#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Target/TargetData.h"
-#include "llvm/Target/TargetFrameInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
+STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
namespace {
static cl::opt<bool>
SDValue visitANY_EXTEND(SDNode *N);
SDValue visitSIGN_EXTEND_INREG(SDNode *N);
SDValue visitTRUNCATE(SDNode *N);
- SDValue visitBIT_CONVERT(SDNode *N);
+ SDValue visitBITCAST(SDNode *N);
SDValue visitBUILD_PAIR(SDNode *N);
SDValue visitFADD(SDNode *N);
SDValue visitFSUB(SDNode *N);
SDValue visitBUILD_VECTOR(SDNode *N);
SDValue visitCONCAT_VECTORS(SDNode *N);
SDValue visitVECTOR_SHUFFLE(SDNode *N);
+ SDValue visitMEMBARRIER(SDNode *N);
SDValue XformToShuffleWithZero(SDNode *N);
SDValue ReassociateOps(unsigned Opc, DebugLoc DL, SDValue LHS, SDValue RHS);
SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
unsigned HiOp);
SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
- SDValue ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *, EVT);
+ SDValue ConstantFoldBITCASTofBUILD_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 TransformFPLoadStorePair(SDNode *N);
SDValue GetDemandedBits(SDValue V, const APInt &Mask);
bool isAlias(SDValue Ptr1, int64_t Size1,
const Value *SrcValue1, int SrcValueOffset1,
unsigned SrcValueAlign1,
+ const MDNode *TBAAInfo1,
SDValue Ptr2, int64_t Size2,
const Value *SrcValue2, int SrcValueOffset2,
- unsigned SrcValueAlign2) const;
+ unsigned SrcValueAlign2,
+ const MDNode *TBAAInfo2) 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,
- unsigned &SrcValueAlignment) const;
+ unsigned &SrcValueAlignment,
+ const MDNode *&TBAAInfo) const;
/// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
/// looking for a better chain (aliasing node.)
/// Run - runs the dag combiner on all nodes in the work list
void Run(CombineLevel AtLevel);
-
+
SelectionDAG &getDAG() const { return DAG; }
-
+
/// getShiftAmountTy - Returns a type large enough to hold any valid
/// shift amount - before type legalization these can be huge.
EVT getShiftAmountTy() {
return LegalTypes ? TLI.getShiftAmountTy() : TLI.getPointerTy();
}
-
+
/// isTypeLegal - This method returns true if we are running before type
/// legalization or if the specified VT is legal.
bool isTypeLegal(const EVT &VT) {
// Replace the old value with the new one.
++NodesCombined;
- DEBUG(dbgs() << "\nReplacing.2 ";
+ DEBUG(dbgs() << "\nReplacing.2 ";
TLO.Old.getNode()->dump(&DAG);
dbgs() << "\nWith: ";
TLO.New.getNode()->dump(&DAG);
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
EVT MemVT = LD->getMemoryVT();
ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
- ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD : ISD::EXTLOAD)
+ ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
+ : ISD::EXTLOAD)
: LD->getExtensionType();
Replace = true;
return DAG.getExtLoad(ExtType, dl, PVT,
LD->getChain(), LD->getBasePtr(),
- LD->getSrcValue(), LD->getSrcValueOffset(),
+ LD->getPointerInfo(),
MemVT, LD->isVolatile(),
LD->isNonTemporal(), LD->getAlignment());
}
unsigned ExtOpc =
Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
return DAG.getNode(ExtOpc, dl, PVT, Op);
- }
+ }
}
if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
bool Replace1 = false;
SDValue N1 = Op.getOperand(1);
- SDValue NN1 = PromoteOperand(N1, PVT, Replace1);
- if (NN1.getNode() == 0)
- return SDValue();
+ SDValue NN1;
+ if (N0 == N1)
+ NN1 = NN0;
+ else {
+ NN1 = PromoteOperand(N1, PVT, Replace1);
+ if (NN1.getNode() == 0)
+ return SDValue();
+ }
AddToWorkList(NN0.getNode());
- AddToWorkList(NN1.getNode());
+ if (NN1.getNode())
+ AddToWorkList(NN1.getNode());
if (Replace0)
ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
LoadSDNode *LD = cast<LoadSDNode>(N);
EVT MemVT = LD->getMemoryVT();
ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
- ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD : ISD::EXTLOAD)
+ ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
+ : ISD::EXTLOAD)
: LD->getExtensionType();
SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
LD->getChain(), LD->getBasePtr(),
- LD->getSrcValue(), LD->getSrcValueOffset(),
+ LD->getPointerInfo(),
MemVT, LD->isVolatile(),
LD->isNonTemporal(), LD->getAlignment());
SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
"Node was deleted but visit returned new node!");
- DEBUG(dbgs() << "\nReplacing.3 ";
+ DEBUG(dbgs() << "\nReplacing.3 ";
N->dump(&DAG);
dbgs() << "\nWith: ";
RV.getNode()->dump(&DAG);
}
SDValue DAGCombiner::visit(SDNode *N) {
- switch(N->getOpcode()) {
+ switch (N->getOpcode()) {
default: break;
case ISD::TokenFactor: return visitTokenFactor(N);
case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
case ISD::TRUNCATE: return visitTRUNCATE(N);
- case ISD::BIT_CONVERT: return visitBIT_CONVERT(N);
+ case ISD::BITCAST: return visitBITCAST(N);
case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
case ISD::FADD: return visitFADD(N);
case ISD::FSUB: return visitFSUB(N);
case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
+ case ISD::MEMBARRIER: return visitMEMBARRIER(N);
}
return SDValue();
}
}
}
+ // If nothing happened still, try promoting the operation.
+ if (RV.getNode() == 0) {
+ switch (N->getOpcode()) {
+ default: break;
+ case ISD::ADD:
+ case ISD::SUB:
+ case ISD::MUL:
+ case ISD::AND:
+ case ISD::OR:
+ case ISD::XOR:
+ RV = PromoteIntBinOp(SDValue(N, 0));
+ break;
+ case ISD::SHL:
+ case ISD::SRA:
+ case ISD::SRL:
+ RV = PromoteIntShiftOp(SDValue(N, 0));
+ break;
+ case ISD::SIGN_EXTEND:
+ case ISD::ZERO_EXTEND:
+ case ISD::ANY_EXTEND:
+ RV = PromoteExtend(SDValue(N, 0));
+ break;
+ case ISD::LOAD:
+ if (PromoteLoad(SDValue(N, 0)))
+ RV = SDValue(N, 0);
+ break;
+ }
+ }
+
// If N is a commutative binary node, try commuting it to enable more
// sdisel CSE.
if (RV.getNode() == 0 &&
}
}
}
-
+
SDValue Result;
// If we've change things around then replace token factor.
if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
GA->getOpcode() == ISD::GlobalAddress)
- return DAG.getGlobalAddress(GA->getGlobal(), VT,
+ return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
GA->getOffset() +
(uint64_t)N1C->getSExtValue());
// fold ((c1-A)+c2) -> (c1+c2)-A
N0.getOperand(0).getOperand(1),
N0.getOperand(1)));
- return PromoteIntBinOp(SDValue(N, 0));
+ if (N1.getOpcode() == ISD::AND) {
+ SDValue AndOp0 = N1.getOperand(0);
+ ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
+ unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
+ unsigned DestBits = VT.getScalarType().getSizeInBits();
+
+ // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
+ // and similar xforms where the inner op is either ~0 or 0.
+ if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
+ DebugLoc DL = N->getDebugLoc();
+ return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
+ }
+ }
+
+ // add (sext i1), X -> sub X, (zext i1)
+ if (N0.getOpcode() == ISD::SIGN_EXTEND &&
+ N0.getOperand(0).getValueType() == MVT::i1 &&
+ !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
+ DebugLoc DL = N->getDebugLoc();
+ SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
+ return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
+ }
+
+ return SDValue();
}
SDValue DAGCombiner::visitADDC(SDNode *N) {
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));
+ N->getDebugLoc(), MVT::Glue));
// canonicalize constant to RHS.
if (N0C && !N1C)
// 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));
+ N->getDebugLoc(), MVT::Glue));
// fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
APInt LHSZero, LHSOne;
(LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
return CombineTo(N, DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1),
DAG.getNode(ISD::CARRY_FALSE,
- N->getDebugLoc(), MVT::Flag));
+ N->getDebugLoc(), MVT::Glue));
}
return SDValue();
return SDValue();
}
+// Since it may not be valid to emit a fold to zero for vector initializers
+// check if we can before folding.
+static SDValue tryFoldToZero(DebugLoc DL, const TargetLowering &TLI, EVT VT,
+ SelectionDAG &DAG, bool LegalOperations) {
+ if (!VT.isVector()) {
+ return DAG.getConstant(0, VT);
+ } else if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
+ // Produce a vector of zeros.
+ SDValue El = DAG.getConstant(0, VT.getVectorElementType());
+ std::vector<SDValue> Ops(VT.getVectorNumElements(), El);
+ return DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
+ &Ops[0], Ops.size());
+ }
+ return SDValue();
+}
+
SDValue DAGCombiner::visitSUB(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
}
// fold (sub x, x) -> 0
+ // FIXME: Refactor this and xor and other similar operations together.
if (N0 == N1)
- return DAG.getConstant(0, N->getValueType(0));
+ return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
// fold (sub c1, c2) -> c1-c2
if (N0C && N1C)
return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
// Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
if (N0C && N0C->isAllOnesValue())
return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
+ // fold A-(A-B) -> B
+ if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
+ return N1.getOperand(1);
// fold (A+B)-A -> B
if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
return N0.getOperand(1);
if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
// fold (sub Sym, c) -> Sym-c
if (N1C && GA->getOpcode() == ISD::GlobalAddress)
- return DAG.getGlobalAddress(GA->getGlobal(), VT,
+ return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
GA->getOffset() -
(uint64_t)N1C->getSExtValue());
// fold (sub Sym+c1, Sym+c2) -> c1-c2
VT);
}
- return PromoteIntBinOp(SDValue(N, 0));
+ return SDValue();
}
SDValue DAGCombiner::visitMUL(SDNode *N) {
if (RMUL.getNode() != 0)
return RMUL;
- return PromoteIntBinOp(SDValue(N, 0));
+ return SDValue();
}
SDValue DAGCombiner::visitSDIV(SDNode *N) {
SDValue N1 = N->getOperand(1);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
EVT VT = N->getValueType(0);
+ DebugLoc DL = N->getDebugLoc();
// fold (mulhs x, 0) -> 0
if (N1C && N1C->isNullValue())
if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
return DAG.getConstant(0, VT);
+ // If the type twice as wide is legal, transform the mulhs to a wider multiply
+ // plus a shift.
