[LegalizeTypes][TargetLowering] Merge getShiftAmountTyForConstant into TargetLowering::getShiftAmountTy.

[craig.topper]

Summary

getShiftAmountTyForConstant is a special helper that changes the
shift amount to i32 if the type chosen by
TargetLowering::getShiftAmountTy can't represent all possible values.
This is needed to satisfy an assert in SelectionDAG::getNode.

It requires additional consideration to know when this helper should be used.
I'm not sure that we are always using it when we should.

This patch merges the getShiftAmountTyForConstant handling into
TargetLowering::getShiftAmountTy so we don't need to think about it
anymore.

Technically this may slightly increase compile times since the majority
of callers of getShiftAmountTy won't need this. Hopefully, this isn't
an issue in practice.

[craig.topper]

Use MVT instead of EVT in getShiftAmountTy. Should make getSizeInBits() cheaper.

[RKSimon]

LGTM - cheers

llvm/include/llvm/CodeGen/TargetLowering.h

public:
   }
 
   /// Return the type for operands of fence.
   /// TODO: Let fence operands be of i32 type and remove this.
   virtual MVT getFenceOperandTy(const DataLayout &DL) const {
     return getPointerTy(DL);
   }
 
-  /// EVT is not used in-tree, but is used by out-of-tree target.
-  /// A documentation for this function would be nice...
+  /// Return the type to use for a scalar shift opcode, given the shifted amount
+  /// type. Targets should return a legal type if the input type is legal.
+  /// Targets can return a type that is too small if the input type is illegal.
   virtual MVT getScalarShiftAmountTy(const DataLayout &, EVT) const;
 
+  /// Returns the type for the shift amount of a shift opcode. For vectors,
+  /// returns the input type. For scalars, behavior depends on \p LegalTypes. If
+  /// \p LegalTypes is true, calls getScalarShiftAmountTy, otherwise uses
+  /// pointer type. If getScalarShiftAmountTy or pointer type cannot represent
+  /// all possible shift amounts, returns MVT::i32. In general, \p LegalTypes
+  /// should be set to true for calls during type legalization and after type
+  /// legalization has been completed.
   EVT getShiftAmountTy(EVT LHSTy, const DataLayout &DL,
                        bool LegalTypes = true) const;
 
   /// Return the preferred type to use for a shift opcode, given the shifted
   /// amount type is \p ShiftValueTy.
   LLVM_READONLY
   virtual LLT getPreferredShiftAmountTy(LLT ShiftValueTy) const {
     return ShiftValueTy;

llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

if (NOutVT.isVector()) {
       }
     }
   }
 
   return DAG.getNode(ISD::ANY_EXTEND, dl, NOutVT,
                      CreateStackStoreLoad(InOp, OutVT));
 }
 
-// Helper for BSWAP/BITREVERSE promotion to ensure we can fit any shift amount
-// in the VT returned by getShiftAmountTy and to return a safe VT if we can't.
-static EVT getShiftAmountTyForConstant(EVT VT, const TargetLowering &TLI,
-                                       SelectionDAG &DAG) {
-  EVT ShiftVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
-  // If any possible shift value won't fit in the prefered type, just use
-  // something safe. It will be legalized when the shift is expanded.
-  if (!ShiftVT.isVector() &&
-      ShiftVT.getSizeInBits() < Log2_32_Ceil(VT.getSizeInBits()))
-    ShiftVT = MVT::i32;
-  return ShiftVT;
-}
-
 SDValue DAGTypeLegalizer::PromoteIntRes_FREEZE(SDNode *N) {
   SDValue V = GetPromotedInteger(N->getOperand(0));
   return DAG.getNode(ISD::FREEZE, SDLoc(N),
                      V.getValueType(), V);
 }
 
 SDValue DAGTypeLegalizer::PromoteIntRes_BSWAP(SDNode *N) {
   SDValue Op = GetPromotedInteger(N->getOperand(0));
   EVT OVT = N->getValueType(0);
   EVT NVT = Op.getValueType();
   SDLoc dl(N);
 
   // If the larger BSWAP isn't supported by the target, try to expand now.
   // If we expand later we'll end up with more operations since we lost the
   // original type. We only do this for scalars since we have a shuffle
   // based lowering for vectors in LegalizeVectorOps.
   if (!OVT.isVector() &&
       !TLI.isOperationLegalOrCustomOrPromote(ISD::BSWAP, NVT)) {
     if (SDValue Res = TLI.expandBSWAP(N, DAG))
       return DAG.getNode(ISD::ANY_EXTEND, dl, NVT, Res);
   }
 
   unsigned DiffBits = NVT.getScalarSizeInBits() - OVT.getScalarSizeInBits();
-  EVT ShiftVT = getShiftAmountTyForConstant(NVT, TLI, DAG);
+  EVT ShiftVT = TLI.getShiftAmountTy(NVT, DAG.getDataLayout());
   return DAG.getNode(ISD::SRL, dl, NVT, DAG.getNode(ISD::BSWAP, dl, NVT, Op),
                      DAG.getConstant(DiffBits, dl, ShiftVT));
 }
 
