diff options
Diffstat (limited to 'contrib/llvm/lib/CodeGen/SelectionDAG')
16 files changed, 1314 insertions, 688 deletions
diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp index e671752..c9c4d91 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -4489,6 +4489,16 @@ ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) { // 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::BIT_CONVERT, 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); @@ -4500,8 +4510,6 @@ ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) { 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()); } @@ -5790,7 +5798,8 @@ SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) { 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) return SDValue(); diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp index decaa76..a4eed71 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/FastISel.cpp @@ -94,7 +94,7 @@ bool FastISel::hasTrivialKill(const Value *V) const { !(I->getOpcode() == Instruction::BitCast || I->getOpcode() == Instruction::PtrToInt || I->getOpcode() == Instruction::IntToPtr) && - cast<Instruction>(I->use_begin())->getParent() == I->getParent(); + cast<Instruction>(*I->use_begin())->getParent() == I->getParent(); } unsigned FastISel::getRegForValue(const Value *V) { @@ -146,7 +146,7 @@ unsigned FastISel::getRegForValue(const Value *V) { return Reg; } -/// materializeRegForValue - Helper for getRegForVale. This function is +/// materializeRegForValue - Helper for getRegForValue. This function is /// called when the value isn't already available in a register and must /// be materialized with new instructions. unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) { @@ -276,6 +276,7 @@ std::pair<unsigned, bool> FastISel::getRegForGEPIndex(const Value *Idx) { void FastISel::recomputeInsertPt() { if (getLastLocalValue()) { FuncInfo.InsertPt = getLastLocalValue(); + FuncInfo.MBB = FuncInfo.InsertPt->getParent(); ++FuncInfo.InsertPt; } else FuncInfo.InsertPt = FuncInfo.MBB->getFirstNonPHI(); @@ -472,17 +473,7 @@ bool FastISel::SelectCall(const User *I) { return true; const AllocaInst *AI = dyn_cast<AllocaInst>(Address); // Don't handle byval struct arguments or VLAs, for example. - // Note that if we have a byval struct argument, fast ISel is turned off; - // those are handled in SelectionDAGBuilder. - if (AI) { - DenseMap<const AllocaInst*, int>::iterator SI = - FuncInfo.StaticAllocaMap.find(AI); - if (SI == FuncInfo.StaticAllocaMap.end()) break; // VLAs. - int FI = SI->second; - if (!DI->getDebugLoc().isUnknown()) - FuncInfo.MF->getMMI().setVariableDbgInfo(DI->getVariable(), - FI, DI->getDebugLoc()); - } else + if (!AI) // Building the map above is target independent. Generating DBG_VALUE // inline is target dependent; do this now. (void)TargetSelectInstruction(cast<Instruction>(I)); diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/FunctionLoweringInfo.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/FunctionLoweringInfo.cpp index 928e1ec..5ef6404 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/FunctionLoweringInfo.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/FunctionLoweringInfo.cpp @@ -20,6 +20,7 @@ #include "llvm/IntrinsicInst.h" #include "llvm/LLVMContext.h" #include "llvm/Module.h" +#include "llvm/Analysis/DebugInfo.h" #include "llvm/CodeGen/Analysis.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFrameInfo.h" @@ -111,17 +112,56 @@ void FunctionLoweringInfo::set(const Function &fn, MachineFunction &mf) { TySize *= CUI->getZExtValue(); // Get total allocated size. if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects. + + // The object may need to be placed onto the stack near the stack + // protector if one exists. Determine here if this object is a suitable + // candidate. I.e., it would trigger the creation of a stack protector. + bool MayNeedSP = + (AI->isArrayAllocation() || + (TySize > 8 && isa<ArrayType>(Ty) && + cast<ArrayType>(Ty)->getElementType()->isIntegerTy(8))); StaticAllocaMap[AI] = - MF->getFrameInfo()->CreateStackObject(TySize, Align, false); + MF->getFrameInfo()->CreateStackObject(TySize, Align, false, MayNeedSP); } for (; BB != EB; ++BB) - for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) + for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) { + // Mark values used outside their block as exported, by allocating + // a virtual register for them. if (isUsedOutsideOfDefiningBlock(I)) if (!isa<AllocaInst>(I) || !StaticAllocaMap.count(cast<AllocaInst>(I))) InitializeRegForValue(I); + // Collect llvm.dbg.declare information. This is done now instead of + // during the initial isel pass through the IR so that it is done + // in a predictable order. + if (const DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(I)) { + MachineModuleInfo &MMI = MF->getMMI(); + if (MMI.hasDebugInfo() && + DIVariable(DI->getVariable()).Verify() && + !DI->getDebugLoc().isUnknown()) { + // Don't handle byval struct arguments or VLAs, for example. + // Non-byval arguments are handled here (they refer to the stack + // temporary alloca at this point). + const Value *Address = DI->getAddress(); + if (Address) { + if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) + Address = BCI->getOperand(0); + if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) { + DenseMap<const AllocaInst *, int>::iterator SI = + StaticAllocaMap.find(AI); + if (SI != StaticAllocaMap.end()) { // Check for VLAs. + int FI = SI->second; + MMI.setVariableDbgInfo(DI->getVariable(), + FI, DI->getDebugLoc()); + } + } + } + } + } + } + // Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This // also creates the initial PHI MachineInstrs, though none of the input // operands are populated. @@ -181,6 +221,7 @@ void FunctionLoweringInfo::clear() { #endif LiveOutRegInfo.clear(); ArgDbgValues.clear(); + ByValArgFrameIndexMap.clear(); RegFixups.clear(); } @@ -214,6 +255,28 @@ unsigned FunctionLoweringInfo::CreateRegs(const Type *Ty) { return FirstReg; } +/// setByValArgumentFrameIndex - Record frame index for the byval +/// argument. This overrides previous frame index entry for this argument, +/// if any. +void FunctionLoweringInfo::setByValArgumentFrameIndex(const Argument *A, + int FI) { + assert (A->hasByValAttr() && "Argument does not have byval attribute!"); + ByValArgFrameIndexMap[A] = FI; +} + +/// getByValArgumentFrameIndex - Get frame index for the byval argument. +/// If the argument does not have any assigned frame index then 0 is +/// returned. +int FunctionLoweringInfo::getByValArgumentFrameIndex(const Argument *A) { + assert (A->hasByValAttr() && "Argument does not have byval attribute!"); + DenseMap<const Argument *, int>::iterator I = + ByValArgFrameIndexMap.find(A); + if (I != ByValArgFrameIndexMap.end()) + return I->second; + DEBUG(dbgs() << "Argument does not have assigned frame index!"); + return 0; +} + /// AddCatchInfo - Extract the personality and type infos from an eh.selector /// call, and add them to the specified machine basic block. void llvm::AddCatchInfo(const CallInst &I, MachineModuleInfo *MMI, diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp index 7a47da4..2981cd3 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp @@ -100,8 +100,7 @@ public: /// it is already legal or we need to expand it into multiple registers of /// smaller integer type, or we need to promote it to a larger type. LegalizeAction getTypeAction(EVT VT) const { - return - (LegalizeAction)ValueTypeActions.getTypeAction(*DAG.getContext(), VT); + return (LegalizeAction)ValueTypeActions.getTypeAction(VT); } /// isTypeLegal - Return true if this type is legal on this target. @@ -1314,21 +1313,30 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { } break; case TargetLowering::Expand: - // f64 = EXTLOAD f32 should expand to LOAD, FP_EXTEND - // f128 = EXTLOAD {f32,f64} too - if ((SrcVT == MVT::f32 && (Node->getValueType(0) == MVT::f64 || - Node->getValueType(0) == MVT::f128)) || - (SrcVT == MVT::f64 && Node->getValueType(0) == MVT::f128)) { + if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && isTypeLegal(SrcVT)) { SDValue Load = DAG.getLoad(SrcVT, dl, Tmp1, Tmp2, LD->getSrcValue(), LD->getSrcValueOffset(), LD->isVolatile(), LD->isNonTemporal(), LD->getAlignment()); - Result = DAG.getNode(ISD::FP_EXTEND, dl, - Node->getValueType(0), Load); + unsigned ExtendOp; + switch (ExtType) { + case ISD::EXTLOAD: + ExtendOp = (SrcVT.isFloatingPoint() ? + ISD::FP_EXTEND : ISD::ANY_EXTEND); + break; + case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break; + case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break; + default: llvm_unreachable("Unexpected extend load type!"); + } + Result = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load); Tmp1 = LegalizeOp(Result); // Relegalize new nodes. Tmp2 = LegalizeOp(Load.getValue(1)); break; } + // FIXME: This does not work for vectors on most targets. Sign- and + // zero-extend operations are currently folded into extending loads, + // whether they are legal or not, and then we end up here without any + // support for legalizing them. assert(ExtType != ISD::EXTLOAD && "EXTLOAD should always be supported!"); // Turn the unsupported load into an EXTLOAD followed by an explicit diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp index b94ea9a..f8c5890 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp @@ -234,8 +234,9 @@ SDValue DAGTypeLegalizer::PromoteIntRes_BUILD_PAIR(SDNode *N) { // The pair element type may be legal, or may not promote to the same type as // the result, for example i14 = BUILD_PAIR (i7, i7). Handle all cases. return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), - TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)), - JoinIntegers(N->getOperand(0), N->getOperand(1))); + TLI.getTypeToTransformTo(*DAG.getContext(), + N->getValueType(0)), JoinIntegers(N->getOperand(0), + N->getOperand(1))); } SDValue DAGTypeLegalizer::PromoteIntRes_Constant(SDNode *N) { @@ -245,7 +246,8 @@ SDValue DAGTypeLegalizer::PromoteIntRes_Constant(SDNode *N) { // Zero extend things like i1, sign extend everything else. It shouldn't // matter in theory which one we pick, but this tends to give better code? unsigned Opc = VT.isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; - SDValue Result = DAG.getNode(Opc, dl, TLI.getTypeToTransformTo(*DAG.getContext(), VT), + SDValue Result = DAG.getNode(Opc, dl, + TLI.getTypeToTransformTo(*DAG.getContext(), VT), SDValue(N, 0)); assert(isa<ConstantSDNode>(Result) && "Didn't constant fold ext?"); return Result; @@ -310,8 +312,8 @@ SDValue DAGTypeLegalizer::PromoteIntRes_FP_TO_XINT(SDNode *N) { // If we're promoting a UINT to a larger size and the larger FP_TO_UINT is // not Legal, check to see if we can use FP_TO_SINT instead. (If both UINT - // and SINT conversions are Custom, there is no way to tell which is preferable. - // We choose SINT because that's the right thing on PPC.) + // and SINT conversions are Custom, there is no way to tell which is + // preferable. We choose SINT because that's the right thing on PPC.) if (N->getOpcode() == ISD::FP_TO_UINT && !TLI.isOperationLegal(ISD::FP_TO_UINT, NVT) && TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NVT)) @@ -1030,7 +1032,7 @@ void DAGTypeLegalizer::ExpandShiftByConstant(SDNode *N, unsigned Amt, Hi = InL; } else if (Amt == 1 && TLI.isOperationLegalOrCustom(ISD::ADDC, - TLI.getTypeToExpandTo(*DAG.getContext(), NVT))) { + TLI.getTypeToExpandTo(*DAG.getContext(), NVT))) { // Emit this X << 1 as X+X. SDVTList VTList = DAG.getVTList(NVT, MVT::Flag); SDValue LoOps[2] = { InL, InL }; @@ -1926,7 +1928,8 @@ ExpandIntRes_SIGN_EXTEND_INREG(SDNode *N, SDValue &Lo, SDValue &Hi) { unsigned ExcessBits = EVT.getSizeInBits() - Lo.getValueType().getSizeInBits(); Hi = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, Hi.getValueType(), Hi, - DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), ExcessBits))); + DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), + ExcessBits))); } } @@ -2046,7 +2049,8 @@ void DAGTypeLegalizer::ExpandIntRes_ZERO_EXTEND(SDNode *N, unsigned ExcessBits = Op.getValueType().getSizeInBits() - NVT.getSizeInBits(); Hi = DAG.getZeroExtendInReg(Hi, dl, - EVT::getIntegerVT(*DAG.getContext(), ExcessBits)); + EVT::getIntegerVT(*DAG.getContext(), + ExcessBits)); } } diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h b/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h index bd86694..d560292 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h @@ -75,7 +75,7 @@ private: /// getTypeAction - Return how we should legalize values of this type. LegalizeAction getTypeAction(EVT VT) const { - switch (ValueTypeActions.getTypeAction(*DAG.getContext(), VT)) { + switch (ValueTypeActions.getTypeAction(VT)) { default: assert(false && "Unknown legalize action!"); case TargetLowering::Legal: @@ -86,8 +86,7 @@ private: // 2) For vectors, use a wider vector type (e.g. v3i32 -> v4i32). if (!VT.isVector()) return PromoteInteger; - else - return WidenVector; + return WidenVector; case TargetLowering::Expand: // Expand can mean // 1) split scalar in half, 2) convert a float to an integer, @@ -95,23 +94,21 @@ private: if (!VT.isVector()) { if (VT.isInteger()) return ExpandInteger; - else if (VT.getSizeInBits() == - TLI.getTypeToTransformTo(*DAG.getContext(), VT).getSizeInBits()) + if (VT.getSizeInBits() == + TLI.getTypeToTransformTo(*DAG.getContext(), VT).getSizeInBits()) return SoftenFloat; - else - return ExpandFloat; - } else if (VT.getVectorNumElements() == 1) { - return ScalarizeVector; - } else { - return SplitVector; + return ExpandFloat; } + + if (VT.getVectorNumElements() == 1) + return ScalarizeVector; + return SplitVector; } } /// isTypeLegal - Return true if this type is legal on this target. bool isTypeLegal(EVT VT) const { - return (ValueTypeActions.getTypeAction(*DAG.getContext(), VT) == - TargetLowering::Legal); + return ValueTypeActions.getTypeAction(VT) == TargetLowering::Legal; } /// IgnoreNodeResults - Pretend all of this node's results are legal. @@ -584,6 +581,7 @@ private: SDValue SplitVecOp_EXTRACT_SUBVECTOR(SDNode *N); SDValue SplitVecOp_EXTRACT_VECTOR_ELT(SDNode *N); SDValue SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo); + SDValue SplitVecOp_CONCAT_VECTORS(SDNode *N); //===--------------------------------------------------------------------===// // Vector Widening Support: LegalizeVectorTypes.cpp diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp index 93aeff5..93bc2d0 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp @@ -983,6 +983,7 @@ bool DAGTypeLegalizer::SplitVectorOperand(SDNode *N, unsigned OpNo) { case ISD::BIT_CONVERT: Res = SplitVecOp_BIT_CONVERT(N); break; case ISD::EXTRACT_SUBVECTOR: Res = SplitVecOp_EXTRACT_SUBVECTOR(N); break; case ISD::EXTRACT_VECTOR_ELT:Res = SplitVecOp_EXTRACT_VECTOR_ELT(N); break; + case ISD::CONCAT_VECTORS: Res = SplitVecOp_CONCAT_VECTORS(N); break; case ISD::STORE: Res = SplitVecOp_STORE(cast<StoreSDNode>(N), OpNo); break; @@ -1091,8 +1092,7 @@ SDValue DAGTypeLegalizer::SplitVecOp_EXTRACT_VECTOR_ELT(SDNode *N) { return SDValue(DAG.UpdateNodeOperands(N, Lo, Idx), 0); return SDValue(DAG.UpdateNodeOperands(N, Hi, DAG.getConstant(IdxVal - LoElts, - Idx.getValueType())), - 0); + Idx.getValueType())), 0); } // Store the vector to the stack. @@ -1113,7 +1113,7 @@ SDValue DAGTypeLegalizer::SplitVecOp_EXTRACT_VECTOR_ELT(SDNode *N) { SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) { assert(N->isUnindexed() && "Indexed store of vector?"); assert(OpNo == 1 && "Can only split the stored value"); - DebugLoc dl = N->getDebugLoc(); + DebugLoc DL = N->getDebugLoc(); bool isTruncating = N->isTruncatingStore(); SDValue Ch = N->getChain(); @@ -1132,25 +1132,49 @@ SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) { unsigned IncrementSize = LoMemVT.getSizeInBits()/8; if (isTruncating) - Lo = DAG.getTruncStore(Ch, dl, Lo, Ptr, N->getSrcValue(), SVOffset, + Lo = DAG.getTruncStore(Ch, DL, Lo, Ptr, N->getSrcValue(), SVOffset, LoMemVT, isVol, isNT, Alignment); else - Lo = DAG.getStore(Ch, dl, Lo, Ptr, N->getSrcValue(), SVOffset, + Lo = DAG.getStore(Ch, DL, Lo, Ptr, N->getSrcValue(), SVOffset, isVol, isNT, Alignment); // Increment the pointer to the other half. - Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, + Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr, DAG.getIntPtrConstant(IncrementSize)); SVOffset += IncrementSize; if (isTruncating) - Hi = DAG.getTruncStore(Ch, dl, Hi, Ptr, N->getSrcValue(), SVOffset, + Hi = DAG.getTruncStore(Ch, DL, Hi, Ptr, N->getSrcValue(), SVOffset, HiMemVT, isVol, isNT, Alignment); else - Hi = DAG.getStore(Ch, dl, Hi, Ptr, N->getSrcValue(), SVOffset, + Hi = DAG.getStore(Ch, DL, Hi, Ptr, N->getSrcValue(), SVOffset, isVol, isNT, Alignment); - return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi); + return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi); +} + +SDValue DAGTypeLegalizer::SplitVecOp_CONCAT_VECTORS(SDNode *N) { + DebugLoc DL = N->getDebugLoc(); + + // The input operands all must have the same type, and we know the result the + // result type is valid. Convert this to a buildvector which extracts all the + // input elements. + // TODO: If the input elements are power-two vectors, we could convert this to + // a new CONCAT_VECTORS node with elements that are half-wide. + SmallVector<SDValue, 32> Elts; + EVT EltVT = N->getValueType(0).getVectorElementType(); + for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) { + SDValue Op = N->getOperand(op); + for (unsigned i = 0, e = Op.getValueType().getVectorNumElements(); + i != e; ++i) { + Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, + Op, DAG.getIntPtrConstant(i))); + + } + } + + return DAG.getNode(ISD::BUILD_VECTOR, DL, N->getValueType(0), + &Elts[0], Elts.size()); } @@ -1274,8 +1298,8 @@ SDValue DAGTypeLegalizer::WidenVecRes_Binary(SDNode *N) { EVT VT = WidenVT; unsigned NumElts = VT.getVectorNumElements(); while (!TLI.isTypeSynthesizable(VT) && NumElts != 1) { - NumElts = NumElts / 2; - VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts); + NumElts = NumElts / 2; + VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts); } if (NumElts != 1 && !TLI.canOpTrap(N->getOpcode(), VT)) { @@ -1283,124 +1307,123 @@ SDValue DAGTypeLegalizer::WidenVecRes_Binary(SDNode *N) { SDValue InOp1 = GetWidenedVector(N->getOperand(0)); SDValue InOp2 = GetWidenedVector(N->getOperand(1)); return DAG.getNode(N->getOpcode(), dl, WidenVT, InOp1, InOp2); - } else if (NumElts == 1) { - // No legal vector version so unroll the vector operation and then widen. + } + + // No legal vector version so unroll the vector operation and then widen. + if (NumElts == 1) return DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()); - } else { - // Since the operation can trap, apply operation on the original vector. - EVT MaxVT = VT; - SDValue InOp1 = GetWidenedVector(N->getOperand(0)); - SDValue InOp2 = GetWidenedVector(N->getOperand(1)); - unsigned CurNumElts = N->getValueType(0).getVectorNumElements(); - - SmallVector<SDValue, 16> ConcatOps(CurNumElts); - unsigned ConcatEnd = 0; // Current ConcatOps index. - int Idx = 0; // Current Idx into input vectors. - - // NumElts := greatest synthesizable vector size (at most WidenVT) - // while (orig. vector has unhandled elements) { - // take munches of size NumElts from the beginning and add to ConcatOps - // NumElts := next smaller supported vector size or 1 - // } - while (CurNumElts != 0) { - while (CurNumElts >= NumElts) { - SDValue EOp1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp1, - DAG.getIntPtrConstant(Idx)); - SDValue EOp2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp2, - DAG.getIntPtrConstant(Idx)); - ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, VT, EOp1, EOp2); - Idx += NumElts; - CurNumElts -= NumElts; - } - do { - NumElts = NumElts / 2; - VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts); - } while (!TLI.isTypeSynthesizable(VT) && NumElts != 1); - - if (NumElts == 1) { - for (unsigned i = 0; i != CurNumElts; ++i, ++Idx) { - SDValue EOp1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT, - InOp1, DAG.getIntPtrConstant(Idx)); - SDValue EOp2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT, - InOp2, DAG.getIntPtrConstant(Idx)); - ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, WidenEltVT, - EOp1, EOp2); - } - CurNumElts = 0; + + // Since the operation can trap, apply operation on the original vector. + EVT MaxVT = VT; + SDValue InOp1 = GetWidenedVector(N->getOperand(0)); + SDValue InOp2 = GetWidenedVector(N->getOperand(1)); + unsigned CurNumElts = N->getValueType(0).getVectorNumElements(); + + SmallVector<SDValue, 16> ConcatOps(CurNumElts); + unsigned ConcatEnd = 0; // Current ConcatOps index. + int Idx = 0; // Current Idx into input vectors. + + // NumElts := greatest synthesizable vector size (at most WidenVT) + // while (orig. vector has unhandled elements) { + // take munches of size NumElts from the beginning and add to ConcatOps + // NumElts := next smaller supported vector size or 1 + // } + while (CurNumElts != 0) { + while (CurNumElts >= NumElts) { + SDValue EOp1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp1, + DAG.getIntPtrConstant(Idx)); + SDValue EOp2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp2, + DAG.getIntPtrConstant(Idx)); + ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, VT, EOp1, EOp2); + Idx += NumElts; + CurNumElts -= NumElts; + } + do { + NumElts = NumElts / 2; + VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts); + } while (!TLI.isTypeSynthesizable(VT) && NumElts != 1); + + if (NumElts == 1) { + for (unsigned i = 0; i != CurNumElts; ++i, ++Idx) { + SDValue EOp1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT, + InOp1, DAG.getIntPtrConstant(Idx)); + SDValue EOp2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT, + InOp2, DAG.getIntPtrConstant(Idx)); + ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, WidenEltVT, + EOp1, EOp2); } + CurNumElts = 0; } + } - // Check to see if we have a single operation with the widen type. - if (ConcatEnd == 1) { - VT = ConcatOps[0].getValueType(); - if (VT == WidenVT) - return ConcatOps[0]; - } + // Check to see if we have a single operation with the widen type. + if (ConcatEnd == 1) { + VT = ConcatOps[0].getValueType(); + if (VT == WidenVT) + return ConcatOps[0]; + } - // while (Some element of ConcatOps is not of type MaxVT) { - // From the end of ConcatOps, collect elements of the same type and put - // them into an op of the next larger supported type - // } - while (ConcatOps[ConcatEnd-1].getValueType() != MaxVT) { - Idx = ConcatEnd - 1; - VT = ConcatOps[Idx--].getValueType(); - while (Idx >= 0 && ConcatOps[Idx].getValueType() == VT) - Idx--; - - int NextSize = VT.isVector() ? VT.getVectorNumElements() : 1; - EVT NextVT; - do { - NextSize *= 2; - NextVT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NextSize); - } while (!TLI.isTypeSynthesizable(NextVT)); - - if (!VT.isVector()) { - // Scalar type, create an INSERT_VECTOR_ELEMENT of type NextVT - SDValue VecOp = DAG.getUNDEF(NextVT); - unsigned NumToInsert = ConcatEnd - Idx - 1; - for (unsigned i = 0, OpIdx = Idx+1; i < NumToInsert; i++, OpIdx++) { - VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NextVT, VecOp, - ConcatOps[OpIdx], DAG.getIntPtrConstant(i)); - } - ConcatOps[Idx+1] = VecOp; - ConcatEnd = Idx + 2; - } - else { - // Vector type, create a CONCAT_VECTORS of type NextVT - SDValue undefVec = DAG.getUNDEF(VT); - unsigned OpsToConcat = NextSize/VT.getVectorNumElements(); - SmallVector<SDValue, 16> SubConcatOps(OpsToConcat); - unsigned RealVals = ConcatEnd - Idx - 1; - unsigned SubConcatEnd = 0; - unsigned SubConcatIdx = Idx + 1; - while (SubConcatEnd < RealVals) - SubConcatOps[SubConcatEnd++] = ConcatOps[++Idx]; - while (SubConcatEnd < OpsToConcat) - SubConcatOps[SubConcatEnd++] = undefVec; - ConcatOps[SubConcatIdx] = DAG.getNode(ISD::CONCAT_VECTORS, dl, - NextVT, &SubConcatOps[0], - OpsToConcat); - ConcatEnd = SubConcatIdx + 1; + // while (Some element of ConcatOps is not of type MaxVT) { + // From the end of ConcatOps, collect elements of the same type and put + // them into an op of the next larger supported type + // } + while (ConcatOps[ConcatEnd-1].getValueType() != MaxVT) { + Idx = ConcatEnd - 1; + VT = ConcatOps[Idx--].getValueType(); + while (Idx >= 0 && ConcatOps[Idx].getValueType() == VT) + Idx--; + + int NextSize = VT.isVector() ? VT.getVectorNumElements() : 1; + EVT NextVT; + do { + NextSize *= 2; + NextVT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NextSize); + } while (!TLI.isTypeSynthesizable(NextVT)); + + if (!VT.isVector()) { + // Scalar type, create an INSERT_VECTOR_ELEMENT of type NextVT + SDValue VecOp = DAG.getUNDEF(NextVT); + unsigned NumToInsert = ConcatEnd - Idx - 1; + for (unsigned i = 0, OpIdx = Idx+1; i < NumToInsert; i++, OpIdx++) { + VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NextVT, VecOp, + ConcatOps[OpIdx], DAG.getIntPtrConstant(i)); } + ConcatOps[Idx+1] = VecOp; + ConcatEnd = Idx + 2; + } else { + // Vector type, create a CONCAT_VECTORS of type NextVT + SDValue undefVec = DAG.getUNDEF(VT); + unsigned OpsToConcat = NextSize/VT.getVectorNumElements(); + SmallVector<SDValue, 16> SubConcatOps(OpsToConcat); + unsigned RealVals = ConcatEnd - Idx - 1; + unsigned SubConcatEnd = 0; + unsigned SubConcatIdx = Idx + 1; + while (SubConcatEnd < RealVals) + SubConcatOps[SubConcatEnd++] = ConcatOps[++Idx]; + while (SubConcatEnd < OpsToConcat) + SubConcatOps[SubConcatEnd++] = undefVec; + ConcatOps[SubConcatIdx] = DAG.getNode(ISD::CONCAT_VECTORS, dl, + NextVT, &SubConcatOps[0], + OpsToConcat); + ConcatEnd = SubConcatIdx + 1; } + } - // Check to see if we have a single operation with the widen type. - if (ConcatEnd == 1) { - VT = ConcatOps[0].getValueType(); - if (VT == WidenVT) - return ConcatOps[0]; - } - - // add undefs of size MaxVT until ConcatOps grows to length of WidenVT - unsigned NumOps = - WidenVT.getVectorNumElements()/MaxVT.getVectorNumElements(); - if (NumOps != ConcatEnd ) { - SDValue UndefVal = DAG.getUNDEF(MaxVT); - for (unsigned j = ConcatEnd; j < NumOps; ++j) - ConcatOps[j] = UndefVal; - } - return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &ConcatOps[0], NumOps); + // Check to see if we have a single operation with the widen type. + if (ConcatEnd == 1) { + VT = ConcatOps[0].getValueType(); + if (VT == WidenVT) + return ConcatOps[0]; } + + // add undefs of size MaxVT until ConcatOps grows to length of WidenVT + unsigned NumOps = WidenVT.getVectorNumElements()/MaxVT.getVectorNumElements(); + if (NumOps != ConcatEnd ) { + SDValue UndefVal = DAG.getUNDEF(MaxVT); + for (unsigned j = ConcatEnd; j < NumOps; ++j) + ConcatOps[j] = UndefVal; + } + return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &ConcatOps[0], NumOps); } SDValue DAGTypeLegalizer::WidenVecRes_Convert(SDNode *N) { @@ -1561,8 +1584,8 @@ SDValue DAGTypeLegalizer::WidenVecRes_BIT_CONVERT(SDNode *N) { unsigned NewNumElts = WidenSize / InSize; if (InVT.isVector()) { EVT InEltVT = InVT.getVectorElementType(); - NewInVT= EVT::getVectorVT(*DAG.getContext(), InEltVT, - WidenSize / InEltVT.getSizeInBits()); + NewInVT = EVT::getVectorVT(*DAG.getContext(), InEltVT, + WidenSize / InEltVT.getSizeInBits()); } else { NewInVT = EVT::getVectorVT(*DAG.getContext(), InVT, NewNumElts); } @@ -1686,8 +1709,7 @@ SDValue