+ if (VT.isSimple() && !VT.isVector()) {
+ MVT Simple = VT.getSimpleVT();
+ unsigned SimpleSize = Simple.getSizeInBits();
+ EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
+ if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
+ N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
+ N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
+ N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
+ N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
+ DAG.getConstant(SimpleSize, getShiftAmountTy()));
+ return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
+ }
+ }
+
return SDValue();
}
SDValue N1 = N->getOperand(1);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
EVT VT = N->getValueType(0);
+ DebugLoc DL = N->getDebugLoc();
// fold (mulhu x, 0) -> 0
if (N1C && N1C->isNullValue())
if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
return DAG.getConstant(0, VT);
+ // If the type twice as wide is legal, transform the mulhu to a wider multiply
+ // plus a shift.
+ if (VT.isSimple() && !VT.isVector()) {
+ MVT Simple = VT.getSimpleVT();
+ unsigned SimpleSize = Simple.getSizeInBits();
+ EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
+ if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
+ N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
+ N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
+ N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
+ N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
+ DAG.getConstant(SimpleSize, getShiftAmountTy()));
+ return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
+ }
+ }
+
return SDValue();
}
SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
if (Res.getNode()) return Res;
+ EVT VT = N->getValueType(0);
+ DebugLoc DL = N->getDebugLoc();
+
+ // If the type twice as wide is legal, transform the mulhu to a wider multiply
+ // plus a shift.
+ if (VT.isSimple() && !VT.isVector()) {
+ MVT Simple = VT.getSimpleVT();
+ unsigned SimpleSize = Simple.getSizeInBits();
+ EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
+ if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
+ SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
+ SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
+ Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
+ // Compute the high part as N1.
+ Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
+ DAG.getConstant(SimpleSize, getShiftAmountTy()));
+ Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
+ // Compute the low part as N0.
+ Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
+ return CombineTo(N, Lo, Hi);
+ }
+ }
+
return SDValue();
}
SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
if (Res.getNode()) return Res;
+ EVT VT = N->getValueType(0);
+ DebugLoc DL = N->getDebugLoc();
+
+ // If the type twice as wide is legal, transform the mulhu to a wider multiply
+ // plus a shift.
+ if (VT.isSimple() && !VT.isVector()) {
+ MVT Simple = VT.getSimpleVT();
+ unsigned SimpleSize = Simple.getSizeInBits();
+ EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
+ if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
+ SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
+ SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
+ Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
+ // Compute the high part as N1.
+ Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
+ DAG.getConstant(SimpleSize, getShiftAmountTy()));
+ Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
+ // Compute the low part as N0.
+ Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
+ return CombineTo(N, Lo, Hi);
+ }
+ }
+
return SDValue();
}
// 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)
//
// do not sink logical op inside of a vector extend, since it may combine
// into a vsetcc.
// Avoid infinite looping with PromoteIntBinOp.
(N0.getOpcode() == ISD::ANY_EXTEND &&
(!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
- (N0.getOpcode() == ISD::TRUNCATE && TLI.isTypeLegal(Op0VT))) &&
+ (N0.getOpcode() == ISD::TRUNCATE &&
+ (!TLI.isZExtFree(VT, Op0VT) ||
+ !TLI.isTruncateFree(Op0VT, VT)) &&
+ TLI.isTypeLegal(Op0VT))) &&
!VT.isVector() &&
Op0VT == N1.getOperand(0).getValueType() &&
(!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
SDValue N0Op0 = N0.getOperand(0);
APInt Mask = ~N1C->getAPIntValue();
- Mask.trunc(N0Op0.getValueSizeInBits());
+ Mask = Mask.trunc(N0Op0.getValueSizeInBits());
if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(),
N0.getValueType(), N0Op0);
TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
LN0->getChain(), LN0->getBasePtr(),
- LN0->getSrcValue(),
- LN0->getSrcValueOffset(), MemVT,
+ LN0->getPointerInfo(), MemVT,
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
AddToWorkList(N);
TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(), MemVT,
+ LN0->getBasePtr(), LN0->getPointerInfo(),
+ MemVT,
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
AddToWorkList(N);
if (ExtVT == LoadedVT &&
(!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
-
- SDValue NewLoad =
+
+ SDValue NewLoad =
DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
LN0->getChain(), LN0->getBasePtr(),
- LN0->getSrcValue(), LN0->getSrcValueOffset(),
+ LN0->getPointerInfo(),
ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
AddToWorkList(N);
CombineTo(LN0, NewLoad, NewLoad.getValue(1));
return SDValue(N, 0); // Return N so it doesn't get rechecked!
}
-
+
// Do not change the width of a volatile load.
// 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).
}
AddToWorkList(NewPtr.getNode());
-
+
EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
SDValue Load =
DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
LN0->getChain(), NewPtr,
- LN0->getSrcValue(), LN0->getSrcValueOffset(),
+ LN0->getPointerInfo(),
ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
Alignment);
AddToWorkList(N);
}
}
- return PromoteIntBinOp(SDValue(N, 0));
+ return SDValue();
}
SDValue DAGCombiner::visitOR(SDNode *N) {
}
// fold (or x, undef) -> -1
- if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF) {
+ if (!LegalOperations &&
+ (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
}
if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
return SDValue(Rot, 0);
- return PromoteIntBinOp(SDValue(N, 0));
+ // Simplify the operands using demanded-bits information.
+ if (!VT.isVector() &&
+ SimplifyDemandedBits(SDValue(N, 0)))
+ return SDValue(N, 0);
+
+ return SDValue();
}
/// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
N01C->getAPIntValue(), VT));
}
// fold (xor x, x) -> 0
- if (N0 == N1) {
- if (!VT.isVector()) {
- return DAG.getConstant(0, VT);
- } else if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)){
- // Produce a vector of zeros.
- SDValue El = DAG.getConstant(0, VT.getVectorElementType());
- std::vector<SDValue> Ops(VT.getVectorNumElements(), El);
- return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
- &Ops[0], Ops.size());
- }
- }
+ if (N0 == N1)
+ return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
// Simplify: xor (op x...), (op y...) -> (op (xor x, y))
if (N0.getOpcode() == N1.getOpcode()) {
SimplifyDemandedBits(SDValue(N, 0)))
return SDValue(N, 0);
- return PromoteIntBinOp(SDValue(N, 0));
+ return SDValue();
}
/// visitShiftByConstant - Handle transforms common to the three shifts, when
LHS->getOperand(1), N->getOperand(1));
// Create the new shift.
- SDValue NewShift = DAG.getNode(N->getOpcode(), LHS->getOperand(0).getDebugLoc(),
+ SDValue NewShift = DAG.getNode(N->getOpcode(),
+ LHS->getOperand(0).getDebugLoc(),
VT, LHS->getOperand(0), N->getOperand(1));
// Create the new binop.
EVT TruncVT = N1.getValueType();
SDValue N100 = N1.getOperand(0).getOperand(0);
APInt TruncC = N101C->getAPIntValue();
- TruncC.trunc(TruncVT.getSizeInBits());
+ TruncC = TruncC.trunc(TruncVT.getSizeInBits());
return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
DAG.getNode(ISD::TRUNCATE,
N0.getOperand(1).getOpcode() == ISD::Constant) {
uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
uint64_t c2 = N1C->getZExtValue();
- if (c1 + c2 > OpSizeInBits)
+ if (c1 + c2 >= OpSizeInBits)
return DAG.getConstant(0, VT);
return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
DAG.getConstant(c1 + c2, N1.getValueType()));
}
+
+ // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
+ // For this to be valid, the second form must not preserve any of the bits
+ // that are shifted out by the inner shift in the first form. This means
+ // the outer shift size must be >= the number of bits added by the ext.
+ // As a corollary, we don't care what kind of ext it is.
+ if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
+ N0.getOpcode() == ISD::ANY_EXTEND ||
+ N0.getOpcode() == ISD::SIGN_EXTEND) &&
+ N0.getOperand(0).getOpcode() == ISD::SHL &&
+ isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
+ uint64_t c1 =
+ cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
+ uint64_t c2 = N1C->getZExtValue();
+ EVT InnerShiftVT = N0.getOperand(0).getValueType();
+ uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
+ if (c2 >= OpSizeInBits - InnerShiftSize) {
+ if (c1 + c2 >= OpSizeInBits)
+ return DAG.getConstant(0, VT);
+ return DAG.getNode(ISD::SHL, N0->getDebugLoc(), VT,
+ DAG.getNode(N0.getOpcode(), N0->getDebugLoc(), VT,
+ N0.getOperand(0)->getOperand(0)),
+ DAG.getConstant(c1 + c2, N1.getValueType()));
+ }
+ }
+
// 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 &&
return NewSHL;
}
- return PromoteIntShiftOp(SDValue(N, 0));
+ return SDValue();
}
SDValue DAGCombiner::visitSRA(SDNode *N) {
if (N01C && N1C) {
// Determine what the truncate's result bitsize and type would be.
EVT TruncVT =
- EVT::getIntegerVT(*DAG.getContext(), OpSizeInBits - N1C->getZExtValue());
+ EVT::getIntegerVT(*DAG.getContext(),
+ OpSizeInBits - N1C->getZExtValue());
// Determine the residual right-shift amount.
signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
EVT TruncVT = N1.getValueType();
SDValue N100 = N1.getOperand(0).getOperand(0);
APInt TruncC = N101C->getAPIntValue();
- TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
+ TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
DAG.getNode(ISD::AND, N->getDebugLoc(),
TruncVT,
}
}
+ // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
+ // if c1 is equal to the number of bits the trunc removes
+ if (N0.getOpcode() == ISD::TRUNCATE &&
+ (N0.getOperand(0).getOpcode() == ISD::SRL ||
+ N0.getOperand(0).getOpcode() == ISD::SRA) &&
+ N0.getOperand(0).hasOneUse() &&
+ N0.getOperand(0).getOperand(1).hasOneUse() &&
+ N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
+ EVT LargeVT = N0.getOperand(0).getValueType();
+ ConstantSDNode *LargeShiftAmt =
+ cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
+
+ if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
+ LargeShiftAmt->getZExtValue()) {
+ SDValue Amt =
+ DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
+ getShiftAmountTy());
+ SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), LargeVT,
+ N0.getOperand(0).getOperand(0), Amt);
+ return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, SRA);
+ }
+ }
+
// Simplify, based on bits shifted out of the LHS.
if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
return SDValue(N, 0);
return NewSRA;
}
- return PromoteIntShiftOp(SDValue(N, 0));
+ return SDValue();
}
SDValue DAGCombiner::visitSRL(SDNode *N) {
N0.getOperand(1).getOpcode() == ISD::Constant) {
uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
uint64_t c2 = N1C->getZExtValue();
- if (c1 + c2 > OpSizeInBits)
+ if (c1 + c2 >= OpSizeInBits)
return DAG.getConstant(0, VT);
return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
DAG.getConstant(c1 + c2, N1.getValueType()));
}
-
+
+ // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
+ if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
+ N0.getOperand(0).getOpcode() == ISD::SRL &&
+ isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
+ uint64_t c1 =
+ cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
+ uint64_t c2 = N1C->getZExtValue();
+ EVT InnerShiftVT = N0.getOperand(0).getValueType();
+ EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
+ uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
+ // This is only valid if the OpSizeInBits + c1 = size of inner shift.
