBaseOffset(Ptr2, Base2, Offset2);
-
- // If they have a same base address then...
- if (Base1 == Base2)
- // Check to see if the addresses overlap.
+ GlobalValue *GV1, *GV2;
+ void *CV1, *CV2;
+ bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
+ bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
+
+ // If they have a same base address then check to see if they overlap.
+ if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
-
- // If we know both bases then they can't alias.
- if (KnownBase1 && KnownBase2) return false;
+ // If we know what the bases are, and they aren't identical, then we know they
+ // cannot alias.
+ if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
+ return false;
+
+ // If we know required SrcValue1 and SrcValue2 have relatively large alignment
+ // compared to the size and offset of the access, we may be able to prove they
+ // do not alias. This check is conservative for now to catch cases created by
+ // splitting vector types.
+ if ((SrcValueAlign1 == SrcValueAlign2) &&
+ (SrcValueOffset1 != SrcValueOffset2) &&
+ (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
+ int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
+ int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
+
+ // There is no overlap between these relatively aligned accesses of similar
+ // size, return no alias.
+ if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
+ return false;
+ }
+
if (CombinerGlobalAA) {
// Use alias analysis information.
int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
- AliasAnalysis::AliasResult AAResult =
+ AliasAnalysis::AliasResult AAResult =
AA.alias(SrcValue1, Overlap1, SrcValue2, Overlap2);
if (AAResult == AliasAnalysis::NoAlias)
return false;
/// node. Returns true if the operand was a load.
bool DAGCombiner::FindAliasInfo(SDNode *N,
SDValue &Ptr, int64_t &Size,
- const Value *&SrcValue, int &SrcValueOffset) const {
+ const Value *&SrcValue,
+ int &SrcValueOffset,
+ unsigned &SrcValueAlign) const {
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
Ptr = LD->getBasePtr();
Size = LD->getMemoryVT().getSizeInBits() >> 3;
SrcValue = LD->getSrcValue();
SrcValueOffset = LD->getSrcValueOffset();
+ SrcValueAlign = LD->getOriginalAlignment();
return true;
} else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
Ptr = ST->getBasePtr();
Size = ST->getMemoryVT().getSizeInBits() >> 3;
SrcValue = ST->getSrcValue();
SrcValueOffset = ST->getSrcValueOffset();
+ SrcValueAlign = ST->getOriginalAlignment();
} else {
- assert(0 && "FindAliasInfo expected a memory operand");
+ llvm_unreachable("FindAliasInfo expected a memory operand");
}
-
+
return false;
}
void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
SmallVector<SDValue, 8> &Aliases) {
SmallVector<SDValue, 8> Chains; // List of chains to visit.
- std::set<SDNode *> Visited; // Visited node set.
-
+ SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
+
// Get alias information for node.
SDValue Ptr;
int64_t Size;
const Value *SrcValue;
int SrcValueOffset;
- bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset);
+ unsigned SrcValueAlign;
+ bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
+ SrcValueAlign);
// Starting off.
Chains.push_back(OriginalChain);
+ unsigned Depth = 0;
// Look at each chain and determine if it is an alias. If so, add it to the
// aliases list. If not, then continue up the chain looking for the next
- // candidate.
+ // candidate.
while (!Chains.empty()) {
SDValue Chain = Chains.back();
Chains.pop_back();
- // Don't bother if we've been before.
- if (Visited.find(Chain.getNode()) != Visited.end()) continue;
- Visited.insert(Chain.getNode());
-
+ // For TokenFactor nodes, look at each operand and only continue up the
+ // chain until we find two aliases. If we've seen two aliases, assume we'll
+ // find more and revert to original chain since the xform is unlikely to be
+ // profitable.
+ //
+ // FIXME: The depth check could be made to return the last non-aliasing
+ // chain we found before we hit a tokenfactor rather than the original
+ // chain.
+ if (Depth > 6 || Aliases.size() == 2) {
+ Aliases.clear();
+ Aliases.push_back(OriginalChain);
+ break;
+ }
+
+ // Don't bother if we've been before.
+ if (!Visited.insert(Chain.getNode()))
+ continue;
+
switch (Chain.getOpcode()) {
case ISD::EntryToken:
// Entry token is ideal chain operand, but handled in FindBetterChain.
break;
-
+
case ISD::LOAD:
case ISD::STORE: {
// Get alias information for Chain.
int64_t OpSize;
const Value *OpSrcValue;
int OpSrcValueOffset;
+ unsigned OpSrcValueAlign;
bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
- OpSrcValue, OpSrcValueOffset);
-
+ OpSrcValue, OpSrcValueOffset,
+ OpSrcValueAlign);
+
// If chain is alias then stop here.
if (!(IsLoad && IsOpLoad) &&
- isAlias(Ptr, Size, SrcValue, SrcValueOffset,
- OpPtr, OpSize, OpSrcValue, OpSrcValueOffset)) {
+ isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
+ OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
+ OpSrcValueAlign)) {
Aliases.push_back(Chain);
} else {
// Look further up the chain.
- Chains.push_back(Chain.getOperand(0));
- // Clean up old chain.
- AddToWorkList(Chain.getNode());
+ Chains.push_back(Chain.getOperand(0));
+ ++Depth;
}
break;
}
-
+
case ISD::TokenFactor:
- // We have to check each of the operands of the token factor, so we queue
- // then up. Adding the operands to the queue (stack) in reverse order
- // maintains the original order and increases the likelihood that getNode
- // will find a matching token factor (CSE.)
+ // We have to check each of the operands of the token factor for "small"
+ // token factors, so we queue them up. Adding the operands to the queue
+ // (stack) in reverse order maintains the original order and increases the
+ // likelihood that getNode will find a matching token factor (CSE.)
+ if (Chain.getNumOperands() > 16) {
+ Aliases.push_back(Chain);
+ break;
+ }
for (unsigned n = Chain.getNumOperands(); n;)
Chains.push_back(Chain.getOperand(--n));
- // Eliminate the token factor if we can.
- AddToWorkList(Chain.getNode());
+ ++Depth;
break;
-
+
default:
// For all other instructions we will just have to take what we can get.
Aliases.push_back(Chain);
/// for a better chain (aliasing node.)
SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
-
+
// Accumulate all the aliases to this node.
GatherAllAliases(N, OldChain, Aliases);
-
+
if (Aliases.size() == 0) {
// If no operands then chain to entry token.
return DAG.getEntryNode();
// If a single operand then chain to it. We don't need to revisit it.
return Aliases[0];
}
-
- // Construct a custom tailored token factor.
- SDValue NewChain = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
- &Aliases[0], Aliases.size());
-
- // Make sure the old chain gets cleaned up.
- if (NewChain != OldChain) AddToWorkList(OldChain.getNode());
- return NewChain;
+ // Construct a custom tailored token factor.
+ return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
+ &Aliases[0], Aliases.size());
}
// SelectionDAG::Combine - This is the entry point for the file.
//
-void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA, bool Fast) {
+void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
+ CodeGenOpt::Level OptLevel) {
/// run - This is the main entry point to this class.
///
- DAGCombiner(*this, AA, Fast).Run(Level);
+ DAGCombiner(*this, AA, OptLevel).Run(Level);
}