DAGTypeLegalizer::WidenVecRes_CONVERT_RNDSAT(SDNode *N) { SDValue RndOp = N->getOperand(3); SDValue SatOp = N->getOperand(4); - EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), - N->getValueType(0)); + EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); unsigned WidenNumElts = WidenVT.getVectorNumElements(); EVT InVT = InOp.getValueType(); @@ -1720,9 +1742,9 @@ SDValue DAGTypeLegalizer::WidenVecRes_CONVERT_RNDSAT(SDNode *N) { SmallVector<SDValue, 16> Ops(NumConcat); Ops[0] = InOp; SDValue UndefVal = DAG.getUNDEF(InVT); - for (unsigned i = 1; i != NumConcat; ++i) { + for (unsigned i = 1; i != NumConcat; ++i) Ops[i] = UndefVal; - } + InOp = DAG.getNode(ISD::CONCAT_VECTORS, dl, InWidenVT, &Ops[0],NumConcat); return DAG.getConvertRndSat(WidenVT, dl, InOp, DTyOp, STyOp, RndOp, SatOp, CvtCode); @@ -2225,25 +2247,24 @@ SDValue DAGTypeLegalizer::GenWidenVectorLoads(SmallVector<SDValue, 16>& LdChain, // Check if we can load the element with one instruction if (LdWidth <= NewVTWidth) { - if (NewVT.isVector()) { - if (NewVT != WidenVT) { - assert(WidenWidth % NewVTWidth == 0); - unsigned NumConcat = WidenWidth / NewVTWidth; - SmallVector<SDValue, 16> ConcatOps(NumConcat); - SDValue UndefVal = DAG.getUNDEF(NewVT); - ConcatOps[0] = LdOp; - for (unsigned i = 1; i != NumConcat; ++i) - ConcatOps[i] = UndefVal; - return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &ConcatOps[0], - NumConcat); - } else - return LdOp; - } else { + if (!NewVT.isVector()) { unsigned NumElts = WidenWidth / NewVTWidth; EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewVT, NumElts); SDValue VecOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NewVecVT, LdOp); return DAG.getNode(ISD::BIT_CONVERT, dl, WidenVT, VecOp); } + if (NewVT == WidenVT) + return LdOp; + + assert(WidenWidth % NewVTWidth == 0); + unsigned NumConcat = WidenWidth / NewVTWidth; + SmallVector<SDValue, 16> ConcatOps(NumConcat); + SDValue UndefVal = DAG.getUNDEF(NewVT); + ConcatOps[0] = LdOp; + for (unsigned i = 1; i != NumConcat; ++i) + ConcatOps[i] = UndefVal; + return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &ConcatOps[0], + NumConcat); } // Load vector by using multiple loads from largest vector to scalar @@ -2276,52 +2297,55 @@ SDValue DAGTypeLegalizer::GenWidenVectorLoads(SmallVector<SDValue, 16>& LdChain, // Build the vector from the loads operations unsigned End = LdOps.size(); - if (LdOps[0].getValueType().isVector()) { - // If the load contains vectors, build the vector using concat vector. - // All of the vectors used to loads are power of 2 and the scalars load - // can be combined to make a power of 2 vector. - SmallVector<SDValue, 16> ConcatOps(End); - int i = End - 1; - int Idx = End; - EVT LdTy = LdOps[i].getValueType(); - // First combine the scalar loads to a vector - if (!LdTy.isVector()) { - for (--i; i >= 0; --i) { - LdTy = LdOps[i].getValueType(); - if (LdTy.isVector()) - break; - } - ConcatOps[--Idx] = BuildVectorFromScalar(DAG, LdTy, LdOps, i+1, End); - } - ConcatOps[--Idx] = LdOps[i]; + if (!LdOps[0].getValueType().isVector()) + // All the loads are scalar loads. + return BuildVectorFromScalar(DAG, WidenVT, LdOps, 0, End); + + // If the load contains vectors, build the vector using concat vector. + // All of the vectors used to loads are power of 2 and the scalars load + // can be combined to make a power of 2 vector. + SmallVector<SDValue, 16> ConcatOps(End); + int i = End - 1; + int Idx = End; + EVT LdTy = LdOps[i].getValueType(); + // First combine the scalar loads to a vector + if (!LdTy.isVector()) { for (--i; i >= 0; --i) { - EVT NewLdTy = LdOps[i].getValueType(); - if (NewLdTy != LdTy) { - // Create a larger vector - ConcatOps[End-1] = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewLdTy, - &ConcatOps[Idx], End - Idx); - Idx = End - 1; - LdTy = NewLdTy; - } - ConcatOps[--Idx] = LdOps[i]; + LdTy = LdOps[i].getValueType(); + if (LdTy.isVector()) + break; } + ConcatOps[--Idx] = BuildVectorFromScalar(DAG, LdTy, LdOps, i+1, End); + } + ConcatOps[--Idx] = LdOps[i]; + for (--i; i >= 0; --i) { + EVT NewLdTy = LdOps[i].getValueType(); + if (NewLdTy != LdTy) { + // Create a larger vector + ConcatOps[End-1] = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewLdTy, + &ConcatOps[Idx], End - Idx); + Idx = End - 1; + LdTy = NewLdTy; + } + ConcatOps[--Idx] = LdOps[i]; + } - if (WidenWidth != LdTy.getSizeInBits()*(End - Idx)) { - // We need to fill the rest with undefs to build the vector - unsigned NumOps = WidenWidth / LdTy.getSizeInBits(); - SmallVector<SDValue, 16> WidenOps(NumOps); - SDValue UndefVal = DAG.getUNDEF(LdTy); - unsigned i = 0; - for (; i != End-Idx; ++i) - WidenOps[i] = ConcatOps[Idx+i]; - for (; i != NumOps; ++i) - WidenOps[i] = UndefVal; - return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &WidenOps[0],NumOps); - } else - return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, - &ConcatOps[Idx], End - Idx); - } else // All the loads are scalar loads. - return BuildVectorFromScalar(DAG, WidenVT, LdOps, 0, End); + if (WidenWidth == LdTy.getSizeInBits()*(End - Idx)) + return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, + &ConcatOps[Idx], End - Idx); + + // We need to fill the rest with undefs to build the vector + unsigned NumOps = WidenWidth / LdTy.getSizeInBits(); + SmallVector<SDValue, 16> WidenOps(NumOps); + SDValue UndefVal = DAG.getUNDEF(LdTy); + { + unsigned i = 0; + for (; i != End-Idx; ++i) + WidenOps[i] = ConcatOps[Idx+i]; + for (; i != NumOps; ++i) + WidenOps[i] = UndefVal; + } + return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &WidenOps[0],NumOps); } SDValue diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGFast.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGFast.cpp index 3b86c32..fae2729 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGFast.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGFast.cpp @@ -13,6 +13,7 @@ #define DEBUG_TYPE "pre-RA-sched" #include "ScheduleDAGSDNodes.h" +#include "llvm/InlineAsm.h" #include "llvm/CodeGen/SchedulerRegistry.h" #include "llvm/CodeGen/SelectionDAGISel.h" #include "llvm/Target/TargetRegisterInfo.h" @@ -432,6 +433,30 @@ static EVT getPhysicalRegisterVT(SDNode *N, unsigned Reg, return N->getValueType(NumRes); } +/// CheckForLiveRegDef - Return true and update live register vector if the +/// specified register def of the specified SUnit clobbers any "live" registers. +static bool CheckForLiveRegDef(SUnit *SU, unsigned Reg, + std::vector<SUnit*> &LiveRegDefs, + SmallSet<unsigned, 4> &RegAdded, + SmallVector<unsigned, 4> &LRegs, + const TargetRegisterInfo *TRI) { + bool Added = false; + if (LiveRegDefs[Reg] && LiveRegDefs[Reg] != SU) { + if (RegAdded.insert(Reg)) { + LRegs.push_back(Reg); + Added = true; + } + } + for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) + if (LiveRegDefs[*Alias] && LiveRegDefs[*Alias] != SU) { + if (RegAdded.insert(*Alias)) { + LRegs.push_back(*Alias); + Added = true; + } + } + return Added; +} + /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay /// scheduling of the given node to satisfy live physical register dependencies. /// If the specific node is the last one that's available to schedule, do @@ -446,37 +471,44 @@ bool ScheduleDAGFast::DelayForLiveRegsBottomUp(SUnit *SU, for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); I != E; ++I) { if (I->isAssignedRegDep()) { - unsigned Reg = I->getReg(); - if (LiveRegDefs[Reg] && LiveRegDefs[Reg] != I->getSUnit()) { - if (RegAdded.insert(Reg)) - LRegs.push_back(Reg); - } - for (const unsigned *Alias = TRI->getAliasSet(Reg); - *Alias; ++Alias) - if (LiveRegDefs[*Alias] && LiveRegDefs[*Alias] != I->getSUnit()) { - if (RegAdded.insert(*Alias)) - LRegs.push_back(*Alias); - } + CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs, + RegAdded, LRegs, TRI); } } for (SDNode *Node = SU->getNode(); Node; Node = Node->getFlaggedNode()) { + if (Node->getOpcode() == ISD::INLINEASM) { + // Inline asm can clobber physical defs. + unsigned NumOps = Node->getNumOperands(); + if (Node->getOperand(NumOps-1).getValueType() == MVT::Flag) + --NumOps; // Ignore the flag operand. + + for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) { + unsigned Flags = + cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue(); + unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags); + + ++i; // Skip the ID value. + if (InlineAsm::isRegDefKind(Flags) || + InlineAsm::isRegDefEarlyClobberKind(Flags)) { + // Check for def of register or earlyclobber register. + for (; NumVals; --NumVals, ++i) { + unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg(); + if (TargetRegisterInfo::isPhysicalRegister(Reg)) + CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI); + } + } else + i += NumVals; + } + continue; + } if (!Node->isMachineOpcode()) continue; const TargetInstrDesc &TID = TII->get(Node->getMachineOpcode()); if (!TID.ImplicitDefs) continue; for (const unsigned *Reg = TID.ImplicitDefs; *Reg; ++Reg) { - if (LiveRegDefs[*Reg] && LiveRegDefs[*Reg] != SU) { - if (RegAdded.insert(*Reg)) - LRegs.push_back(*Reg); - } - for (const unsigned *Alias = TRI->getAliasSet(*Reg); - *Alias; ++Alias) - if (LiveRegDefs[*Alias] && LiveRegDefs[*Alias] != SU) { - if (RegAdded.insert(*Alias)) - LRegs.push_back(*Alias); - } + CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI); } } return !LRegs.empty(); diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp index 3ef521c..4c3e4e3 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp @@ -24,6 +24,7 @@ #include "llvm/Target/TargetData.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Target/TargetLowering.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/STLExtras.h" @@ -54,10 +55,16 @@ static RegisterScheduler static RegisterScheduler hybridListDAGScheduler("list-hybrid", - "Bottom-up rr list scheduling which avoid stalls for " - "long latency instructions", + "Bottom-up register pressure aware list scheduling " + "which tries to balance latency and register pressure", createHybridListDAGScheduler); +static RegisterScheduler + ILPListDAGScheduler("list-ilp", + "Bottom-up register pressure aware list scheduling " + "which tries to balance ILP and register pressure", + createILPListDAGScheduler); + namespace { //===----------------------------------------------------------------------===// /// ScheduleDAGRRList - The actual register reduction list scheduler @@ -181,7 +188,9 @@ private: /// Schedule - Schedule the DAG using list scheduling. void ScheduleDAGRRList::Schedule() { - DEBUG(dbgs() << "********** List Scheduling **********\n"); + DEBUG(dbgs() + << "********** List Scheduling BB#" << BB->getNumber() + << " **********\n"); NumLiveRegs = 0; LiveRegDefs.resize(TRI->getNumRegs(), NULL); @@ -273,6 +282,8 @@ void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) { SU->setHeightToAtLeast(CurCycle); Sequence.push_back(SU); + AvailableQueue->ScheduledNode(SU); + ReleasePredecessors(SU, CurCycle); // Release all the implicit physical register defs that are live. @@ -291,7 +302,6 @@ void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) { } SU->isScheduled = true; - AvailableQueue->ScheduledNode(SU); } /// CapturePred - This does the opposite of ReleasePred. Since SU is being @@ -315,8 +325,6 @@ void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) { DEBUG(dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: "); DEBUG(SU->dump(this)); - AvailableQueue->UnscheduledNode(SU); - for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); I != E; ++I) { CapturePred(&*I); @@ -346,6 +354,7 @@ void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) { SU->isScheduled = false; SU->isAvailable = true; AvailableQueue->push(SU); + AvailableQueue->UnscheduledNode(SU); } /// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in @@ -956,7 +965,8 @@ namespace { template<class SF> class RegReductionPriorityQueue; - /// Sorting functions for the Available queue. + /// bu_ls_rr_sort - Priority function for bottom up register pressure + // reduction scheduler. struct bu_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> { RegReductionPriorityQueue<bu_ls_rr_sort> *SPQ; bu_ls_rr_sort(RegReductionPriorityQueue<bu_ls_rr_sort> *spq) : SPQ(spq) {} @@ -965,6 +975,8 @@ namespace { bool operator()(const SUnit* left, const SUnit* right) const; }; + // td_ls_rr_sort - Priority function for top down register pressure reduction + // scheduler. struct td_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> { RegReductionPriorityQueue<td_ls_rr_sort> *SPQ; td_ls_rr_sort(RegReductionPriorityQueue<td_ls_rr_sort> *spq) : SPQ(spq) {} @@ -973,6 +985,7 @@ namespace { bool operator()(const SUnit* left, const SUnit* right) const; }; + // src_ls_rr_sort - Priority function for source order scheduler. struct src_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> { RegReductionPriorityQueue<src_ls_rr_sort> *SPQ; src_ls_rr_sort(RegReductionPriorityQueue<src_ls_rr_sort> *spq) @@ -983,13 +996,26 @@ namespace { bool operator()(const SUnit* left, const SUnit* right) const; }; + // hybrid_ls_rr_sort - Priority function for hybrid scheduler. struct hybrid_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> { RegReductionPriorityQueue<hybrid_ls_rr_sort> *SPQ; hybrid_ls_rr_sort(RegReductionPriorityQueue<hybrid_ls_rr_sort> *spq) : SPQ(spq) {} hybrid_ls_rr_sort(const hybrid_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {} - + + bool operator()(const SUnit* left, const SUnit* right) const; + }; + + // ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism) + // scheduler. + struct ilp_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> { + RegReductionPriorityQueue<ilp_ls_rr_sort> *SPQ; + ilp_ls_rr_sort(RegReductionPriorityQueue<ilp_ls_rr_sort> *spq) + : SPQ(spq) {} + ilp_ls_rr_sort(const