+ if (c1 + OpSizeInBits == InnerShiftSize) {
+ if (c1 + c2 >= InnerShiftSize)
+ return DAG.getConstant(0, VT);
+ return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
+ DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT,
+ N0.getOperand(0)->getOperand(0),
+ DAG.getConstant(c1 + c2, ShiftCountVT)));
+ }
+ }
+
// fold (srl (shl x, c), c) -> (and x, cst2)
if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
N0.getValueSizeInBits() <= 64) {
return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
DAG.getConstant(~0ULL >> ShAmt, VT));
}
-
+
// fold (srl (anyextend x), c) -> (anyextend (srl x, c))
if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
EVT TruncVT = N1.getValueType();
SDValue N100 = N1.getOperand(0).getOperand(0);
APInt TruncC = N101C->getAPIntValue();
- TruncC.trunc(TruncVT.getSizeInBits());
+ TruncC = TruncC.trunc(TruncVT.getSizeInBits());
return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
DAG.getNode(ISD::AND, N->getDebugLoc(),
TruncVT,
return NewSRL;
}
+ // Attempt to convert a srl of a load into a narrower zero-extending load.
+ SDValue NarrowLoad = ReduceLoadWidth(N);
+ if (NarrowLoad.getNode())
+ return NarrowLoad;
+
// Here is a common situation. We want to optimize:
//
// %a = ...
// brcond i32 %c ...
//
// into
- //
+ //
// %a = ...
// %b = and %a, 2
// %c = setcc eq %b, 0
}
}
- return PromoteIntShiftOp(SDValue(N, 0));
+ return SDValue();
}
SDValue DAGCombiner::visitCTLZ(SDNode *N) {
}
if (BothLiveOut)
// Both unextended and extended values are live out. There had better be
- // good a reason for the transformation.
+ // a good reason for the transformation.
return ExtendNodes.size();
}
return true;
// fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
if (NarrowLoad.getNode()) {
- if (NarrowLoad.getNode() != N0.getNode())
+ SDNode* oye = N0.getNode()->getOperand(0).getNode();
+ if (NarrowLoad.getNode() != N0.getNode()) {
CombineTo(N0.getNode(), NarrowLoad);
+ // CombineTo deleted the truncate, if needed, but not what's under it.
+ AddToWorkList(oye);
+ }
return SDValue(N, 0); // Return N so it doesn't get rechecked!
}
}
// fold (sext (load x)) -> (sext (truncate (sextload x)))
- if (ISD::isNON_EXTLoad(N0.getNode()) &&
+ // None of the supported targets knows how to perform load and sign extend
+ // in one instruction. We only perform this transformation on scalars.
+ if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
bool DoXform = true;
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(),
+ LN0->getBasePtr(), LN0->getPointerInfo(),
N0.getValueType(),
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(), MemVT,
+ LN0->getBasePtr(), LN0->getPointerInfo(),
+ MemVT,
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
CombineTo(N, ExtLoad);
if (N0.getOpcode() == ISD::SETCC) {
// sext(setcc) -> sext_in_reg(vsetcc) for vectors.
- if (VT.isVector() &&
+ // Only do this before legalize for now.
+ if (VT.isVector() && !LegalOperations) {
+ EVT N0VT = N0.getOperand(0).getValueType();
// 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());
+ if (VT.getSizeInBits() == N0VT.getSizeInBits())
+ return DAG.getVSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
+ N0.getOperand(1),
+ cast<CondCodeSDNode>(N0.getOperand(2))->get());
+ // If the desired elements are smaller or larger than the source
+ // elements we can use a matching integer vector type and then
+ // truncate/sign extend
+ else {
+ EVT MatchingElementType =
+ EVT::getIntegerVT(*DAG.getContext(),
+ N0VT.getScalarType().getSizeInBits());
+ EVT MatchingVectorType =
+ EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
+ N0VT.getVectorNumElements());
+ SDValue VsetCC =
+ DAG.getVSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
+ N0.getOperand(1),
+ cast<CondCodeSDNode>(N0.getOperand(2))->get());
+ return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
+ }
}
-
+
// sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
unsigned ElementWidth = VT.getScalarType().getSizeInBits();
SDValue NegOne =
cast<CondCodeSDNode>(N0.getOperand(2))->get()),
NegOne, DAG.getConstant(0, VT));
}
-
-
// 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 PromoteExtend(SDValue(N, 0));
+ return SDValue();
}
SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
if (N0.getOpcode() == ISD::TRUNCATE) {
SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
if (NarrowLoad.getNode()) {
- if (NarrowLoad.getNode() != N0.getNode())
+ SDNode* oye = N0.getNode()->getOperand(0).getNode();
+ if (NarrowLoad.getNode() != N0.getNode()) {
CombineTo(N0.getNode(), NarrowLoad);
+ // CombineTo deleted the truncate, if needed, but not what's under it.
+ AddToWorkList(oye);
+ }
return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, NarrowLoad);
}
}
// fold (zext (truncate x)) -> (and x, mask)
if (N0.getOpcode() == ISD::TRUNCATE &&
- (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT)) &&
- (!TLI.isTruncateFree(N0.getOperand(0).getValueType(),
- N0.getValueType()) ||
- !TLI.isZExtFree(N0.getValueType(), VT))) {
+ (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
+
+ // fold (zext (truncate (load x))) -> (zext (smaller load x))
+ // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
+ SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
+ if (NarrowLoad.getNode()) {
+ SDNode* oye = N0.getNode()->getOperand(0).getNode();
+ if (NarrowLoad.getNode() != N0.getNode()) {
+ CombineTo(N0.getNode(), NarrowLoad);
+ // CombineTo deleted the truncate, if needed, but not what's under it.
+ AddToWorkList(oye);
+ }
+ return SDValue(N, 0); // Return N so it doesn't get rechecked!
+ }
+
SDValue Op = N0.getOperand(0);
if (Op.getValueType().bitsLT(VT)) {
Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
}
APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
- Mask.zext(VT.getSizeInBits());
+ Mask = Mask.zext(VT.getSizeInBits());
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()) &&
+ // None of the supported targets knows how to perform load and vector_zext
+ // in one instruction. We only perform this transformation on scalar zext.
+ if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
bool DoXform = true;
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(),
+ LN0->getBasePtr(), LN0->getPointerInfo(),
N0.getValueType(),
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(), MemVT,
+ LN0->getBasePtr(), LN0->getPointerInfo(),
+ MemVT,
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
CombineTo(N, ExtLoad);
}
}
- // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
if (N0.getOpcode() == ISD::SETCC) {
+ if (!LegalOperations && VT.isVector()) {
+ // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
+ // Only do this before legalize for now.
+ EVT N0VT = N0.getOperand(0).getValueType();
+ EVT EltVT = VT.getVectorElementType();
+ SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
+ DAG.getConstant(1, EltVT));
+ if (VT.getSizeInBits() == N0VT.getSizeInBits()) {
+ // 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.
+ return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
+ DAG.getVSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
+ N0.getOperand(1),
+ cast<CondCodeSDNode>(N0.getOperand(2))->get()),
+ DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
+ &OneOps[0], OneOps.size()));
+ } else {
+ // If the desired elements are smaller or larger than the source
+ // elements we can use a matching integer vector type and then
+ // truncate/sign extend
+ EVT MatchingElementType =
+ EVT::getIntegerVT(*DAG.getContext(),
+ N0VT.getScalarType().getSizeInBits());
+ EVT MatchingVectorType =
+ EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
+ N0VT.getVectorNumElements());
+ SDValue VsetCC =
+ DAG.getVSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
+ N0.getOperand(1),
+ cast<CondCodeSDNode>(N0.getOperand(2))->get());
+ return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
+ DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT),
+ DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
+ &OneOps[0], OneOps.size()));
+ }
+ }
+
+ // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
SDValue SCC =
SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
DAG.getConstant(1, VT), DAG.getConstant(0, VT),
isa<ConstantSDNode>(N0.getOperand(1)) &&
N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
N0.hasOneUse()) {
+ SDValue ShAmt = N0.getOperand(1);
+ unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
if (N0.getOpcode() == ISD::SHL) {
+ SDValue InnerZExt = N0.getOperand(0);
// If the original shl may be shifting out bits, do not perform this
// transformation.
- unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
- unsigned KnownZeroBits = N0.getOperand(0).getValueType().getSizeInBits() -
- N0.getOperand(0).getOperand(0).getValueType().getSizeInBits();
- if (ShAmt > KnownZeroBits)
+ unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
+ InnerZExt.getOperand(0).getValueType().getSizeInBits();
+ if (ShAmtVal > KnownZeroBits)
return SDValue();
}
- DebugLoc dl = N->getDebugLoc();
- return DAG.getNode(N0.getOpcode(), dl, VT,
- DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0.getOperand(0)),
- DAG.getNode(ISD::ZERO_EXTEND, dl,
- N0.getOperand(1).getValueType(),
- N0.getOperand(1)));
+
+ DebugLoc DL = N->getDebugLoc();
+
+ // Ensure that the shift amount is wide enough for the shifted value.
+ if (VT.getSizeInBits() >= 256)
+ ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
+
+ return DAG.getNode(N0.getOpcode(), DL, VT,
+ DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
+ ShAmt);
}
- return PromoteExtend(SDValue(N, 0));
+ return SDValue();
}
SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
if (N0.getOpcode() == ISD::TRUNCATE) {
SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
if (NarrowLoad.getNode()) {
- if (NarrowLoad.getNode() != N0.getNode())
+ SDNode* oye = N0.getNode()->getOperand(0).getNode();
+ if (NarrowLoad.getNode() != N0.getNode()) {
CombineTo(N0.getNode(), NarrowLoad);
+ // CombineTo deleted the truncate, if needed, but not what's under it.
+ AddToWorkList(oye);
+ }
return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, NarrowLoad);
}
}
X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
}
APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
- Mask.zext(VT.getSizeInBits());
+ Mask = Mask.zext(VT.getSizeInBits());
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()) &&
+ // None of the supported targets knows how to perform load and any_ext
+ // in one instruction. We only perform this transformation on scalars.
+ if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
bool DoXform = true;
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(),
+ LN0->getBasePtr(), LN0->getPointerInfo(),
N0.getValueType(),
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
EVT MemVT = LN0->getMemoryVT();
SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
VT, LN0->getChain(), LN0->getBasePtr(),
- LN0->getSrcValue(),
- LN0->getSrcValueOffset(), MemVT,
+ LN0->getPointerInfo(), MemVT,
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
CombineTo(N, ExtLoad);
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) {
+ // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
+ // Only do this before legalize for now.
+ if (VT.isVector() && !LegalOperations) {
+ EVT N0VT = N0.getOperand(0).getValueType();
+ // 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.
+ if (VT.getSizeInBits() == N0VT.getSizeInBits())
+ return DAG.getVSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
+ N0.getOperand(1),
+ cast<CondCodeSDNode>(N0.getOperand(2))->get());
+ // If the desired elements are smaller or larger than the source
+ // elements we can use a matching integer vector type and then
+ // truncate/sign extend
+ else {
+ EVT MatchingElementType =
+ EVT::getIntegerVT(*DAG.getContext(),
+ N0VT.getScalarType().getSizeInBits());
+ EVT MatchingVectorType =
+ EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
+ N0VT.getVectorNumElements());
+ SDValue VsetCC =
+ DAG.getVSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
+ N0.getOperand(1),
+ cast<CondCodeSDNode>(N0.getOperand(2))->get());
+ return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
+ }
+ }
+
+ // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
SDValue SCC =
SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
DAG.getConstant(1, VT), DAG.getConstant(0, VT),
return SCC;
}
- return PromoteExtend(SDValue(N, 0));
+ return SDValue();
}
/// GetDemandedBits - See if the specified operand can be simplified with the
/// extended, also fold the extension to form a extending load.
SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
unsigned Opc = N->getOpcode();
+
ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
SDValue N0 = N->getOperand(0);
EVT VT = N->getValueType(0);
if (Opc == ISD::SIGN_EXTEND_INREG) {
ExtType = ISD::SEXTLOAD;
ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
- if (LegalOperations && !TLI.isLoadExtLegal(ISD::SEXTLOAD, ExtVT))
- return SDValue();
- }
+ } else if (Opc == ISD::SRL) {
+ // Another special-case: SRL is basically zero-extending a narrower value.
+ ExtType = ISD::ZEXTLOAD;
+ N0 = SDValue(N, 0);
+ ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+ if (!N01) return SDValue();
+ ExtVT = EVT::getIntegerVT(*DAG.getContext(),
+ VT.getSizeInBits() - N01->getZExtValue());
+ }
+ if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
+ return SDValue();
unsigned EVTBits = ExtVT.getSizeInBits();
+
+ // 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 (!ExtVT.isRound())
+ return SDValue();
+
unsigned ShAmt = 0;
- if (N0.getOpcode() == ISD::SRL && N0.hasOneUse() && ExtVT.isRound()) {
+ if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
ShAmt = N01->getZExtValue();
// Is the shift amount a multiple of size of VT?
if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
return SDValue();
}
+
+ // At this point, we must have a load or else we can't do the transform.
+ if (!isa<LoadSDNode>(N0)) return SDValue();
+
+ // If the shift amount is larger than the input type then we're not
+ // accessing any of the loaded bytes. If the load was a zextload/extload
+ // then the result of the shift+trunc is zero/undef (handled elsewhere).
+ // If the load was a sextload then the result is a splat of the sign bit
+ // of the extended byte. This is not worth optimizing for.
+ if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
+ 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() && 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);
- 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 = ExtVT.getStoreSizeInBits();
- ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
- }
-
- uint64_t PtrOff = ShAmt / 8;
- unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
- SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
- PtrType, LN0->getBasePtr(),
- DAG.getConstant(PtrOff, PtrType));
- AddToWorkList(NewPtr.getNode());
-
- SDValue Load = (ExtType == ISD::NON_EXTLOAD)
- ? DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
- LN0->getSrcValue(), LN0->getSrcValueOffset() + PtrOff,
- LN0->isVolatile(), LN0->isNonTemporal(), NewAlign)
- : DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(), NewPtr,
- LN0->getSrcValue(), LN0->getSrcValueOffset() + PtrOff,
- ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
- NewAlign);
-
- // Replace the old load's chain with the new load's chain.
- WorkListRemover DeadNodes(*this);
- DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1),
- &DeadNodes);
+ // If the load is shifted left (and the result isn't shifted back right),
+ // we can fold the truncate through the shift.
+ unsigned ShLeftAmt = 0;
+ if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
+ ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
+ if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
+ ShLeftAmt = N01->getZExtValue();
+ N0 = N0.getOperand(0);
+ }
+ }
+
+ // If we haven't found a load, we can't narrow it. Don't transform one with
+ // multiple uses, this would require adding a new load.
+ if (!isa<LoadSDNode>(N0) || !N0.hasOneUse() ||
+ // Don't change the width of a volatile load.
+ cast<LoadSDNode>(N0)->isVolatile())
+ return SDValue();
+
+ // Verify that we are actually reducing a load width here.
+ if (cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() < EVTBits)
+ return SDValue();
+
+ LoadSDNode *LN0 = cast<LoadSDNode>(N0);
+ 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 = ExtVT.getStoreSizeInBits();
+ ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
+ }
+
+ uint64_t PtrOff = ShAmt / 8;
+ unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
+ SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
+ PtrType, LN0->getBasePtr(),
+ DAG.getConstant(PtrOff, PtrType));
+ AddToWorkList(NewPtr.getNode());
+
+ SDValue Load;
+ if (ExtType == ISD::NON_EXTLOAD)
+ Load = DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
+ LN0->getPointerInfo().getWithOffset(PtrOff),
+ LN0->isVolatile(), LN0->isNonTemporal(), NewAlign);
+ else
+ Load = DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(),NewPtr,
+ LN0->getPointerInfo().getWithOffset(PtrOff),
+ ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
+ NewAlign);
- // Return the new loaded value.
- return Load;
+ // Replace the old load's chain with the new load's chain.
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1),
+ &DeadNodes);
+
+ // Shift the result left, if we've swallowed a left shift.
+ SDValue Result = Load;
+ if (ShLeftAmt != 0) {
+ EVT ShImmTy = getShiftAmountTy();
+ if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
+ ShImmTy = VT;
+ Result = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT,
+ Result, DAG.getConstant(ShLeftAmt, ShImmTy));
}
- return SDValue();
+ // Return the new loaded value.
+ return Result;
}
SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(), EVT,
+ LN0->getBasePtr(), LN0->getPointerInfo(),
+ EVT,
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
CombineTo(N, ExtLoad);
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(), EVT,
+ LN0->getBasePtr(), LN0->getPointerInfo(),
+ EVT,
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
CombineTo(N, ExtLoad);
}
// See if we can simplify the input to this truncate through knowledge that
- // only the low bits are being used. For example "trunc (or (shl x, 8), y)"
- // -> trunc y
- SDValue Shorter =
- GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
- VT.getSizeInBits()));
- if (Shorter.getNode())
- return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
-
+ // only the low bits are being used.
+ // For example "trunc (or (shl x, 8), y)" // -> trunc y
+ // Currenly we only perform this optimization on scalars because vectors
+ // may have different active low bits.
+ if (!VT.isVector()) {
+ SDValue Shorter =
+ GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
+ VT.getSizeInBits()));
+ if (Shorter.getNode())
+ return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
+ }
// fold (truncate (load x)) -> (smaller load x)
// fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
- if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT))
- return ReduceLoadWidth(N);
+ if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
+ SDValue Reduced = ReduceLoadWidth(N);
+ if (Reduced.getNode())
+ return Reduced;
+ }
+
+ // Simplify the operands using demanded-bits information.
+ if (!VT.isVector() &&
+ SimplifyDemandedBits(SDValue(N, 0)))
+ return SDValue(N, 0);
+
return SDValue();
}
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())
+ if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
+ LD1->getPointerInfo().getAddrSpace() !=
+ LD2->getPointerInfo().getAddrSpace())
return SDValue();
EVT LD1VT = LD1->getValueType(0);
if (NewAlign <= Align &&
(!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
- LD1->getBasePtr(), LD1->getSrcValue(),
- LD1->getSrcValueOffset(), false, false, Align);
+ LD1->getBasePtr(), LD1->getPointerInfo(),
+ false, false, Align);
}
return SDValue();
}
-SDValue DAGCombiner::visitBIT_CONVERT(SDNode *N) {
+SDValue DAGCombiner::visitBITCAST(SDNode *N) {
SDValue N0 = N->getOperand(0);
EVT VT = N->getValueType(0);
assert(!DestEltVT.isVector() &&
"Element type of vector ValueType must not be vector!");
if (isSimple)
- return ConstantFoldBIT_CONVERTofBUILD_VECTOR(N0.getNode(), DestEltVT);
+ return ConstantFoldBITCASTofBUILD_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);
+ SDValue Res = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, N0);
if (Res.getNode() != N) {
if (!LegalOperations ||
TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
}
// (conv (conv x, t1), t2) -> (conv x, t2)
- if (N0.getOpcode() == ISD::BIT_CONVERT)
- return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT,
+ if (N0.getOpcode() == ISD::BITCAST)
+ return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT,
N0.getOperand(0));
// fold (conv (load x)) -> (load (conv*)x)
if (Align <= OrigAlign) {
SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
- LN0->getBasePtr(),
- LN0->getSrcValue(), LN0->getSrcValueOffset(),
+ LN0->getBasePtr(), LN0->getPointerInfo(),
LN0->isVolatile(), LN0->isNonTemporal(),
OrigAlign);
AddToWorkList(N);
CombineTo(N0.getNode(),
- DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(),
+ DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
N0.getValueType(), Load),
Load.getValue(1));
return Load;
// This often reduces constant pool loads.
if ((N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FABS) &&
N0.getNode()->hasOneUse() && VT.isInteger() && !VT.isVector()) {
- SDValue NewConv = DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(), VT,
+ SDValue NewConv = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(), VT,
N0.getOperand(0));
AddToWorkList(NewConv.getNode());
unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
if (isTypeLegal(IntXVT)) {
- SDValue X = DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(),
+ SDValue X = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
IntXVT, N0.getOperand(1));
AddToWorkList(X.getNode());
X, DAG.getConstant(SignBit, VT));
AddToWorkList(X.getNode());
- SDValue Cst = DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(),
+ SDValue Cst = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
VT, N0.getOperand(0));
Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
Cst, DAG.getConstant(~SignBit, VT));
return CombineConsecutiveLoads(N, VT);
}
-/// ConstantFoldBIT_CONVERTofBUILD_VECTOR - We know that BV is a build_vector
+/// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
/// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
/// destination element value type.