ilp_ls_rr_sort &RHS) + : SPQ(RHS.SPQ) {} + bool operator()(const SUnit* left, const SUnit* right) const; }; } // end anonymous namespace @@ -1029,23 +1055,48 @@ namespace { std::vector<SUnit*> Queue; SF Picker; unsigned CurQueueId; + bool TracksRegPressure; protected: // SUnits - The SUnits for the current graph. std::vector<SUnit> *SUnits; - + + MachineFunction &MF; const TargetInstrInfo *TII; const TargetRegisterInfo *TRI; + const TargetLowering *TLI; ScheduleDAGRRList *scheduleDAG; // SethiUllmanNumbers - The SethiUllman number for each node. std::vector<unsigned> SethiUllmanNumbers; + /// RegPressure - Tracking current reg pressure per register class. + /// + std::vector<unsigned> RegPressure; + + /// RegLimit - Tracking the number of allocatable registers per register + /// class. + std::vector<unsigned> RegLimit; + public: - RegReductionPriorityQueue(const TargetInstrInfo *tii, - const TargetRegisterInfo *tri) - : Picker(this), CurQueueId(0), - TII(tii), TRI(tri), scheduleDAG(NULL) {} + RegReductionPriorityQueue(MachineFunction &mf, + bool tracksrp, + const TargetInstrInfo *tii, + const TargetRegisterInfo *tri, + const TargetLowering *tli) + : Picker(this), CurQueueId(0), TracksRegPressure(tracksrp), + MF(mf), TII(tii), TRI(tri), TLI(tli), scheduleDAG(NULL) { + if (TracksRegPressure) { + unsigned NumRC = TRI->getNumRegClasses(); + RegLimit.resize(NumRC); + RegPressure.resize(NumRC); + std::fill(RegLimit.begin(), RegLimit.end(), 0); + std::fill(RegPressure.begin(), RegPressure.end(), 0); + for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(), + E = TRI->regclass_end(); I != E; ++I) + RegLimit[(*I)->getID()] = tli->getRegPressureLimit(*I, MF); + } + } void initNodes(std::vector<SUnit> &sunits) { SUnits = &sunits; @@ -1072,6 +1123,7 @@ namespace { void releaseState() { SUnits = 0; SethiUllmanNumbers.clear(); + std::fill(RegPressure.begin(), RegPressure.end(), 0); } unsigned getNodePriority(const SUnit *SU) const { @@ -1139,10 +1191,244 @@ namespace { SU->NodeQueueId = 0; } + bool HighRegPressure(const SUnit *SU) const { + if (!TLI) + return false; + + for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end(); + I != E; ++I) { + if (I->isCtrl()) + continue; + SUnit *PredSU = I->getSUnit(); + const SDNode *PN = PredSU->getNode(); + if (!PN->isMachineOpcode()) { + if (PN->getOpcode() == ISD::CopyFromReg) { + EVT VT = PN->getValueType(0); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + unsigned Cost = TLI->getRepRegClassCostFor(VT); + if ((RegPressure[RCId] + Cost) >= RegLimit[RCId]) + return true; + } + continue; + } + unsigned POpc = PN->getMachineOpcode(); + if (POpc == TargetOpcode::IMPLICIT_DEF) + continue; + if (POpc == TargetOpcode::EXTRACT_SUBREG) { + EVT VT = PN->getOperand(0).getValueType(); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + unsigned Cost = TLI->getRepRegClassCostFor(VT); + // Check if this increases register pressure of the specific register + // class to the point where it would cause spills. + if ((RegPressure[RCId] + Cost) >= RegLimit[RCId]) + return true; + continue; + } else if (POpc == TargetOpcode::INSERT_SUBREG || + POpc == TargetOpcode::SUBREG_TO_REG) { + EVT VT = PN->getValueType(0); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + unsigned Cost = TLI->getRepRegClassCostFor(VT); + // Check if this increases register pressure of the specific register + // class to the point where it would cause spills. + if ((RegPressure[RCId] + Cost) >= RegLimit[RCId]) + return true; + continue; + } + unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs(); + for (unsigned i = 0; i != NumDefs; ++i) { + EVT VT = PN->getValueType(i); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + if (RegPressure[RCId] >= RegLimit[RCId]) + return true; // Reg pressure already high. + unsigned Cost = TLI->getRepRegClassCostFor(VT); + if (!PN->hasAnyUseOfValue(i)) + continue; + // Check if this increases register pressure of the specific register + // class to the point where it would cause spills. + if ((RegPressure[RCId] + Cost) >= RegLimit[RCId]) + return true; + } + } + + return false; + } + + void ScheduledNode(SUnit *SU) { + if (!TracksRegPressure) + return; + + const SDNode *N = SU->getNode(); + if (!N->isMachineOpcode()) { + if (N->getOpcode() != ISD::CopyToReg) + return; + } else { + unsigned Opc = N->getMachineOpcode(); + if (Opc == TargetOpcode::EXTRACT_SUBREG || + Opc == TargetOpcode::INSERT_SUBREG || + Opc == TargetOpcode::SUBREG_TO_REG || + Opc == TargetOpcode::REG_SEQUENCE || + Opc == TargetOpcode::IMPLICIT_DEF) + return; + } + + for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); + I != E; ++I) { + if (I->isCtrl()) + continue; + SUnit *PredSU = I->getSUnit(); + if (PredSU->NumSuccsLeft != PredSU->NumSuccs) + continue; + const SDNode *PN = PredSU->getNode(); + if (!PN->isMachineOpcode()) { + if (PN->getOpcode() == ISD::CopyFromReg) { + EVT VT = PN->getValueType(0); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + } + continue; + } + unsigned POpc = PN->getMachineOpcode(); + if (POpc == TargetOpcode::IMPLICIT_DEF) + continue; + if (POpc == TargetOpcode::EXTRACT_SUBREG) { + EVT VT = PN->getOperand(0).getValueType(); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + continue; + } else if (POpc == TargetOpcode::INSERT_SUBREG || + POpc == TargetOpcode::SUBREG_TO_REG) { + EVT VT = PN->getValueType(0); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + continue; + } + unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs(); + for (unsigned i = 0; i != NumDefs; ++i) { + EVT VT = PN->getValueType(i); + if (!PN->hasAnyUseOfValue(i)) + continue; + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + } + } + + // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses() + // may transfer data dependencies to CopyToReg. + if (SU->NumSuccs && N->isMachineOpcode()) { + unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs(); + for (unsigned i = 0; i != NumDefs; ++i) { + EVT VT = N->getValueType(i); + if (!N->hasAnyUseOfValue(i)) + continue; + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT)) + // Register pressure tracking is imprecise. This can happen. + RegPressure[RCId] = 0; + else + RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT); + } + } + + dumpRegPressure(); + } + + void UnscheduledNode(SUnit *SU) { + if (!TracksRegPressure) + return; + + const SDNode *N = SU->getNode(); + if (!N->isMachineOpcode()) { + if (N->getOpcode() != ISD::CopyToReg) + return; + } else { + unsigned Opc = N->getMachineOpcode(); + if (Opc == TargetOpcode::EXTRACT_SUBREG || + Opc == TargetOpcode::INSERT_SUBREG || + Opc == TargetOpcode::SUBREG_TO_REG || + Opc == TargetOpcode::REG_SEQUENCE || + Opc == TargetOpcode::IMPLICIT_DEF) + return; + } + + for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); + I != E; ++I) { + if (I->isCtrl()) + continue; + SUnit *PredSU = I->getSUnit(); + if (PredSU->NumSuccsLeft != PredSU->NumSuccs) + continue; + const SDNode *PN = PredSU->getNode(); + if (!PN->isMachineOpcode()) { + if (PN->getOpcode() == ISD::CopyFromReg) { + EVT VT = PN->getValueType(0); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + } + continue; + } + unsigned POpc = PN->getMachineOpcode(); + if (POpc == TargetOpcode::IMPLICIT_DEF) + continue; + if (POpc == TargetOpcode::EXTRACT_SUBREG) { + EVT VT = PN->getOperand(0).getValueType(); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + continue; + } else if (POpc == TargetOpcode::INSERT_SUBREG || + POpc == TargetOpcode::SUBREG_TO_REG) { + EVT VT = PN->getValueType(0); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + continue; + } + unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs(); + for (unsigned i = 0; i != NumDefs; ++i) { + EVT VT = PN->getValueType(i); + if (!PN->hasAnyUseOfValue(i)) + continue; + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT)) + // Register pressure tracking is imprecise. This can happen. + RegPressure[RCId] = 0; + else + RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT); + } + } + + // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses() + // may transfer data dependencies to CopyToReg. + if (SU->NumSuccs && N->isMachineOpcode()) { + unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs(); + for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) { + EVT VT = N->getValueType(i); + if (VT == MVT::Flag || VT == MVT::Other) + continue; + if (!N->hasAnyUseOfValue(i)) + continue; + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + } + } + + dumpRegPressure(); + } + void setScheduleDAG(ScheduleDAGRRList *scheduleDag) { scheduleDAG = scheduleDag; } + void dumpRegPressure() const { + for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(), + E = TRI->regclass_end(); I != E; ++I) { + const TargetRegisterClass *RC = *I; + unsigned Id = RC->getID(); + unsigned RP = RegPressure[Id]; + if (!RP) continue; + DEBUG(dbgs() << RC->getName() << ": " << RP << " / " << RegLimit[Id] + << '\n'); + } + } + protected: bool canClobber(const SUnit *SU, const SUnit *Op); void AddPseudoTwoAddrDeps(); @@ -1161,6 +1447,9 @@ namespace { typedef RegReductionPriorityQueue<hybrid_ls_rr_sort> HybridBURRPriorityQueue; + + typedef RegReductionPriorityQueue<ilp_ls_rr_sort> + ILPBURRPriorityQueue; } /// closestSucc - Returns the scheduled cycle of the successor which is @@ -1260,30 +1549,63 @@ bool src_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const { } bool hybrid_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const{ - bool LStall = left->SchedulingPref == Sched::Latency && - SPQ->getCurCycle() < left->getHeight(); - bool RStall = right->SchedulingPref == Sched::Latency && - SPQ->getCurCycle() < right->getHeight(); - // If scheduling one of the node will cause a pipeline stall, delay it. - // If scheduling either one of the node will cause a pipeline stall, sort them - // according to their height. - // If neither will cause a pipeline stall, try to reduce register pressure. - if (LStall) { - if (!RStall) - return true; - if (left->getHeight() != right->getHeight()) - return left->getHeight() > right->getHeight(); - } else if (RStall) + bool LHigh = SPQ->HighRegPressure(left); + bool RHigh = SPQ->HighRegPressure(right); + // Avoid causing spills. If register pressure is high, schedule for + // register pressure reduction. + if (LHigh && !RHigh) + return true; + else if (!LHigh && RHigh) + return false; + else if (!LHigh && !RHigh) { + // Low register pressure situation, schedule for latency if possible. + bool LStall = left->SchedulingPref == Sched::Latency && + SPQ->getCurCycle() < left->getHeight(); + bool RStall = right->SchedulingPref == Sched::Latency && + SPQ->getCurCycle() < right->getHeight(); + // If scheduling one of the node will cause a pipeline stall, delay it. + // If scheduling either one of the node will cause a pipeline stall, sort + // them according to their height. + // If neither will cause a pipeline stall, try to reduce register pressure. + if (LStall) { + if (!RStall) + return true; + if (left->getHeight() != right->getHeight()) + return left->getHeight() > right->getHeight(); + } else if (RStall) return false; - // If either node is scheduling for latency, sort them by height and latency - // first. - if (left->SchedulingPref == Sched::Latency || - right->SchedulingPref == Sched::Latency) { - if (left->getHeight() != right->getHeight()) - return left->getHeight() > right->getHeight(); - if (left->Latency != right->Latency) - return left->Latency > right->Latency; + // If either node is scheduling for latency, sort them by height and latency + // first. + if (left->SchedulingPref == Sched::Latency || + right->SchedulingPref == Sched::Latency) { + if (left->getHeight() != right->getHeight()) + return left->getHeight() > right->getHeight(); + if (left->Latency != right->Latency) + return left->Latency > right->Latency; + } + } + + return BURRSort(left, right, SPQ); +} + +bool ilp_ls_rr_sort::operator()(const SUnit *left, + const SUnit *right) const { + bool LHigh = SPQ->HighRegPressure(left); + bool RHigh = SPQ->HighRegPressure(right); + // Avoid causing spills. If register pressure is high, schedule for + // register pressure reduction. + if (LHigh && !RHigh) + return true; + else if (!LHigh && RHigh) + return false; + else if (!LHigh && !RHigh) { + // Low register pressure situation, schedule to maximize instruction level + // parallelism. + if (left->NumPreds > right->NumPreds) + return false; + else if (left->NumPreds < right->NumPreds) + return false; } return BURRSort(left, right, SPQ); @@ -1635,8 +1957,8 @@ llvm::createBURRListDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) { const TargetInstrInfo *TII = TM.getInstrInfo(); const TargetRegisterInfo *TRI = TM.getRegisterInfo(); - BURegReductionPriorityQueue *PQ = new BURegReductionPriorityQueue(TII, TRI); - + BURegReductionPriorityQueue *PQ = + new BURegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0); ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, false, PQ); PQ->setScheduleDAG(SD); return SD; @@ -1648,8 +1970,8 @@ llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) { const TargetInstrInfo *TII = TM.getInstrInfo(); const TargetRegisterInfo *TRI = TM.getRegisterInfo(); - TDRegReductionPriorityQueue *PQ = new TDRegReductionPriorityQueue(TII, TRI); - + TDRegReductionPriorityQueue *PQ = + new TDRegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0); ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, false, PQ); PQ->setScheduleDAG(SD); return SD; @@ -1661,8 +1983,8 @@ llvm::createSourceListDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) { const TargetInstrInfo *TII = TM.getInstrInfo(); const TargetRegisterInfo *TRI = TM.getRegisterInfo(); - SrcRegReductionPriorityQueue *PQ = new SrcRegReductionPriorityQueue(TII, TRI); - + SrcRegReductionPriorityQueue *PQ = + new SrcRegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0); ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, false, PQ); PQ->setScheduleDAG(SD); return SD; @@ -1673,9 +1995,24 @@ llvm::createHybridListDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) { const TargetMachine &TM = IS->TM; const TargetInstrInfo *TII = TM.getInstrInfo(); const TargetRegisterInfo *TRI = TM.getRegisterInfo(); + const TargetLowering *TLI = &IS->getTargetLowering(); - HybridBURRPriorityQueue *PQ = new HybridBURRPriorityQueue(TII, TRI); + HybridBURRPriorityQueue *PQ = + new HybridBURRPriorityQueue(*IS->MF, true, TII, TRI, TLI); + ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, true, PQ); + PQ->setScheduleDAG(SD); + return SD; +} +llvm::ScheduleDAGSDNodes * +llvm::createILPListDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) { + const TargetMachine &TM = IS->TM; + const TargetInstrInfo *TII = TM.getInstrInfo(); + const TargetRegisterInfo *TRI = TM.getRegisterInfo(); + const TargetLowering *TLI = &IS->getTargetLowering(); + + ILPBURRPriorityQueue *PQ = + new ILPBURRPriorityQueue(*IS->MF, true, TII, TRI, TLI); ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, true, PQ); PQ->setScheduleDAG(SD); return SD; diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp index 06cf053..f1bf82a 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp @@ -59,8 +59,9 @@ SUnit *ScheduleDAGSDNodes::NewSUnit(SDNode *N) { SUnits.back().OrigNode = &SUnits.back(); SUnit *SU = &SUnits.back(); const TargetLowering &TLI = DAG->getTargetLoweringInfo(); - if (N->isMachineOpcode() && - N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF) + if (!N || + (N->isMachineOpcode() && + N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF)) SU->SchedulingPref = Sched::None; else SU->SchedulingPref = TLI.getSchedulingPreference(N); diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp index e83a034..ad06ebd 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp @@ -2236,7 +2236,7 @@ unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{ bool SelectionDAG::isKnownNeverNaN(SDValue Op) const { // If we're told that NaNs won't happen, assume they won't. - if (FiniteOnlyFPMath()) + if (NoNaNsFPMath) return true; // If the value is a constant, we can obviously see if it is a NaN or not. @@ -2281,35 +2281,6 @@ bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const { } -/// getShuffleScalarElt - Returns the scalar element that will make up the ith -/// element of the result of the vector shuffle. -SDValue SelectionDAG::getShuffleScalarElt(const ShuffleVectorSDNode *N, - unsigned i) { - EVT VT = N->getValueType(0); - if (N->getMaskElt(i) < 0) - return getUNDEF(VT.getVectorElementType()); - unsigned Index = N->getMaskElt(i); - unsigned NumElems = VT.getVectorNumElements(); - SDValue V = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1); - Index %= NumElems; - - if (V.getOpcode() == ISD::BIT_CONVERT) { - V = V.getOperand(0); - EVT VVT = V.getValueType(); - if (!VVT.isVector() || VVT.getVectorNumElements() != (unsigned)NumElems) - return SDValue(); - } - if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) - return (Index == 0) ? V.getOperand(0) - : getUNDEF(VT.getVectorElementType()); - if (V.getOpcode() == ISD::BUILD_VECTOR) - return V.getOperand(Index); - if (const ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(V)) - return getShuffleScalarElt(SVN, Index); - return SDValue(); -} - - /// getNode - Gets or creates the specified node. /// SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT) { @@ -2624,7 +2595,8 @@ SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, // one big BUILD_VECTOR. if (N1.getOpcode() == ISD::BUILD_VECTOR && N2.getOpcode() == ISD::BUILD_VECTOR) { - SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end()); + SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), + N1.getNode()->op_end()); Elts.append(N2.getNode()->op_begin(), N2.getNode()->op_end()); return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size()); } @@ -3021,7 +2993,8 @@ SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, if (N1.getOpcode() == ISD::BUILD_VECTOR && N2.getOpcode() == ISD::BUILD_VECTOR && N3.getOpcode() == ISD::BUILD_VECTOR) { - SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end()); + SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), + N1.getNode()->op_end()); Elts.append(N2.getNode()->op_begin(), N2.getNode()->op_end()); Elts.append(N3.getNode()->op_begin(), N3.getNode()->op_end()); return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size()); @@ -5872,6 +5845,7 @@ std::string ISD::ArgFlagsTy::getArgFlagsString() { void SDNode::dump() const { dump(0); } void SDNode::dump(const SelectionDAG *G) const { print(dbgs(), G); + dbgs() << '\n'; } void SDNode::print_types(raw_ostream &OS, const SelectionDAG *G) const { @@ -5895,7 +5869,7 @@ void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const { for (MachineSDNode::mmo_iterator i = MN->memoperands_begin(), e = MN->memoperands_end(); i != e; ++i) { OS << **i; - if (next(i) != e) + if (llvm::next(i) != e) OS << " "; } OS << ">"; diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp index 458e865..e657445 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp @@ -70,22 +70,29 @@ LimitFPPrecision("limit-float-precision", cl::location(LimitFloatPrecision), cl::init(0)); +static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL, + const SDValue *Parts, unsigned NumParts, + EVT PartVT, EVT ValueVT); + /// getCopyFromParts - Create a value that contains the specified legal parts /// combined into the value they represent. If the parts combine to a type /// larger then ValueVT then AssertOp can be used to specify whether the extra /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT /// (ISD::AssertSext). -static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, +static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc DL, const SDValue *Parts, unsigned NumParts, EVT PartVT, EVT ValueVT, ISD::NodeType AssertOp = ISD::DELETED_NODE) { + if (ValueVT.isVector()) + return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT); + assert(NumParts > 0 && "No parts to assemble!"); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); SDValue Val = Parts[0]; if (NumParts > 1) { // Assemble the value from multiple parts. - if (!ValueVT.isVector() && ValueVT.isInteger()) { + if (ValueVT.isInteger()) { unsigned PartBits = PartVT.getSizeInBits(); unsigned ValueBits = ValueVT.getSizeInBits(); @@ -100,25 +107,25 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2); if (RoundParts > 2) { - Lo = getCopyFromParts(DAG, dl, Parts, RoundParts / 2, + Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2, PartVT, HalfVT); - Hi = getCopyFromParts(DAG, dl, Parts + RoundParts / 2, + Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2, RoundParts / 2, PartVT, HalfVT); } else { - Lo = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[0]); - Hi = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[1]); + Lo = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[0]); + Hi = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[1]); } if (TLI.isBigEndian()) std::swap(Lo, Hi); - Val = DAG.getNode(ISD::BUILD_PAIR, dl, RoundVT, Lo, Hi); + Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi); if (RoundParts < NumParts) { // Assemble the trailing non-power-of-2 part. unsigned OddParts = NumParts - RoundParts; EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits); - Hi = getCopyFromParts(DAG, dl, + Hi = getCopyFromParts(DAG, DL, Parts + RoundParts, OddParts, PartVT, OddVT); // Combine the round and odd parts. @@ -126,68 +133,29 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, if (TLI.isBigEndian()) std::swap(Lo, Hi); EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); - Hi = DAG.getNode(ISD::ANY_EXTEND, dl, TotalVT, Hi); - Hi = DAG.getNode(ISD::SHL, dl, TotalVT, Hi, + Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi); + Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi, DAG.getConstant(Lo.getValueType().getSizeInBits(), TLI.getPointerTy())); - Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, TotalVT, Lo); - Val = DAG.getNode(ISD::OR, dl, TotalVT, Lo, Hi); - } - } else if (ValueVT.isVector()) { - // Handle a multi-element vector. - EVT IntermediateVT, RegisterVT; - unsigned NumIntermediates; - unsigned NumRegs = - TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT, - NumIntermediates, RegisterVT); - assert(NumRegs == NumParts - && "Part count doesn't match vector breakdown!"); - NumParts = NumRegs; // Silence a compiler warning. - assert(RegisterVT == PartVT - && "Part type doesn't match vector breakdown!"); - assert(RegisterVT == Parts[0].getValueType() && - "Part type doesn't match part!"); - - // Assemble the parts into intermediate operands. - SmallVector<SDValue, 8> Ops(NumIntermediates); - if (NumIntermediates == NumParts) { - // If the register was not expanded, truncate or copy the value, - // as appropriate. - for (unsigned i = 0; i != NumParts; ++i) - Ops[i] = getCopyFromParts(DAG, dl, &Parts[i], 1, - PartVT, IntermediateVT); - } else if (NumParts > 0) { - // If the intermediate type was expanded, build the intermediate - // operands from the parts. - assert(NumParts % NumIntermediates == 0 && - "Must expand into a divisible number of parts!"); - unsigned Factor = NumParts / NumIntermediates; - for (unsigned i = 0; i != NumIntermediates; ++i) - Ops[i] = getCopyFromParts(DAG, dl, &Parts[i * Factor], Factor, - PartVT, IntermediateVT); + Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo); + Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi); } - - // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the - // intermediate operands. - Val = DAG.getNode(IntermediateVT.isVector() ? - ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, dl, - ValueVT, &Ops[0], NumIntermediates); } else if (PartVT.isFloatingPoint()) { // FP split into multiple FP parts (for ppcf128) assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) && "Unexpected split"); SDValue Lo, Hi; - Lo = DAG.getNode(ISD::BIT_CONVERT, dl, EVT(MVT::f64), Parts[0]); - Hi = DAG.getNode(ISD::BIT_CONVERT, dl, EVT(MVT::f64), Parts[1]); + Lo = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[0]); + Hi = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[1]); if (TLI.isBigEndian()) std::swap(Lo, Hi); - Val = DAG.getNode(ISD::BUILD_PAIR, dl, ValueVT, Lo, Hi); + Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi); } else { // FP split into integer parts (soft fp) assert(ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && "Unexpected split"); EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits()); - Val = getCopyFromParts(DAG, dl, Parts, NumParts, PartVT, IntVT); + Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT); } } @@ -197,219 +165,315 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, if (PartVT == ValueVT) return Val; - if (PartVT.isVector()) { - assert(ValueVT.isVector() && "Unknown vector conversion!"); - return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val); - } - - if (ValueVT.isVector()) { - assert(ValueVT.getVectorElementType() == PartVT && - ValueVT.getVectorNumElements() == 1 && - "Only trivial scalar-to-vector conversions should get here!"); - return DAG.getNode(ISD::BUILD_VECTOR, dl, ValueVT, Val); - } - - if (PartVT.isInteger() && - ValueVT.isInteger()) { + if (PartVT.isInteger() && ValueVT.isInteger()) { if (ValueVT.bitsLT(PartVT)) { // For a truncate, see if we have any information to // indicate whether the truncated bits will always be // zero or sign-extension. if (AssertOp != ISD::DELETED_NODE) - Val = DAG.getNode(AssertOp, dl, PartVT, Val, + Val = DAG.getNode(AssertOp, DL, PartVT, Val, DAG.getValueType(ValueVT)); - return DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val); - } else { - return DAG.getNode(ISD::ANY_EXTEND, dl, ValueVT, Val); + return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); } + return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val); } if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { - if (ValueVT.bitsLT(Val.getValueType())) { - // FP_ROUND's are always exact here. - return DAG.getNode(ISD::FP_ROUND, dl, ValueVT, Val, + // FP_ROUND's are always exact here. + if (ValueVT.bitsLT(Val.getValueType())) + return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val, DAG.getIntPtrConstant(1)); - } - return DAG.getNode(ISD::FP_EXTEND, dl, ValueVT, Val); + return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val); } if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) - return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val); + return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val); llvm_unreachable("Unknown mismatch!"); return SDValue(); } +/// getCopyFromParts - Create a value that contains the specified legal parts +/// combined into the value they represent. If the parts combine to a type +/// larger then ValueVT then AssertOp can be used to specify whether the extra +/// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT +/// (ISD::AssertSext). +static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL, + const SDValue *Parts, unsigned NumParts, + EVT PartVT, EVT ValueVT) { + assert(ValueVT.isVector() && "Not a vector value"); + assert(NumParts > 0 && "No parts to assemble!"); + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); + SDValue Val = Parts[0]; + + // Handle a multi-element vector. + if (NumParts > 1) { + EVT IntermediateVT, RegisterVT; + unsigned NumIntermediates; + unsigned NumRegs = + TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT, + NumIntermediates, RegisterVT); + assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!"); + NumParts = NumRegs; // Silence a compiler warning. + assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!"); + assert(RegisterVT == Parts[0].getValueType() && + "Part type doesn't match part!"); + + // Assemble the parts into intermediate operands. + SmallVector<SDValue, 8> Ops(NumIntermediates); + if (NumIntermediates == NumParts) { + // If the register was not expanded, truncate or copy the value, + // as appropriate. + for (unsigned i = 0; i != NumParts; ++i) + Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1, + PartVT, IntermediateVT); + } else if (NumParts > 0) { + // If the intermediate type was expanded, build the intermediate + // operands from the parts. + assert(NumParts % NumIntermediates == 0 && + "Must expand into a divisible number of parts!"); + unsigned Factor = NumParts / NumIntermediates; + for (unsigned i = 0; i != NumIntermediates; ++i) + Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor, + PartVT, IntermediateVT); + } + + // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the + // intermediate operands. + Val = DAG.getNode(IntermediateVT.isVector() ? + ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL, + ValueVT, &Ops[0], NumIntermediates); + } + + // There is now one part, held in Val. Correct it to match ValueVT. + PartVT = Val.getValueType(); + + if (PartVT == ValueVT) + return Val; + + if (PartVT.isVector()) { + // If the element type of the source/dest vectors are the same, but the + // parts vector has more elements than the value vector, then we have a + // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the + // elements we want. + if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) { + assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() && + "Cannot narrow, it would be a lossy transformation"); + return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val, + DAG.getIntPtrConstant(0)); + } + + // Vector/Vector bitcast. + return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val); + } + + assert(ValueVT.getVectorElementType() == PartVT && + ValueVT.getVectorNumElements() == 1 && + "Only trivial scalar-to-vector conversions should get here!"); + return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val); +} + + + + +static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc dl, + SDValue Val, SDValue *Parts, unsigned NumParts, + EVT PartVT); + /// getCopyToParts - Create a series of nodes that contain the specified value /// split into legal parts. If the parts contain more bits than Val, then, for /// integers, ExtendKind can be used to specify how to generate the extra bits. -static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, +static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL, SDValue Val, SDValue *Parts, unsigned NumParts, EVT PartVT, ISD::NodeType ExtendKind = ISD::ANY_EXTEND) { - const TargetLowering &TLI = DAG.getTargetLoweringInfo(); - EVT PtrVT = TLI.getPointerTy(); EVT ValueVT = Val.getValueType(); + + // Handle the vector case separately. + if (ValueVT.isVector()) + return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT); + + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); unsigned PartBits = PartVT.getSizeInBits(); unsigned OrigNumParts = NumParts; assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!"); - if (!NumParts) + if (NumParts == 0) return; - if (!ValueVT.isVector()) { - if (PartVT == ValueVT) { - assert(NumParts == 1 && "No-op copy with multiple parts!"); - Parts[0] = Val; - return; - } - - if (NumParts * PartBits > ValueVT.getSizeInBits()) { - // If the parts cover more bits than the value has, promote the value. - if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { - assert(NumParts == 1 && "Do not know what to promote to!"); - Val = DAG.getNode(ISD::FP_EXTEND, dl, PartVT, Val); - } else if (PartVT.isInteger() && ValueVT.isInteger()) { - ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); - Val = DAG.getNode(ExtendKind, dl, ValueVT, Val); - } else { - llvm_unreachable("Unknown mismatch!"); - } - } else if (PartBits == ValueVT.getSizeInBits()) { - // Different types of the same size. - assert(NumParts == 1 && PartVT != ValueVT); - Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val); - } else if (NumParts * PartBits < ValueVT.getSizeInBits()) { - // If the parts cover less bits than value has, truncate the value. - if (PartVT.isInteger() && ValueVT.isInteger()) { - ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); - Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val); - } else { - llvm_unreachable("Unknown mismatch!"); - } - } - - // The value may have changed - recompute ValueVT. - ValueVT = Val.getValueType(); - assert(NumParts * PartBits == ValueVT.getSizeInBits() && - "Failed to tile the value with PartVT!"); - - if (NumParts == 1) { - assert(PartVT == ValueVT && "Type conversion failed!"); - Parts[0] = Val; - return; - } + assert(!ValueVT.isVector() && "Vector case handled elsewhere"); + if (PartVT == ValueVT) { + assert(NumParts == 1 && "No-op copy with multiple parts!"); + Parts[0] = Val; + return; + } - // Expand the value into multiple parts. - if (NumParts & (NumParts - 1)) { - // The number of parts is not a power of 2. Split off and copy the tail. + if (NumParts * PartBits > ValueVT.getSizeInBits()) { + // If the parts cover more bits than the value has, promote the value. + if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { + assert(NumParts == 1 && "Do not know what to promote to!"); + Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val); + } else { assert(PartVT.isInteger() && ValueVT.isInteger() && - "Do not know what to expand to!"); - unsigned RoundParts = 1 << Log2_32(NumParts); - unsigned RoundBits = RoundParts * PartBits; - unsigned OddParts = NumParts - RoundParts; - SDValue OddVal = DAG.getNode(ISD::SRL, dl, ValueVT, Val, - DAG.getConstant(RoundBits, - TLI.getPointerTy())); - getCopyToParts(DAG, dl, OddVal, Parts + RoundParts, - OddParts, PartVT); - - if (TLI.isBigEndian()) - // The odd parts were reversed by getCopyToParts - unreverse them. - std::reverse(Parts + RoundParts, Parts + NumParts); - - NumParts = RoundParts; + "Unknown mismatch!"); ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); - Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val); + Val = DAG.getNode(ExtendKind, DL, ValueVT, Val); } + } else if (PartBits == ValueVT.getSizeInBits()) { + // Different types of the same size. + assert(NumParts == 1 && PartVT != ValueVT); + Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val); + } else if (NumParts * PartBits < ValueVT.getSizeInBits()) { + // If the parts cover less bits than value has, truncate the value. + assert(PartVT.isInteger() && ValueVT.isInteger() && + "Unknown mismatch!"); + ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); + Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); + } + + // The value may have changed - recompute ValueVT. + ValueVT = Val.getValueType(); + assert(NumParts * PartBits == ValueVT.getSizeInBits() && + "Failed to tile the value with PartVT!"); - // The number of parts is a power of 2. Repeatedly bisect the value using - // EXTRACT_ELEMENT. - Parts[0] = DAG.getNode(ISD::BIT_CONVERT, dl, - EVT::getIntegerVT(*DAG.getContext(), - ValueVT.getSizeInBits()), - Val); - - for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) { - for (unsigned i = 0; i < NumParts; i += StepSize) { - unsigned ThisBits = StepSize * PartBits / 2; - EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits); - SDValue &Part0 = Parts[i]; - SDValue &Part1 = Parts[i+StepSize/2]; - - Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, - ThisVT, Part0, - DAG.getConstant(1, PtrVT)); - Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, - ThisVT, Part0, - DAG.getConstant(0, PtrVT)); - - if (ThisBits == PartBits && ThisVT != PartVT) { - Part0 = DAG.getNode(ISD::BIT_CONVERT, dl, - PartVT, Part0); - Part1 = DAG.getNode(ISD::BIT_CONVERT, dl, - PartVT, Part1); - } + if (NumParts == 1) { + assert(PartVT == ValueVT && "Type conversion failed!"); + Parts[0] = Val; + return; + } + + // Expand the value into multiple parts. + if (NumParts & (NumParts - 1)) { + // The number of parts is not a power of 2. Split off and copy the tail. + assert(PartVT.isInteger() && ValueVT.isInteger() && + "Do not know what to expand to!"); + unsigned RoundParts = 1 << Log2_32(NumParts); + unsigned RoundBits = RoundParts * PartBits; + unsigned OddParts = NumParts - RoundParts; + SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val, + DAG.getIntPtrConstant(RoundBits)); + getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT); + + if (TLI.isBigEndian()) + // The odd parts were reversed by getCopyToParts - unreverse them. + std::reverse(Parts + RoundParts, Parts + NumParts); + + NumParts = RoundParts; + ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); + Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); + } + + // The number of parts is a power of 2. Repeatedly bisect the value using + // EXTRACT_ELEMENT. + Parts[0] = DAG.getNode(ISD::BIT_CONVERT, DL, + EVT::getIntegerVT(*DAG.getContext(), + ValueVT.getSizeInBits()), + Val); + + for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) { + for (unsigned i = 0; i < NumParts; i += StepSize) { + unsigned ThisBits = StepSize * PartBits / 2; + EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits); + SDValue &Part0 = Parts[i]; + SDValue &Part1 = Parts[i+StepSize/2]; + + Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, + ThisVT, Part0, DAG.getIntPtrConstant(1)); + Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, + ThisVT, Part0, DAG.getIntPtrConstant(0)); + + if (ThisBits == PartBits && ThisVT != PartVT) { + Part0 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part0); + Part1 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part1); } } + } - if (TLI.isBigEndian()) - std::reverse(Parts, Parts + OrigNumParts); + if (TLI.isBigEndian()) + std::reverse(Parts, Parts + OrigNumParts); +} - return; - } - // Vector ValueVT. +/// getCopyToPartsVector - Create a series of nodes that contain the specified +/// value split into legal parts. +static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL, + SDValue Val, SDValue *Parts, unsigned NumParts, + EVT PartVT) { + EVT ValueVT = Val.getValueType(); + assert(ValueVT.isVector() && "Not a vector"); + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); + if (NumParts == 1) { - if (PartVT != ValueVT) { - if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) { - Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val); - } else { - assert(ValueVT.getVectorElementType() == PartVT && - ValueVT.getVectorNumElements() == 1 && - "Only trivial vector-to-scalar conversions should get here!"); - Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, - PartVT, Val, - DAG.getConstant(0, PtrVT)); - } - } + if (PartVT == ValueVT) { + // Nothing to do. + } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) { + // Bitconvert vector->vector case. + Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val); + } else if (PartVT.isVector() && + PartVT.getVectorElementType() == ValueVT.getVectorElementType()&& + PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) { + EVT ElementVT = PartVT.getVectorElementType(); + // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in + // undef elements. + SmallVector<SDValue, 16> Ops; + for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i) + Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, + ElementVT, Val, DAG.getIntPtrConstant(i))); + + for (unsigned i = ValueVT.getVectorNumElements(), + e = PartVT.getVectorNumElements(); i != e; ++i) + Ops.push_back(DAG.getUNDEF(ElementVT)); + + Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size()); + // FIXME: Use CONCAT for 2x -> 4x. + + //SDValue UndefElts = DAG.getUNDEF(VectorTy); + //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts); + } else { + // Vector -> scalar conversion. + assert(ValueVT.getVectorElementType() == PartVT && + ValueVT.getVectorNumElements() == 1 && + "Only trivial vector-to-scalar conversions should get here!"); + Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, + PartVT, Val, DAG.getIntPtrConstant(0)); + } + Parts[0] = Val; return; } - + // Handle a multi-element vector. EVT IntermediateVT, RegisterVT; unsigned NumIntermediates; unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, - IntermediateVT, NumIntermediates, RegisterVT); + IntermediateVT, + NumIntermediates, RegisterVT); unsigned NumElements = ValueVT.getVectorNumElements(); - + assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!"); NumParts = NumRegs; // Silence a compiler warning. assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!"); - + // Split the vector into intermediate operands. SmallVector<SDValue, 8> Ops(NumIntermediates); for (unsigned i = 0; i != NumIntermediates; ++i) { if (IntermediateVT.isVector()) - Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, + Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, IntermediateVT, Val, - DAG.getConstant(i * (NumElements / NumIntermediates), - PtrVT)); + DAG.getIntPtrConstant(i * (NumElements / NumIntermediates))); else - Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, - IntermediateVT, Val, - DAG.getConstant(i, PtrVT)); + Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, + IntermediateVT, Val, DAG.getIntPtrConstant(i)); } - + // Split the intermediate operands into legal parts. if (NumParts == NumIntermediates) { // If the register was not expanded, promote or copy the value, // as appropriate. for (unsigned i = 0; i != NumParts; ++i) - getCopyToParts(DAG, dl, Ops[i], &Parts[i], 1, PartVT); + getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT); } else if (NumParts > 0) { // If the intermediate type was expanded, split each the value into // legal parts. @@ -417,10 +481,13 @@ static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, "Must expand into a divisible number of parts!"); unsigned Factor = NumParts / NumIntermediates; for (unsigned i = 0; i != NumIntermediates; ++i) - getCopyToParts(DAG, dl, Ops[i], &Parts[i*Factor], Factor, PartVT); + getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT); } } + + + namespace { /// RegsForValue - This struct represents the registers (physical or virtual) /// that a particular set of values is assigned, and the type information @@ -460,11 +527,6 @@ namespace { EVT regvt, EVT valuevt) : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {} - RegsForValue(const SmallVector<unsigned, 4> ®s, - const SmallVector<EVT, 4> ®vts, - const SmallVector<EVT, 4> &valuevts) - : ValueVTs(valuevts), RegVTs(regvts), Regs(regs) {} - RegsForValue(LLVMContext &Context, const TargetLowering &tli, unsigned Reg, const Type *Ty) { ComputeValueVTs(tli, Ty, ValueVTs); @@ -530,6 +592,10 @@ SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo, DebugLoc dl, SDValue &Chain, SDValue *Flag) const { + // A Value with type {} or [0 x %t] needs no registers. + if (ValueVTs.empty()) + return SDValue(); + const TargetLowering &TLI = DAG.getTargetLoweringInfo(); // Assemble the legal parts into the final values. @@ -623,8 +689,7 @@ void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl, unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT); EVT RegisterVT = RegVTs[Value]; - getCopyToParts(DAG, dl, - Val.getValue(Val.getResNo() + Value), + getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value), &Parts[Part], NumParts, RegisterVT); Part += NumParts; } @@ -701,6 +766,7 @@ void SelectionDAGBuilder::clear() { UnusedArgNodeMap.clear(); PendingLoads.clear(); PendingExports.clear(); + DanglingDebugInfoMap.clear(); CurDebugLoc = DebugLoc(); HasTailCall = false; } @@ -805,6 +871,33 @@ void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) { } } +// resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V, +// generate the debug data structures now that we've seen its definition. +void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V, + SDValue Val) { + DanglingDebugInfo &DDI = DanglingDebugInfoMap[V]; + if (DDI.getDI()) { + const DbgValueInst *DI = DDI.getDI(); + DebugLoc dl = DDI.getdl(); + unsigned DbgSDNodeOrder = DDI.getSDNodeOrder(); + MDNode *Variable = DI->getVariable(); + uint64_t Offset = DI->getOffset(); + SDDbgValue *SDV; + if (Val.getNode()) { + if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) { + SDV = DAG.getDbgValue(Variable, Val.getNode(), + Val.getResNo(), Offset, dl, DbgSDNodeOrder); + DAG.AddDbgValue(SDV, Val.getNode(), false); + } + } else { + SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()), + Offset, dl, SDNodeOrder); + DAG.AddDbgValue(SDV, 0, false); + } + DanglingDebugInfoMap[V] = DanglingDebugInfo(); + } +} + // getValue - Return an SDValue for the given Value. SDValue SelectionDAGBuilder::getValue(const Value *V) { // If we already have an SDValue for this value, use it. It's important @@ -826,6 +919,7 @@ SDValue SelectionDAGBuilder::getValue(const Value *V) { // Otherwise create a new SDValue and remember it. SDValue Val = getValueImpl(V); NodeMap[V] = Val; + resolveDanglingDebugInfo(V, Val); return Val; } @@ -839,10 +933,11 @@ SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) { // Otherwise create a new SDValue and remember it. SDValue Val = getValueImpl(V); NodeMap[V] = Val; + resolveDanglingDebugInfo(V, Val); return Val; } -/// getValueImpl - Helper function for getValue and getMaterializedValue. +/// getValueImpl - Helper function for getValue and getNonRegisterValue. /// Create an SDValue for the given value. SDValue SelectionDAGBuilder::getValueImpl(const Value *V) { if (const Constant *C = dyn_cast<Constant>(V)) { @@ -986,10 +1081,10 @@ void SelectionDAGBuilder::visitRet(const ReturnInst &I) { unsigned NumValues = ValueVTs.size(); SmallVector<SDValue, 4> Chains(NumValues); - EVT PtrVT = PtrValueVTs[0]; for (unsigned i = 0; i != NumValues; ++i) { - SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, RetPtr, - DAG.getConstant(Offsets[i], PtrVT)); + SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), + RetPtr.getValueType(), RetPtr, + DAG.getIntPtrConstant(Offsets[i])); Chains[i] = DAG.getStore(Chain, getCurDebugLoc(), SDValue(RetOp.getNode(), RetOp.getResNo() + i), @@ -2709,11 +2804,6 @@ void SelectionDAGBuilder::visitGetElementPtr(const User &I) { } Ty = StTy->getElementType(Field); - } else if (const UnionType *UnTy = dyn_cast<UnionType>(Ty)) { - unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); - - // Offset canonically 0 for unions, but type changes - Ty = UnTy->getElementType(Field); } else { Ty = cast<SequentialType>(Ty)->getElementType(); @@ -2818,7 +2908,7 @@ void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) { // Inform the Frame Information that we have just allocated a variable-sized // object. - FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(); + FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1); } void SelectionDAGBuilder::visitLoad(const LoadInst &I) { @@ -3824,11 +3914,11 @@ static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS, /// argument, create the corresponding DBG_VALUE machine instruction for it now. /// At the end of instruction selection, they will be inserted to the entry BB. bool -SelectionDAGBuilder::EmitFuncArgumentDbgValue(const DbgValueInst &DI, - const Value *V, MDNode *Variable, - uint64_t Offset, +SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable, + int64_t Offset, const SDValue &N) { - if (!isa<Argument>(V)) + const Argument *Arg = dyn_cast<Argument>(V); + if (!Arg) return false; MachineFunction &MF = DAG.getMachineFunction(); @@ -3842,7 +3932,15 @@ SelectionDAGBuilder::EmitFuncArgumentDbgValue(const DbgValueInst &DI, return false; unsigned Reg = 0; - if (N.getOpcode() == ISD::CopyFromReg) { + if (Arg->hasByValAttr()) { + // Byval arguments' frame index is recorded during argument lowering. + // Use this info directly. + const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo(); + Reg = TRI->getFrameRegister(MF); + Offset = FuncInfo.getByValArgumentFrameIndex(Arg); + } + + if (N.getNode() && N.getOpcode() == ISD::CopyFromReg) { Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg(); if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) { MachineRegisterInfo &RegInfo = MF.getRegInfo(); @@ -3966,42 +4064,40 @@ SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { } case Intrinsic::dbg_declare: { const DbgDeclareInst &DI = cast<DbgDeclareInst>(I); - if (!DIVariable(DI.getVariable()).Verify()) - return 0; - MDNode *Variable = DI.getVariable(); - // Parameters are handled specially. - bool isParameter = - DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable; const Value *Address = DI.getAddress(); - if (!Address) + if (!Address || !DIVariable(DI.getVariable()).Verify()) return 0; - if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) - Address = BCI->getOperand(0); - const AllocaInst *AI = dyn_cast<AllocaInst>(Address); - if (AI) { - // Don't handle byval arguments or VLAs, for example. - // Non-byval arguments are handled here (they refer to the stack temporary - // alloca at this point). - DenseMap<const AllocaInst*, int>::iterator SI = - FuncInfo.StaticAllocaMap.find(AI); - if (SI == FuncInfo.StaticAllocaMap.end()) - return 0; // VLAs. - int FI = SI->second; - - MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); - if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo()) - MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc()); - } // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder // but do not always have a corresponding SDNode built. The SDNodeOrder // absolute, but not relative, values are different depending on whether // debug info exists. ++SDNodeOrder; + + // Check if address has undef value. + if (isa<UndefValue>(Address) || + (Address->use_empty() && !isa<Argument>(Address))) { + SDDbgValue*SDV = + DAG.getDbgValue(Variable, UndefValue::get(Address->getType()), + 0, dl, SDNodeOrder); + DAG.AddDbgValue(SDV, 0, false); + return 0; + } + SDValue &N = NodeMap[Address]; + if (!N.getNode() && isa<Argument>(Address)) + // Check unused arguments map. + N = UnusedArgNodeMap[Address]; SDDbgValue *SDV; if (N.getNode()) { + // Parameters are handled specially. + bool isParameter = + DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable; + if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) + Address = BCI->getOperand(0); + const AllocaInst *AI = dyn_cast<AllocaInst>(Address); + if (isParameter && !AI) { FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode()); if (FINode) @@ -4020,10 +4116,14 @@ SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { return 0; DAG.AddDbgValue(SDV, N.getNode(), isParameter); } else { - // This isn't useful, but it shows what we're missing. - SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()), - 0, dl, SDNodeOrder); - DAG.AddDbgValue(SDV, 0, isParameter); + // If Address is an arugment then try to emits its dbg value using + // virtual register info from the FuncInfo.ValueMap. Otherwise add undef + // to help track missing debug info. + if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) { + SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()), + 0, dl, SDNodeOrder); + DAG.AddDbgValue(SDV, 0, false); + } } return 0; } @@ -4048,31 +4148,24 @@ SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder); DAG.AddDbgValue(SDV, 0, false); } else { - bool createUndef = false; - // FIXME : Why not use getValue() directly ? + // Do not use getValue() in here; we don't want to generate code at + // this point if it hasn't been done yet. SDValue N = NodeMap[V]; if (!N.getNode() && isa<Argument>(V)) // Check unused arguments map. N = UnusedArgNodeMap[V]; if (N.getNode()) { - if (!EmitFuncArgumentDbgValue(DI, V, Variable, Offset, N)) { + if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) { SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(), Offset, dl, SDNodeOrder); DAG.AddDbgValue(SDV, N.getNode(), false); } - } else if (isa<PHINode>(V) && !V->use_empty()) { - SDValue N = getValue(V); - if (N.getNode()) { - if (!EmitFuncArgumentDbgValue(DI, V, Variable, Offset, N)) { - SDV = DAG.getDbgValue(Variable, N.getNode(), - N.getResNo(), Offset, dl, SDNodeOrder); - DAG.AddDbgValue(SDV, N.getNode(), false); - } - } else - createUndef = true; - } else - createUndef = true; - if (createUndef) { + } else if (isa<PHINode>(V) && !V->use_empty() ) { + // Do not call getValue(V) yet, as we don't want to generate code. + // Remember it for later. + DanglingDebugInfo DDI(&DI, dl, SDNodeOrder); + DanglingDebugInfoMap[V] = DDI; + } else { // We may expand this to cover more cases. One case where we have no // data available is an unreferenced parameter; we need this fallback. SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()), @@ -4572,6 +4665,11 @@ void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee, !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI)) isTailCall = false; + // If there's a possibility that fast-isel has already selected some amount + // of the current basic block, don't emit a tail call. + if (isTailCall && EnableFastISel) + isTailCall = false; + std::pair<SDValue,SDValue> Result = TLI.LowerCallTo(getRoot(), RetTy, CS.paramHasAttr(0, Attribute::SExt), @@ -6054,6 +6152,12 @@ void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) { i += NumParts; } + // Note down frame index for byval arguments. + if (I->hasByValAttr() && !ArgValues.empty()) + if (FrameIndexSDNode *FI = + dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode())) + FuncInfo->setByValArgumentFrameIndex(I, FI->getIndex()); + if (!I->use_empty()) { SDValue Res; if (!ArgValues.empty()) diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h index 46733d6..5f400e9 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h @@ -18,9 +18,6 @@ #include "llvm/CodeGen/SelectionDAG.h" #include "llvm/ADT/APInt.h" #include "llvm/ADT/DenseMap.h" -#ifndef NDEBUG -#include "llvm/ADT/SmallSet.h" -#endif #include "llvm/CodeGen/SelectionDAGNodes.h" #include "llvm/CodeGen/ValueTypes.h" #include "llvm/Support/CallSite.h" @@ -64,6 +61,7 @@ class PHINode; class PtrToIntInst; class ReturnInst; class SDISelAsmOperandInfo; +class SDDbgValue; class SExtInst; class SelectInst; class ShuffleVectorInst; @@ -93,6 +91,24 @@ class SelectionDAGBuilder { /// to preserve debug information for incoming arguments. DenseMap<const Value*, SDValue> UnusedArgNodeMap; + /// DanglingDebugInfo - Helper type for DanglingDebugInfoMap. + class DanglingDebugInfo { + const DbgValueInst* DI; + DebugLoc dl; + unsigned SDNodeOrder; + public: + DanglingDebugInfo() : DI(0), dl(DebugLoc()), SDNodeOrder(0) { } + DanglingDebugInfo(const DbgValueInst *di, DebugLoc DL, unsigned SDNO) : + DI(di), dl(DL), SDNodeOrder(SDNO) { } + const DbgValueInst* getDI() { return DI; } + DebugLoc getdl() { return dl; } + unsigned getSDNodeOrder() { return SDNodeOrder; } + }; + + /// DanglingDebugInfoMap - Keeps track of dbg_values for which we have not + /// yet seen the referent. We defer handling these until we do see it. + DenseMap<const Value*, DanglingDebugInfo> DanglingDebugInfoMap; + public: /// PendingLoads - Loads are not emitted to the program immediately. We bunch /// them up and then emit token factor nodes when possible. This allows us to @@ -345,6 +361,9 @@ public: void visit(unsigned