SDValue DAGCombiner::
-ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
+ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
// If this is already the right type, we're done.
// 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) {
+ EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
+ BV->getValueType(0).getVectorNumElements());
+
+ // Due to the FP element handling below calling this routine recursively,
+ // we can end up with a scalar-to-vector node here.
+ if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
+ return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
+ DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
+ DstEltVT, BV->getOperand(0)));
+
SmallVector<SDValue, 8> Ops;
for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
SDValue Op = BV->getOperand(i);
// 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(),
+ Ops.push_back(DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
DstEltVT, Op));
AddToWorkList(Ops.back().getNode());
}
- EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
- BV->getValueType(0).getVectorNumElements());
return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
&Ops[0], Ops.size());
}
// same sizes.
assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
- BV = ConstantFoldBIT_CONVERTofBUILD_VECTOR(BV, IntVT).getNode();
+ BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
SrcEltVT = IntVT;
}
if (DstEltVT.isFloatingPoint()) {
assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
- SDNode *Tmp = ConstantFoldBIT_CONVERTofBUILD_VECTOR(BV, TmpVT).getNode();
+ SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
// Next, convert to FP elements of the same size.
- return ConstantFoldBIT_CONVERTofBUILD_VECTOR(Tmp, DstEltVT);
+ return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
}
// Okay, we know the src/dst types are both integers of differing types.
if (Op.getOpcode() == ISD::UNDEF) continue;
EltIsUndef = false;
- NewBits |= APInt(cast<ConstantSDNode>(Op)->getAPIntValue()).
+ NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
zextOrTrunc(SrcBitSize).zext(DstBitSize);
}
continue;
}
- APInt OpVal = APInt(cast<ConstantSDNode>(BV->getOperand(i))->
- getAPIntValue()).zextOrTrunc(SrcBitSize);
+ APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
+ getAPIntValue().zextOrTrunc(SrcBitSize);
for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
- APInt ThisVal = APInt(OpVal).trunc(DstBitSize);
+ APInt ThisVal = OpVal.trunc(DstBitSize);
Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
- if (isS2V && i == 0 && j == 0 && APInt(ThisVal).zext(SrcBitSize) == OpVal)
+ if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
// Simply turn this into a SCALAR_TO_VECTOR of the new type.
return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
Ops[0]);
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(),
+ LN0->getBasePtr(), LN0->getPointerInfo(),
N0.getValueType(),
LN0->isVolatile(), LN0->isNonTemporal(),
LN0->getAlignment());
// Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
// constant pool values.
- if (N0.getOpcode() == ISD::BIT_CONVERT &&
+ if (N0.getOpcode() == ISD::BITCAST &&
!VT.isVector() &&
N0.getNode()->hasOneUse() &&
N0.getOperand(0).getValueType().isInteger()) {
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(),
+ return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
VT, Int);
}
}
// Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
// constant pool values.
- if (N0.getOpcode() == ISD::BIT_CONVERT && N0.getNode()->hasOneUse() &&
+ if (N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
N0.getOperand(0).getValueType().isInteger() &&
!N0.getOperand(0).getValueType().isVector()) {
SDValue Int = N0.getOperand(0);
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(),
+ return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
N->getValueType(0), Int);
}
}
N1.getOperand(0), N1.getOperand(1), N2);
}
- SDNode *Trunc = 0;
- if (N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) {
- // Look past truncate.
- Trunc = N1.getNode();
- N1 = N1.getOperand(0);
- }
+ if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
+ ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
+ (N1.getOperand(0).hasOneUse() &&
+ N1.getOperand(0).getOpcode() == ISD::SRL))) {
+ SDNode *Trunc = 0;
+ if (N1.getOpcode() == ISD::TRUNCATE) {
+ // Look pass the truncate.
+ Trunc = N1.getNode();
+ N1 = N1.getOperand(0);
+ }
- if (N1.hasOneUse() && N1.getOpcode() == ISD::SRL) {
// Match this pattern so that we can generate simpler code:
//
// %a = ...
// brcond i32 %c ...
//
// into
- //
+ //
// %a = ...
// %b = and i32 %a, 2
// %c = setcc eq %b, 0
}
}
}
+
+ if (Trunc)
+ // Restore N1 if the above transformation doesn't match.
+ N1 = N->getOperand(1);
}
-
+
// Transform br(xor(x, y)) -> br(x != y)
// Transform br(xor(xor(x,y), 1)) -> br (x == y)
if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
Equal = true;
}
- SDValue NodeToReplace = Trunc ? SDValue(Trunc, 0) : N1;
-
- EVT SetCCVT = NodeToReplace.getValueType();
+ EVT SetCCVT = N1.getValueType();
if (LegalTypes)
SetCCVT = TLI.getSetCCResultType(SetCCVT);
SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
Equal ? ISD::SETEQ : ISD::SETNE);
// Replace the uses of XOR with SETCC
WorkListRemover DeadNodes(*this);
- DAG.ReplaceAllUsesOfValueWith(NodeToReplace, SetCC, &DeadNodes);
- removeFromWorkList(NodeToReplace.getNode());
- DAG.DeleteNode(NodeToReplace.getNode());
+ DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes);
+ removeFromWorkList(N1.getNode());
+ DAG.DeleteNode(N1.getNode());
return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
MVT::Other, Chain, SetCC, N2);
}
SDValue Offset;
ISD::MemIndexedMode AM = ISD::UNINDEXED;
if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
- if (Ptr == Offset && Op->getOpcode() == ISD::ADD)
- std::swap(BasePtr, Offset);
- if (Ptr != BasePtr)
- continue;
// Don't create a indexed load / store with zero offset.
if (isa<ConstantSDNode>(Offset) &&
cast<ConstantSDNode>(Offset)->isNullValue())
if (Align > LD->getAlignment())
return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
LD->getValueType(0),
- Chain, Ptr, LD->getSrcValue(),
- LD->getSrcValueOffset(), LD->getMemoryVT(),
+ Chain, Ptr, LD->getPointerInfo(),
+ LD->getMemoryVT(),
LD->isVolatile(), LD->isNonTemporal(), Align);
}
}
// Replace the chain to void dependency.
if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
- BetterChain, Ptr,
- LD->getSrcValue(), LD->getSrcValueOffset(),
+ BetterChain, Ptr, LD->getPointerInfo(),
LD->isVolatile(), LD->isNonTemporal(),
LD->getAlignment());
} else {
ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
LD->getValueType(0),
- BetterChain, Ptr, LD->getSrcValue(),
- LD->getSrcValueOffset(),
+ BetterChain, Ptr, LD->getPointerInfo(),
LD->getMemoryVT(),
LD->isVolatile(),
LD->isNonTemporal(),
// Create token factor to keep old chain connected.
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);
if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
return SDValue(N, 0);
- if (PromoteLoad(SDValue(N, 0)))
- return SDValue(N, 0);
return SDValue();
}
static std::pair<unsigned, unsigned>
CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
std::pair<unsigned, unsigned> Result(0, 0);
-
+
// Check for the structure we're looking for.
if (V->getOpcode() != ISD::AND ||
!isa<ConstantSDNode>(V->getOperand(1)) ||
!ISD::isNormalLoad(V->getOperand(0).getNode()))
return Result;
-
+
// Check the chain and pointer.
LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
-
+
// The store should be chained directly to the load or be an operand of a
// tokenfactor.
if (LD == Chain.getNode())
}
if (!isOk) return Result;
}
-
+
// This only handles simple types.
if (V.getValueType() != MVT::i16 &&
V.getValueType() != MVT::i32 &&
unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
if (NotMaskLZ == 64) return Result; // All zero mask.
-
+
// See if we have a continuous run of bits. If so, we have 0*1+0*
if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
return Result;
// Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
if (V.getValueType() != MVT::i64 && NotMaskLZ)
NotMaskLZ -= 64-V.getValueSizeInBits();
-
+
unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
switch (MaskedBytes) {
- case 1:
- case 2:
+ case 1:
+ case 2:
case 4: break;
default: return Result; // All one mask, or 5-byte mask.
}
-
+
// Verify that the first bit starts at a multiple of mask so that the access
// is aligned the same as the access width.
if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
-
+
Result.first = MaskedBytes;
Result.second = NotMaskTZ/8;
return Result;
unsigned NumBytes = MaskInfo.first;
unsigned ByteShift = MaskInfo.second;
SelectionDAG &DAG = DC->getDAG();
-
+
// Check to see if IVal is all zeros in the part being masked in by the 'or'
// that uses this. If not, this is not a replacement.
APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
ByteShift*8, (ByteShift+NumBytes)*8);
if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
-
+
// Check that it is legal on the target to do this. It is legal if the new
// VT we're shrinking to (i8/i16/i32) is legal or we're still before type
// legalization.
MVT VT = MVT::getIntegerVT(NumBytes*8);
if (!DC->isTypeLegal(VT))
return 0;
-
+
// Okay, we can do this! Replace the 'St' store with a store of IVal that is
// shifted by ByteShift and truncated down to NumBytes.
if (ByteShift)
StOffset = ByteShift;
else
StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
-
+
SDValue Ptr = St->getBasePtr();
if (StOffset) {
Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
NewAlign = MinAlign(NewAlign, StOffset);
}
-
+
// Truncate down to the new size.
IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
-
+
++OpsNarrowed;
- return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
- St->getSrcValue(), St->getSrcValueOffset()+StOffset,
+ return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
+ St->getPointerInfo().getWithOffset(StOffset),
false, false, NewAlign).getNode();
}
return SDValue();
unsigned Opc = Value.getOpcode();
-
+
// If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
// is a byte mask indicating a consecutive number of bytes, check to see if
// Y is known to provide just those bytes. If so, we try to replace the
if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
Value.getOperand(1), ST,this))
return SDValue(NewST, 0);
-
+
// Or is commutative, so try swapping X and Y.
MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
if (MaskedLoad.first)
Value.getOperand(0), ST,this))
return SDValue(NewST, 0);
}
-
+
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()) {
+ if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
+ Chain == SDValue(N0.getNode(), 1)) {
LoadSDNode *LD = cast<LoadSDNode>(N0);
- if (LD->getBasePtr() != Ptr)
+ if (LD->getBasePtr() != Ptr ||
+ LD->getPointerInfo().getAddrSpace() !=
+ ST->getPointerInfo().getAddrSpace())
return SDValue();
// Find the type to narrow it the load / op / store to.
DAG.getConstant(PtrOff, Ptr.getValueType()));
SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
LD->getChain(), NewPtr,
- LD->getSrcValue(), LD->getSrcValueOffset(),
+ LD->getPointerInfo().getWithOffset(PtrOff),
LD->isVolatile(), LD->isNonTemporal(),
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(),
+ ST->getPointerInfo().getWithOffset(PtrOff),
false, false, NewAlign);
AddToWorkList(NewPtr.getNode());
return SDValue();
}
+/// TransformFPLoadStorePair - For a given floating point load / store pair,
+/// if the load value isn't used by any other operations, then consider
+/// transforming the pair to integer load / store operations if the target
+/// deems the transformation profitable.
+SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
+ StoreSDNode *ST = cast<StoreSDNode>(N);
+ SDValue Chain = ST->getChain();
+ SDValue Value = ST->getValue();
+ if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
+ Value.hasOneUse() &&
+ Chain == SDValue(Value.getNode(), 1)) {
+ LoadSDNode *LD = cast<LoadSDNode>(Value);
+ EVT VT = LD->getMemoryVT();
+ if (!VT.isFloatingPoint() ||
+ VT != ST->getMemoryVT() ||
+ LD->isNonTemporal() ||
+ ST->isNonTemporal() ||
+ LD->getPointerInfo().getAddrSpace() != 0 ||
+ ST->getPointerInfo().getAddrSpace() != 0)
+ return SDValue();
+
+ EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
+ if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
+ !TLI.isOperationLegal(ISD::STORE, IntVT) ||
+ !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
+ !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
+ return SDValue();
+
+ unsigned LDAlign = LD->getAlignment();
+ unsigned STAlign = ST->getAlignment();
+ const Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
+ unsigned ABIAlign = TLI.getTargetData()->getABITypeAlignment(IntVTTy);
+ if (LDAlign < ABIAlign || STAlign < ABIAlign)
+ return SDValue();
+
+ SDValue NewLD = DAG.getLoad(IntVT, Value.getDebugLoc(),
+ LD->getChain(), LD->getBasePtr(),
+ LD->getPointerInfo(),
+ false, false, LDAlign);
+
+ SDValue NewST = DAG.getStore(NewLD.getValue(1), N->getDebugLoc(),
+ NewLD, ST->getBasePtr(),
+ ST->getPointerInfo(),
+ false, false, STAlign);
+
+ AddToWorkList(NewLD.getNode());
+ AddToWorkList(NewST.getNode());
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1),
+ &DeadNodes);
+ ++LdStFP2Int;
+ return NewST;
+ }
+
+ return SDValue();
+}
+
SDValue DAGCombiner::visitSTORE(SDNode *N) {
StoreSDNode *ST = cast<StoreSDNode>(N);
SDValue Chain = ST->getChain();
// If this is a store of a bit convert, store the input value if the
// resultant store does not need a higher alignment than the original.
- if (Value.getOpcode() == ISD::BIT_CONVERT && !ST->isTruncatingStore() &&
+ if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
ST->isUnindexed()) {
unsigned OrigAlign = ST->getAlignment();
EVT SVT = Value.getOperand(0).getValueType();
((!LegalOperations && !ST->isVolatile()) ||
TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
- Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(), ST->isVolatile(),
+ Ptr, ST->getPointerInfo(), ST->isVolatile(),
ST->isNonTemporal(), OrigAlign);
}
Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
bitcastToAPInt().getZExtValue(), MVT::i32);
return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
- Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(), ST->isVolatile(),
+ Ptr, ST->getPointerInfo(), ST->isVolatile(),
ST->isNonTemporal(), ST->getAlignment());
}
break;
Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
getZExtValue(), MVT::i64);
return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
- Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(), ST->isVolatile(),
+ Ptr, ST->getPointerInfo(), ST->isVolatile(),
ST->isNonTemporal(), ST->getAlignment());
} else if (!ST->isVolatile() &&
TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
if (TLI.isBigEndian()) std::swap(Lo, Hi);
- int SVOffset = ST->getSrcValueOffset();
unsigned Alignment = ST->getAlignment();
bool isVolatile = ST->isVolatile();
bool isNonTemporal = ST->isNonTemporal();
SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
- Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(),
+ Ptr, ST->getPointerInfo(),
isVolatile, isNonTemporal,
ST->getAlignment());
Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
DAG.getConstant(4, Ptr.getValueType()));
- SVOffset += 4;
Alignment = MinAlign(Alignment, 4U);
SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
- Ptr, ST->getSrcValue(),
- SVOffset, isVolatile, isNonTemporal,
+ Ptr, ST->getPointerInfo().getWithOffset(4),
+ isVolatile, isNonTemporal,
Alignment);
return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
St0, St1);
if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
if (Align > ST->getAlignment())
return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
- Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(), ST->getMemoryVT(),
+ Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
ST->isVolatile(), ST->isNonTemporal(), Align);
}
}
+ // Try transforming a pair floating point load / store ops to integer
+ // load / store ops.
+ SDValue NewST = TransformFPLoadStorePair(N);
+ if (NewST.getNode())
+ return NewST;
+
if (CombinerAA) {
// Walk up chain skipping non-aliasing memory nodes.
SDValue BetterChain = FindBetterChain(N, Chain);
// Replace the chain to avoid dependency.
if (ST->isTruncatingStore()) {
ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
- ST->getSrcValue(),ST->getSrcValueOffset(),
+ ST->getPointerInfo(),
ST->getMemoryVT(), ST->isVolatile(),
ST->isNonTemporal(), ST->getAlignment());
} else {
ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
- ST->getSrcValue(), ST->getSrcValueOffset(),
+ ST->getPointerInfo(),
ST->isVolatile(), ST->isNonTemporal(),
ST->getAlignment());
}
AddToWorkList(Value.getNode());
if (Shorter.getNode())
return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
- Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(), ST->getMemoryVT(),
+ Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
ST->isVolatile(), ST->isNonTemporal(),
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,
- APInt::getLowBitsSet(
- Value.getValueType().getScalarType().getSizeInBits(),
- ST->getMemoryVT().getScalarType().getSizeInBits())))
+ APInt::getLowBitsSet(
+ Value.getValueType().getScalarType().getSizeInBits(),
+ ST->getMemoryVT().getScalarType().getSizeInBits())))
return SDValue(N, 0);
}
TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
ST->getMemoryVT())) {
return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
- Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(), ST->getMemoryVT(),
+ Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
ST->isVolatile(), ST->isNonTemporal(),
ST->getAlignment());
}
SDValue InVal = N->getOperand(1);
SDValue EltNo = N->getOperand(2);
+ // If the inserted element is an UNDEF, just use the input vector.
+ if (InVal.getOpcode() == ISD::UNDEF)
+ return InVec;
+
+ EVT VT = InVec.getValueType();
+
+ // If we can't generate a legal BUILD_VECTOR, exit
+ if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
+ return SDValue();
+
// 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)) {
if (Elt < Ops.size())
Ops[Elt] = InVal;
return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
- InVec.getValueType(), &Ops[0], Ops.size());
+ VT, &Ops[0], Ops.size());
}
- // If the invec is an UNDEF and if EltNo is a constant, create a new
+ // 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 &&
+ if (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));
if (Elt < Ops.size())
Ops[Elt] = InVal;
return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
- InVec.getValueType(), &Ops[0], Ops.size());
+ VT, &Ops[0], Ops.size());
}
return SDValue();
}
// Check if the result type doesn't match the inserted element type. A
// SCALAR_TO_VECTOR may truncate the inserted element and the
// EXTRACT_VECTOR_ELT may widen the extracted vector.
- EVT EltVT = InVec.getValueType().getVectorElementType();
SDValue InOp = InVec.getOperand(0);
EVT NVT = N->getValueType(0);
if (InOp.getValueType() != NVT) {
SDValue EltNo = N->getOperand(1);
if (isa<ConstantSDNode>(EltNo)) {
- unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
+ int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
bool NewLoad = false;
bool BCNumEltsChanged = false;
EVT VT = InVec.getValueType();
EVT ExtVT = VT.getVectorElementType();
EVT LVT = ExtVT;
- if (InVec.getOpcode() == ISD::BIT_CONVERT) {
+ if (InVec.getOpcode() == ISD::BITCAST) {
EVT BCVT = InVec.getOperand(0).getValueType();
if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
return SDValue();
// Select the input vector, guarding against out of range extract vector.
unsigned NumElems = VT.getVectorNumElements();
- int Idx = (Elt > NumElems) ? -1 : SVN->getMaskElt(Elt);
+ int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
- if (InVec.getOpcode() == ISD::BIT_CONVERT)
+ if (InVec.getOpcode() == ISD::BITCAST)
InVec = InVec.getOperand(0);
if (ISD::isNormalLoad(InVec.getNode())) {
LN0 = cast<LoadSDNode>(InVec);
if (!LN0 || !LN0->hasOneUse() || LN0->isVolatile())
return SDValue();
+ // If Idx was -1 above, Elt is going to be -1, so just return undef.
+ if (Elt == -1)
+ return DAG.getUNDEF(LN0->getBasePtr().getValueType());
+
unsigned Align = LN0->getAlignment();
if (NewLoad) {
// Check the resultant load doesn't need a higher alignment than the
// original load.
unsigned NewAlign =
- TLI.getTargetData()->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
+ TLI.getTargetData()
+ ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
return SDValue();
}
SDValue NewPtr = LN0->getBasePtr();
+ unsigned PtrOff = 0;
+
if (Elt) {
- unsigned PtrOff = LVT.getSizeInBits() * Elt / 8;
+ PtrOff = LVT.getSizeInBits() * Elt / 8;
EVT PtrType = NewPtr.getValueType();
if (TLI.isBigEndian())
PtrOff = VT.getSizeInBits() / 8 - PtrOff;
}
return DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
- LN0->getSrcValue(), LN0->getSrcValueOffset(),
+ LN0->getPointerInfo().getWithOffset(PtrOff),
LN0->isVolatile(), LN0->isNonTemporal(), Align);
}
unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
if (ExtIndex > VT.getVectorNumElements())
return SDValue();
-
+
Mask.push_back(ExtIndex);
continue;
}
}
SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
- return SDValue();
-
EVT VT = N->getValueType(0);
unsigned NumElts = VT.getVectorNumElements();
// 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 (cast<ShuffleVectorSDNode>(N)->isSplat()) {
+ // If it is a splat, check if the argument vector is another splat or a
+ // build_vector with all scalar elements the same.
+ ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
+ if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
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
// look though conversions that change things like v4f32 to v2f64.
- if (V->getOpcode() == ISD::BIT_CONVERT) {
+ if (V->getOpcode() == ISD::BITCAST) {
SDValue ConvInput = V->getOperand(0);
if (ConvInput.getValueType().isVector() &&
ConvInput.getValueType().getVectorNumElements() == NumElts)
}
if (V->getOpcode() == ISD::BUILD_VECTOR) {
- unsigned NumElems = V->getNumOperands();
- unsigned BaseIdx = cast<ShuffleVectorSDNode>(N)->getSplatIndex();
- if (NumElems > BaseIdx) {
- SDValue Base;
- bool AllSame = true;
- for (unsigned i = 0; i != NumElems; ++i) {
- if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
- Base = V->getOperand(i);
- break;
- }
+ assert(V->getNumOperands() == NumElts &&
+ "BUILD_VECTOR has wrong number of operands");
+ SDValue Base;
+ bool AllSame = true;
+ for (unsigned i = 0; i != NumElts; ++i) {
+ if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
+ Base = V->getOperand(i);
+ break;
}
- // Splat of <u, u, u, u>, return <u, u, u, u>
- if (!Base.getNode())
- return N0;
- for (unsigned i = 0; i != NumElems; ++i) {
- if (V->getOperand(i) != Base) {
- AllSame = false;
- break;
- }
+ }
+ // Splat of <u, u, u, u>, return <u, u, u, u>
+ if (!Base.getNode())
+ return N0;
+ for (unsigned i = 0; i != NumElts; ++i) {
+ if (V->getOperand(i) != Base) {
+ AllSame = false;
+ break;
}
- // Splat of <x, x, x, x>, return <x, x, x, x>
- if (AllSame)
- return N0;
}
+ // Splat of <x, x, x, x>, return <x, x, x, x>
+ if (AllSame)
+ return N0;
}
}
return SDValue();
}
+SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
+ if (!TLI.getShouldFoldAtomicFences())
+ return SDValue();
+
+ SDValue atomic = N->getOperand(0);
+ switch (atomic.getOpcode()) {
+ case ISD::ATOMIC_CMP_SWAP:
+ case ISD::ATOMIC_SWAP:
+ case ISD::ATOMIC_LOAD_ADD:
+ case ISD::ATOMIC_LOAD_SUB:
+ case ISD::ATOMIC_LOAD_AND:
+ case ISD::ATOMIC_LOAD_OR:
+ case ISD::ATOMIC_LOAD_XOR:
+ case ISD::ATOMIC_LOAD_NAND:
+ case ISD::ATOMIC_LOAD_MIN:
+ case ISD::ATOMIC_LOAD_MAX:
+ case ISD::ATOMIC_LOAD_UMIN:
+ case ISD::ATOMIC_LOAD_UMAX:
+ break;
+ default:
+ return SDValue();
+ }
+
+ SDValue fence = atomic.getOperand(0);
+ if (fence.getOpcode() != ISD::MEMBARRIER)
+ return SDValue();
+
+ switch (atomic.getOpcode()) {
+ case ISD::ATOMIC_CMP_SWAP:
+ return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
+ fence.getOperand(0),
+ atomic.getOperand(1), atomic.getOperand(2),
+ atomic.getOperand(3)), atomic.getResNo());
+ case ISD::ATOMIC_SWAP:
+ case ISD::ATOMIC_LOAD_ADD:
+ case ISD::ATOMIC_LOAD_SUB:
+ case ISD::ATOMIC_LOAD_AND:
+ case ISD::ATOMIC_LOAD_OR:
+ case ISD::ATOMIC_LOAD_XOR:
+ case ISD::ATOMIC_LOAD_NAND:
+ case ISD::ATOMIC_LOAD_MIN:
+ case ISD::ATOMIC_LOAD_MAX:
+ case ISD::ATOMIC_LOAD_UMIN:
+ case ISD::ATOMIC_LOAD_UMAX:
+ return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
+ fence.getOperand(0),
+ atomic.getOperand(1), atomic.getOperand(2)),
+ atomic.getResNo());
+ default:
+ return SDValue();
+ }
+}
+
/// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
/// an AND to a vector_shuffle with the destination vector and a zero vector.
/// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
if (N->getOpcode() == ISD::AND) {
- if (RHS.getOpcode() == ISD::BIT_CONVERT)
+ if (RHS.getOpcode() == ISD::BITCAST)
RHS = RHS.getOperand(0);
if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
SmallVector<int, 8> Indices;
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);
+ LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuf);
+ return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
}
}
// things. Simplifying them may result in a loss of legality.
if (LegalOperations) return SDValue();
- EVT VT = N->getValueType(0);
- assert(VT.isVector() && "SimplifyVBinOp only works on vectors!");
+ assert(N->getValueType(0).isVector() &&
+ "SimplifyVBinOp only works on vectors!");
- EVT EltType = VT.getVectorElementType();
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
SDValue Shuffle = XformToShuffleWithZero(N);
break;
}
- Ops.push_back(DAG.getNode(N->getOpcode(), LHS.getDebugLoc(),
- EltType, LHSOp, RHSOp));
- AddToWorkList(Ops.back().getNode());
- assert((Ops.back().getOpcode() == ISD::UNDEF ||
- Ops.back().getOpcode() == ISD::Constant ||
- Ops.back().getOpcode() == ISD::ConstantFP) &&
- "Scalar binop didn't fold!");
+ EVT VT = LHSOp.getValueType();
+ assert(RHSOp.getValueType() == VT &&
+ "SimplifyVBinOp with different BUILD_VECTOR element types");
+ SDValue FoldOp = DAG.getNode(N->getOpcode(), LHS.getDebugLoc(), VT,
+ LHSOp, RHSOp);
+ if (FoldOp.getOpcode() != ISD::UNDEF &&
+ FoldOp.getOpcode() != ISD::Constant &&
+ FoldOp.getOpcode() != ISD::ConstantFP)
+ break;
+ Ops.push_back(FoldOp);
+ AddToWorkList(FoldOp.getNode());
}
- if (Ops.size() == LHS.getNumOperands()) {
- EVT VT = LHS.getValueType();
- return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
- &Ops[0], Ops.size());
- }
+ if (Ops.size() == LHS.getNumOperands())
+ return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
+ LHS.getValueType(), &Ops[0], Ops.size());
}
return SDValue();
bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
SDValue RHS) {
+ // Cannot simplify select with vector condition
+ if (TheSelect->getOperand(0).getValueType().isVector()) return false;
+
// 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()){
- // If this is a load and the token chain is identical, replace the select
- // of two loads with a load through a select of the address to load from.
- // This triggers in things like "select bool X, 10.0, 123.0" after the FP
- // constants have been dropped into the constant pool.
- if (LHS.getOpcode() == ISD::LOAD &&
+ if (LHS.getOpcode() != RHS.getOpcode() ||
+ !LHS.hasOneUse() || !RHS.hasOneUse())
+ return false;
+
+ // If this is a load and the token chain is identical, replace the select
+ // of two loads with a load through a select of the address to load from.
+ // This triggers in things like "select bool X, 10.0, 123.0" after the FP
+ // constants have been dropped into the constant pool.
+ if (LHS.getOpcode() == ISD::LOAD) {
+ LoadSDNode *LLD = cast<LoadSDNode>(LHS);
+ LoadSDNode *RLD = cast<LoadSDNode>(RHS);
+
+ // Token chains must be identical.
+ if (LHS.getOperand(0) != RHS.getOperand(0) ||
// Do not let this transformation reduce the number of volatile loads.
- !cast<LoadSDNode>(LHS)->isVolatile() &&
- !cast<LoadSDNode>(RHS)->isVolatile() &&
- // Token chains must be identical.
- LHS.getOperand(0) == RHS.getOperand(0)) {
- LoadSDNode *LLD = cast<LoadSDNode>(LHS);
- LoadSDNode *RLD = cast<LoadSDNode>(RHS);
-
- // If this is an EXTLOAD, the VT's must match.
- if (LLD->getMemoryVT() == RLD->getMemoryVT()) {
+ LLD->isVolatile() || RLD->isVolatile() ||
+ // If this is an EXTLOAD, the VT's must match.
+ LLD->getMemoryVT() != RLD->getMemoryVT() ||
+ // If this is an EXTLOAD, the kind of extension must match.
+ (LLD->getExtensionType() != RLD->getExtensionType() &&
+ // The only exception is if one of the extensions is anyext.
+ LLD->getExtensionType() != ISD::EXTLOAD &&
+ RLD->getExtensionType() != ISD::EXTLOAD) ||
// FIXME: this discards src value information. This is
// over-conservative. It would be beneficial to be able to remember
// both potential memory locations. Since we are discarding
// src value info, don't do the transformation if the memory
// locations are not in the default address space.
- unsigned LLDAddrSpace = 0, RLDAddrSpace = 0;
- if (const Value *LLDVal = LLD->getMemOperand()->getValue()) {
- if (const PointerType *PT = dyn_cast<PointerType>(LLDVal->getType()))
- LLDAddrSpace = PT->getAddressSpace();
- }
- if (const Value *RLDVal = RLD->getMemOperand()->getValue()) {
- if (const PointerType *PT = dyn_cast<PointerType>(RLDVal->getType()))
- RLDAddrSpace = PT->getAddressSpace();
- }
- SDValue Addr;
- if (LLDAddrSpace == 0 && RLDAddrSpace == 0) {
- if (TheSelect->getOpcode() == ISD::SELECT) {
- // Check that the condition doesn't reach either load. If so, folding
- // this will induce a cycle into the DAG.