Opcode, const User &I); + // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V, + // generate the debug data structures now that we've seen its definition. + void resolveDanglingDebugInfo(const Value *V, SDValue Val); SDValue getValue(const Value *V); SDValue getNonRegisterValue(const Value *V); SDValue getValueImpl(const Value *V); @@ -506,13 +525,11 @@ private: void HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB); - /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a - /// function argument, create the corresponding DBG_VALUE machine instruction - /// for it now. At the end of instruction selection, they will be inserted to - /// the entry BB. - bool EmitFuncArgumentDbgValue(const DbgValueInst &DI, - const Value *V, MDNode *Variable, - uint64_t Offset, const SDValue &N); + /// EmitFuncArgumentDbgValue - If V is an function argument then create + /// corresponding DBG_VALUE machine instruction for it now. At the end of + /// instruction selection, they will be inserted to the entry BB. + bool EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable, + int64_t Offset, const SDValue &N); }; } // end namespace llvm diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp index 08ba548..66cb5ce 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp @@ -132,14 +132,16 @@ namespace llvm { const TargetLowering &TLI = IS->getTargetLowering(); if (OptLevel == CodeGenOpt::None) - return createFastDAGScheduler(IS, OptLevel); + return createSourceListDAGScheduler(IS, OptLevel); if (TLI.getSchedulingPreference() == Sched::Latency) return createTDListDAGScheduler(IS, OptLevel); if (TLI.getSchedulingPreference() == Sched::RegPressure) return createBURRListDAGScheduler(IS, OptLevel); - assert(TLI.getSchedulingPreference() == Sched::Hybrid && + if (TLI.getSchedulingPreference() == Sched::Hybrid) + return createHybridListDAGScheduler(IS, OptLevel); + assert(TLI.getSchedulingPreference() == Sched::ILP && "Unknown sched type!"); - return createHybridListDAGScheduler(IS, OptLevel); + return createILPListDAGScheduler(IS, OptLevel); } } @@ -169,7 +171,7 @@ TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, //===----------------------------------------------------------------------===// SelectionDAGISel::SelectionDAGISel(const TargetMachine &tm, CodeGenOpt::Level OL) : - MachineFunctionPass(&ID), TM(tm), TLI(*tm.getTargetLowering()), + MachineFunctionPass(ID), TM(tm), TLI(*tm.getTargetLowering()), FuncInfo(new FunctionLoweringInfo(TLI)), CurDAG(new SelectionDAG(tm)), SDB(new SelectionDAGBuilder(*CurDAG, *FuncInfo, OL)), @@ -216,7 +218,7 @@ static bool FunctionCallsSetJmp(const Function *F) { for (Value::const_use_iterator I = Callee->use_begin(), E = Callee->use_end(); I != E; ++I) - if (const CallInst *CI = dyn_cast<CallInst>(I)) + if (const CallInst *CI = dyn_cast<CallInst>(*I)) if (CI->getParent()->getParent() == F) return true; } @@ -362,38 +364,6 @@ SelectionDAGISel::SelectBasicBlock(BasicBlock::const_iterator Begin, CodeGenAndEmitDAG(); } -namespace { -/// WorkListRemover - This class is a DAGUpdateListener that removes any deleted -/// nodes from the worklist. -class SDOPsWorkListRemover : public SelectionDAG::DAGUpdateListener { - SmallVector<SDNode*, 128> &Worklist; - SmallPtrSet<SDNode*, 128> &InWorklist; -public: - SDOPsWorkListRemover(SmallVector<SDNode*, 128> &wl, - SmallPtrSet<SDNode*, 128> &inwl) - : Worklist(wl), InWorklist(inwl) {} - - void RemoveFromWorklist(SDNode *N) { - if (!InWorklist.erase(N)) return; - - SmallVector<SDNode*, 128>::iterator I = - std::find(Worklist.begin(), Worklist.end(), N); - assert(I != Worklist.end() && "Not in worklist"); - - *I = Worklist.back(); - Worklist.pop_back(); - } - - virtual void NodeDeleted(SDNode *N, SDNode *E) { - RemoveFromWorklist(N); - } - - virtual void NodeUpdated(SDNode *N) { - // Ignore updates. - } -}; -} - void SelectionDAGISel::ComputeLiveOutVRegInfo() { SmallPtrSet<SDNode*, 128> VisitedNodes; SmallVector<SDNode*, 128> Worklist; diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGPrinter.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGPrinter.cpp index 6cae804..8313de5 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGPrinter.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/SelectionDAGPrinter.cpp @@ -199,7 +199,7 @@ const std::string SelectionDAG::getGraphAttrs(const SDNode *N) const { #else errs() << "SelectionDAG::getGraphAttrs is only available in debug builds" << " on systems with Graphviz or gv!\n"; - return std::string(""); + return std::string(); #endif } diff --git a/contrib/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/contrib/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp index 4f38669..b74f600 100644 --- a/contrib/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp +++ b/contrib/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp @@ -651,6 +651,53 @@ static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, return NumVectorRegs; } +/// isLegalRC - Return true if the value types that can be represented by the +/// specified register class are all legal. +bool TargetLowering::isLegalRC(const TargetRegisterClass *RC) const { + for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); + I != E; ++I) { + if (isTypeLegal(*I)) + return true; + } + return false; +} + +/// hasLegalSuperRegRegClasses - Return true if the specified register class +/// has one or more super-reg register classes that are legal. +bool +TargetLowering::hasLegalSuperRegRegClasses(const TargetRegisterClass *RC) const{ + if (*RC->superregclasses_begin() == 0) + return false; + for (TargetRegisterInfo::regclass_iterator I = RC->superregclasses_begin(), + E = RC->superregclasses_end(); I != E; ++I) { + const TargetRegisterClass *RRC = *I; + if (isLegalRC(RRC)) + return true; + } + return false; +} + +/// findRepresentativeClass - Return the largest legal super-reg register class +/// of the register class for the specified type and its associated "cost". +std::pair<const TargetRegisterClass*, uint8_t> +TargetLowering::findRepresentativeClass(EVT VT) const { + const TargetRegisterClass *RC = RegClassForVT[VT.getSimpleVT().SimpleTy]; + if (!RC) + return std::make_pair(RC, 0); + const TargetRegisterClass *BestRC = RC; + for (TargetRegisterInfo::regclass_iterator I = RC->superregclasses_begin(), + E = RC->superregclasses_end(); I != E; ++I) { + const TargetRegisterClass *RRC = *I; + if (RRC->isASubClass() || !isLegalRC(RRC)) + continue; + if (!hasLegalSuperRegRegClasses(RRC)) + return std::make_pair(RRC, 1); + BestRC = RRC; + } + return std::make_pair(BestRC, 1); +} + + /// computeRegisterProperties - Once all of the register classes are added, /// this allows us to compute derived properties we expose. void TargetLowering::computeRegisterProperties() { @@ -736,6 +783,28 @@ void TargetLowering::computeRegisterProperties() { MVT VT = (MVT::SimpleValueType)i; if (isTypeLegal(VT)) continue; + // Determine if there is a legal wider type. If so, we should promote to + // that wider vector type. + EVT EltVT = VT.getVectorElementType(); + unsigned NElts = VT.getVectorNumElements(); + if (NElts != 1) { + bool IsLegalWiderType = false; + for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { + EVT SVT = (MVT::SimpleValueType)nVT; + if (SVT.getVectorElementType() == EltVT && + SVT.getVectorNumElements() > NElts && + isTypeSynthesizable(SVT)) { + TransformToType[i] = SVT; + RegisterTypeForVT[i] = SVT; + NumRegistersForVT[i] = 1; + ValueTypeActions.setTypeAction(VT, Promote); + IsLegalWiderType = true; + break; + } + } + if (IsLegalWiderType) continue; + } + MVT IntermediateVT; EVT RegisterVT; unsigned NumIntermediates; @@ -744,32 +813,29 @@ void TargetLowering::computeRegisterProperties() { RegisterVT, this); RegisterTypeForVT[i] = RegisterVT; - // Determine if there is a legal wider type. - bool IsLegalWiderType = false; - EVT EltVT = VT.getVectorElementType(); - unsigned NElts = VT.getVectorNumElements(); - for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { - EVT SVT = (MVT::SimpleValueType)nVT; - if (isTypeSynthesizable(SVT) && SVT.getVectorElementType() == EltVT && - SVT.getVectorNumElements() > NElts && NElts != 1) { - TransformToType[i] = SVT; - ValueTypeActions.setTypeAction(VT, Promote); - IsLegalWiderType = true; - break; - } - } - if (!IsLegalWiderType) { - EVT NVT = VT.getPow2VectorType(); - if (NVT == VT) { - // Type is already a power of 2. The default action is to split. - TransformToType[i] = MVT::Other; - ValueTypeActions.setTypeAction(VT, Expand); - } else { - TransformToType[i] = NVT; - ValueTypeActions.setTypeAction(VT, Promote); - } + EVT NVT = VT.getPow2VectorType(); + if (NVT == VT) { + // Type is already a power of 2. The default action is to split. + TransformToType[i] = MVT::Other; + ValueTypeActions.setTypeAction(VT, Expand); + } else { + TransformToType[i] = NVT; + ValueTypeActions.setTypeAction(VT, Promote); } } + + // Determine the 'representative' register class for each value type. + // An representative register class is the largest (meaning one which is + // not a sub-register class / subreg register class) legal register class for + // a group of value types. For example, on i386, i8, i16, and i32 + // representative would be GR32; while on x86_64 it's GR64. + for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { + const TargetRegisterClass* RRC; + uint8_t Cost; + tie(RRC, Cost) = findRepresentativeClass((MVT::SimpleValueType)i); + RepRegClassForVT[i] = RRC; + RepRegClassCostForVT[i] = Cost; + } } const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { @@ -798,8 +864,21 @@ unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, EVT &IntermediateVT, unsigned &NumIntermediates, EVT &RegisterVT) const { - // Figure out the right, legal destination reg to copy into. unsigned NumElts = VT.getVectorNumElements(); + + // If there is a wider vector type with the same element type as this one, + // we should widen to that legal vector type. This handles things like + // <2 x float> -> <4 x float>. + if (NumElts != 1 && getTypeAction(VT) == Promote) { + RegisterVT = getTypeToTransformTo(Context, VT); + if (isTypeLegal(RegisterVT)) { + IntermediateVT = RegisterVT; + NumIntermediates = 1; + return 1; + } + } + + // Figure out the right, legal destination reg to copy into. EVT EltTy = VT.getVectorElementType(); unsigned NumVectorRegs = 1; @@ -828,16 +907,12 @@ unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, EVT DestVT = getRegisterType(Context, NewVT); RegisterVT = DestVT; - if (DestVT.bitsLT(NewVT)) { - // Value is expanded, e.g. i64 -> i16. + if (DestVT.bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits()); - } else { - // Otherwise, promotion or legal types use the same number of registers as - // the vector decimated to the appropriate level. - return NumVectorRegs; - } - return 1; + // Otherwise, promotion or legal types use the same number of registers as + // the vector decimated to the appropriate level. + return NumVectorRegs; } /// Get the EVTs and ArgFlags collections that represent the legalized return @@ -1308,9 +1383,32 @@ bool TargetLowering::SimplifyDemandedBits(SDValue Op, } } - if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt), + if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt), KnownZero, KnownOne, TLO, Depth+1)) return true; + + // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits + // are not demanded. This will likely allow the anyext to be folded away. + if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) { + SDValue InnerOp = InOp.getNode()->getOperand(0); + EVT InnerVT = InnerOp.getValueType(); + if ((APInt::getHighBitsSet(BitWidth, + BitWidth - InnerVT.getSizeInBits()) & + DemandedMask) == 0 && + isTypeDesirableForOp(ISD::SHL, InnerVT)) { + EVT ShTy = getShiftAmountTy(); + if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits())) + ShTy = InnerVT; + SDValue NarrowShl = + TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp, + TLO.DAG.getConstant(ShAmt, ShTy)); + return + TLO.CombineTo(Op, + TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(), + NarrowShl)); + } + } + KnownZero <<= SA->getZExtValue(); KnownOne <<= SA->getZExtValue(); // low bits known zero. @@ -1415,11 +1513,10 @@ bool TargetLowering::SimplifyDemandedBits(SDValue Op, // present in the input. APInt NewBits = APInt::getHighBitsSet(BitWidth, - BitWidth - EVT.getScalarType().getSizeInBits()) & - NewMask; + BitWidth - EVT.getScalarType().getSizeInBits()); // If none of the extended bits are demanded, eliminate the sextinreg. - if (NewBits == 0) + if ((NewBits & NewMask) == 0) return TLO.CombineTo(Op, Op.getOperand(0)); APInt InSignBit = APInt::getSignBit(EVT.getScalarType().getSizeInBits()); @@ -1886,12 +1983,9 @@ TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, EVT ExtDstTy = N0.getValueType(); unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); - // If the extended part has any inconsistent bits, it cannot ever - // compare equal. In other words, they have to be all ones or all - // zeros. - APInt ExtBits = - APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits); - if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits) + // If the constant doesn't fit into the number of bits for the source of + // the sign extension, it is impossible for both sides to be equal. + if (C1.getMinSignedBits() > ExtSrcTyBits) return DAG.getConstant(Cond == ISD::SETNE, VT); SDValue ZextOp; @@ -2476,7 +2570,7 @@ void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, int64_t Offs = GA->getOffset(); if (C) Offs += C->getZExtValue(); Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), - C->getDebugLoc(), + C ? C->getDebugLoc() : DebugLoc(), Op.getValueType(), Offs)); return; } |