- if ((!LLD->hasAnyUseOfValue(1) ||
- !LLD->isPredecessorOf(TheSelect->getOperand(0).getNode())) &&
- (!RLD->hasAnyUseOfValue(1) ||
- !RLD->isPredecessorOf(TheSelect->getOperand(0).getNode()))) {
- Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
- LLD->getBasePtr().getValueType(),
- TheSelect->getOperand(0), LLD->getBasePtr(),
- RLD->getBasePtr());
- }
- } else {
- // Check that the condition doesn't reach either load. If so, folding
- // this will induce a cycle into the DAG.
- if ((!LLD->hasAnyUseOfValue(1) ||
- (!LLD->isPredecessorOf(TheSelect->getOperand(0).getNode()) &&
- !LLD->isPredecessorOf(TheSelect->getOperand(1).getNode()))) &&
- (!RLD->hasAnyUseOfValue(1) ||
- (!RLD->isPredecessorOf(TheSelect->getOperand(0).getNode()) &&
- !RLD->isPredecessorOf(TheSelect->getOperand(1).getNode())))) {
- Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
- LLD->getBasePtr().getValueType(),
- TheSelect->getOperand(0),
- 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, 0, 0,
- LLD->isVolatile(),
- LLD->isNonTemporal(),
- LLD->getAlignment());
- } else {
- Load = DAG.getExtLoad(LLD->getExtensionType(),
- TheSelect->getDebugLoc(),
- TheSelect->getValueType(0),
- LLD->getChain(), Addr, 0, 0,
- LLD->getMemoryVT(),
- LLD->isVolatile(),
- LLD->isNonTemporal(),
- LLD->getAlignment());
- }
+ LLD->getPointerInfo().getAddrSpace() != 0 ||
+ RLD->getPointerInfo().getAddrSpace() != 0)
+ return false;
- // Users of the select now use the result of the load.
- CombineTo(TheSelect, Load);
+ // Check that the select condition doesn't reach either load. If so,
+ // folding this will induce a cycle into the DAG. If not, this is safe to
+ // xform, so create a select of the addresses.
+ SDValue Addr;
+ if (TheSelect->getOpcode() == ISD::SELECT) {
+ SDNode *CondNode = TheSelect->getOperand(0).getNode();
+ if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
+ (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
+ return false;
+ Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
+ LLD->getBasePtr().getValueType(),
+ TheSelect->getOperand(0), LLD->getBasePtr(),
+ RLD->getBasePtr());
+ } else { // Otherwise SELECT_CC
+ SDNode *CondLHS = TheSelect->getOperand(0).getNode();
+ SDNode *CondRHS = TheSelect->getOperand(1).getNode();
+
+ if ((LLD->hasAnyUseOfValue(1) &&
+ (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
+ (LLD->hasAnyUseOfValue(1) &&
+ (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))))
+ return false;
- // 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));
- CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
- return true;
- }
- }
- }
+ Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
+ LLD->getBasePtr().getValueType(),
+ TheSelect->getOperand(0),
+ TheSelect->getOperand(1),
+ LLD->getBasePtr(), RLD->getBasePtr(),
+ TheSelect->getOperand(4));
+ }
+
+ SDValue Load;
+ if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
+ Load = DAG.getLoad(TheSelect->getValueType(0),
+ TheSelect->getDebugLoc(),
+ // FIXME: Discards pointer info.
+ LLD->getChain(), Addr, MachinePointerInfo(),
+ LLD->isVolatile(), LLD->isNonTemporal(),
+ LLD->getAlignment());
+ } else {
+ Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
+ RLD->getExtensionType() : LLD->getExtensionType(),
+ TheSelect->getDebugLoc(),
+ TheSelect->getValueType(0),
+ // FIXME: Discards pointer info.
+ LLD->getChain(), Addr, MachinePointerInfo(),
+ LLD->getMemoryVT(), LLD->isVolatile(),
+ LLD->isNonTemporal(), 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));
+ CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
+ return true;
}
return false;
ISD::CondCode CC, bool NotExtCompare) {
// (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());
return DAG.getNode(ISD::FABS, DL, VT, N3);
}
}
-
+
// 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
};
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(),
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);
CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
CstOffset);
return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
- PseudoSourceValue::getConstantPool(), 0, false,
+ MachinePointerInfo::getConstantPool(), false,
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)
}
}
+ // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
+ // where y is has a single bit set.
+ // A plaintext description would be, we can turn the SELECT_CC into an AND
+ // when the condition can be materialized as an all-ones register. Any
+ // single bit-test can be materialized as an all-ones register with
+ // shift-left and shift-right-arith.
+ if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
+ N0->getValueType(0) == VT &&
+ N1C && N1C->isNullValue() &&
+ N2C && N2C->isNullValue()) {
+ SDValue AndLHS = N0->getOperand(0);
+ ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
+ if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
+ // Shift the tested bit over the sign bit.
+ APInt AndMask = ConstAndRHS->getAPIntValue();
+ SDValue ShlAmt =
+ DAG.getConstant(AndMask.countLeadingZeros(), getShiftAmountTy());
+ SDValue Shl = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT, AndLHS, ShlAmt);
+
+ // Now arithmetic right shift it all the way over, so the result is either
+ // all-ones, or zero.
+ SDValue ShrAmt =
+ DAG.getConstant(AndMask.getBitWidth()-1, getShiftAmountTy());
+ SDValue Shr = DAG.getNode(ISD::SRA, N0.getDebugLoc(), VT, Shl, ShrAmt);
+
+ return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
+ }
+ }
+
// fold select C, 16, 0 -> shl C, 4
if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent) {
}
}
- // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
+ // Check to see if this is an integer abs.
+ // select_cc setg[te] X, 0, X, -X ->
+ // select_cc setgt X, -1, X, -X ->
+ // select_cc setl[te] X, 0, -X, X ->
+ // select_cc setlt X, 1, -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()) {
+ if (N1C) {
+ ConstantSDNode *SubC = NULL;
+ if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
+ (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
+ N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
+ SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
+ else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
+ (N1C->isOne() && CC == ISD::SETLT)) &&
+ N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
+ SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
+
EVT XType = N0.getValueType();
- SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType, N0,
- DAG.getConstant(XType.getSizeInBits()-1,
- getShiftAmountTy()));
- SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(), XType,
- N0, Shift);
- AddToWorkList(Shift.getNode());
- AddToWorkList(Add.getNode());
- return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
- }
- // Check to see if this is an integer abs. select_cc setgt X, -1, X, -X ->
- // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
- 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))) {
- EVT XType = N0.getValueType();
- if (SubC->isNullValue() && XType.isInteger()) {
- SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
- N0,
- DAG.getConstant(XType.getSizeInBits()-1,
- getShiftAmountTy()));
- SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
- XType, N0, Shift);
- AddToWorkList(Shift.getNode());
- AddToWorkList(Add.getNode());
- return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
- }
+ if (SubC && SubC->isNullValue() && XType.isInteger()) {
+ SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
+ N0,
+ DAG.getConstant(XType.getSizeInBits()-1,
+ getShiftAmountTy()));
+ SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
+ XType, N0, Shift);
+ AddToWorkList(Shift.getNode());
+ AddToWorkList(Add.getNode());
+ return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
}
}
}
/// 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.
+// to alias with anything but itself. Provides base object and offset as
+// results.
static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
const GlobalValue *&GV, void *&CV) {
// Assume it is a primitive operation.
Offset += C->getZExtValue();
}
}
-
+
// 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.
bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
const Value *SrcValue1, int SrcValueOffset1,
unsigned SrcValueAlign1,
+ const MDNode *TBAAInfo1,
SDValue Ptr2, int64_t Size2,
const Value *SrcValue2, int SrcValueOffset2,
- unsigned SrcValueAlign2) const {
+ unsigned SrcValueAlign2,
+ const MDNode *TBAAInfo2) const {
// If they are the same then they must be aliases.
if (Ptr1 == Ptr2) return true;
if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
- // If we know what the bases are, and they aren't identical, then we know they
- // cannot alias.
+ // It is possible for different frame indices to alias each other, mostly
+ // when tail call optimization reuses return address slots for arguments.
+ // To catch this case, look up the actual index of frame indices to compute
+ // the real alias relationship.
+ if (isFrameIndex1 && isFrameIndex2) {
+ MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
+ Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
+ Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
+ return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
+ }
+
+ // Otherwise, 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;
(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 =
- AA.alias(SrcValue1, Overlap1, SrcValue2, Overlap2);
+ AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
+ AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
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,
+ const Value *&SrcValue,
int &SrcValueOffset,
- unsigned &SrcValueAlign) const {
+ unsigned &SrcValueAlign,
+ const MDNode *&TBAAInfo) 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();
+ TBAAInfo = LD->getTBAAInfo();
return true;
} else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
Ptr = ST->getBasePtr();
SrcValue = ST->getSrcValue();
SrcValueOffset = ST->getSrcValueOffset();
SrcValueAlign = ST->getOriginalAlignment();
+ TBAAInfo = ST->getTBAAInfo();
} else {
llvm_unreachable("FindAliasInfo expected a memory operand");
}
const Value *SrcValue;
int SrcValueOffset;
unsigned SrcValueAlign;
- bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
- SrcValueAlign);
+ const MDNode *SrcTBAAInfo;
+ bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
+ SrcValueAlign, SrcTBAAInfo);
// 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.
while (!Chains.empty()) {
SDValue Chain = Chains.back();
Chains.pop_back();
-
- // 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
+
+ // 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
+ //
+ // 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) {
const Value *OpSrcValue;
int OpSrcValueOffset;
unsigned OpSrcValueAlign;
+ const MDNode *OpSrcTBAAInfo;
bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
OpSrcValue, OpSrcValueOffset,
- OpSrcValueAlign);
+ OpSrcValueAlign,
+ OpSrcTBAAInfo);
// If chain is alias then stop here.
if (!(IsLoad && IsOpLoad) &&
isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
+ SrcTBAAInfo,
OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
- OpSrcValueAlign)) {
+ OpSrcValueAlign, OpSrcTBAAInfo)) {
Aliases.push_back(Chain);
} else {
// Look further up the chain.
// If a single operand then chain to it. We don't need to revisit it.
return Aliases[0];
}
-
+
// Construct a custom tailored token factor.
- return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
+ return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
&Aliases[0], Aliases.size());
}