 SDValue DAGTypeLegalizer::PromoteIntRes_BITREVERSE(SDNode *N) {
   SDValue Op = GetPromotedInteger(N->getOperand(0));
   EVT OVT = N->getValueType(0);
   EVT NVT = Op.getValueType();
   SDLoc dl(N);
 
   // If the larger BITREVERSE isn't supported by the target, try to expand now.
   // If we expand later we'll end up with more operations since we lost the
   // original type. We only do this for scalars since we have a shuffle
   // based lowering for vectors in LegalizeVectorOps.
   if (!OVT.isVector() && OVT.isSimple() &&
       !TLI.isOperationLegalOrCustomOrPromote(ISD::BITREVERSE, NVT)) {
     if (SDValue Res = TLI.expandBITREVERSE(N, DAG))
       return DAG.getNode(ISD::ANY_EXTEND, dl, NVT, Res);
   }
 
   unsigned DiffBits = NVT.getScalarSizeInBits() - OVT.getScalarSizeInBits();
-  EVT ShiftVT = getShiftAmountTyForConstant(NVT, TLI, DAG);
+  EVT ShiftVT = TLI.getShiftAmountTy(NVT, DAG.getDataLayout());
   return DAG.getNode(ISD::SRL, dl, NVT,
                      DAG.getNode(ISD::BITREVERSE, dl, NVT, Op),
                      DAG.getConstant(DiffBits, dl, ShiftVT));
 }
 
 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.
SDValue DAGTypeLegalizer::PromoteIntRes_XMULO(SDNode *N, unsigned ResNo) {
 
   // Overflow occurred if it occurred in the larger type, or if the high part
   // of the result does not zero/sign-extend the low part.  Check this second
   // possibility first.
   SDValue Overflow;
   if (N->getOpcode() == ISD::UMULO) {
     // Unsigned overflow occurred if the high part is non-zero.
     unsigned Shift = SmallVT.getScalarSizeInBits();
-    EVT ShiftTy = getShiftAmountTyForConstant(Mul.getValueType(), TLI, DAG);
+    EVT ShiftTy = TLI.getShiftAmountTy(Mul.getValueType(), DAG.getDataLayout());
     SDValue Hi = DAG.getNode(ISD::SRL, DL, Mul.getValueType(), Mul,
                              DAG.getConstant(Shift, DL, ShiftTy));
     Overflow = DAG.getSetCC(DL, N->getValueType(1), Hi,
                             DAG.getConstant(0, DL, Hi.getValueType()),
                             ISD::SETNE);
   } else {
     // Signed overflow occurred if the high part does not sign extend the low.
     SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, Mul.getValueType(),
void DAGTypeLegalizer::ExpandIntRes_ABS(SDNode *N, SDValue &Lo, SDValue &Hi) {
   // use in LegalizeDAG. The ADD part of the expansion is based on
   // ExpandIntRes_ADDSUB which also uses ADDCARRY/UADDO after checking that
   // ADDCARRY is LegalOrCustom. Each of the pieces here can be further expanded
   // if needed. Shift expansion has a special case for filling with sign bits
   // so that we will only end up with one SRA.
   bool HasAddCarry = TLI.isOperationLegalOrCustom(
       ISD::ADDCARRY, TLI.getTypeToExpandTo(*DAG.getContext(), NVT));
   if (HasAddCarry) {
-    EVT ShiftAmtTy = getShiftAmountTyForConstant(NVT, TLI, DAG);
+    EVT ShiftAmtTy = TLI.getShiftAmountTy(NVT, DAG.getDataLayout());
     SDValue Sign =
         DAG.getNode(ISD::SRA, dl, NVT, Hi,
                     DAG.getConstant(NVT.getSizeInBits() - 1, dl, ShiftAmtTy));
     SDVTList VTList = DAG.getVTList(NVT, getSetCCResultType(NVT));
     Lo = DAG.getNode(ISD::UADDO, dl, VTList, Lo, Sign);
     Hi = DAG.getNode(ISD::ADDCARRY, dl, VTList, Hi, Sign, Lo.getValue(1));
     Lo = DAG.getNode(ISD::XOR, dl, NVT, Lo, Sign);
     Hi = DAG.getNode(ISD::XOR, dl, NVT, Hi, Sign);
SDValue Mask = DAG.getConstant(APInt::getLowBitsSet(Bits, HalfBits), dl,
                                    NVT);
     SDValue LLL = DAG.getNode(ISD::AND, dl, NVT, LL, Mask);
     SDValue RLL = DAG.getNode(ISD::AND, dl, NVT, RL, Mask);
 
     SDValue T = DAG.getNode(ISD::MUL, dl, NVT, LLL, RLL);
     SDValue TL = DAG.getNode(ISD::AND, dl, NVT, T, Mask);
 
     EVT ShiftAmtTy = TLI.getShiftAmountTy(NVT, DAG.getDataLayout());
-    if (APInt::getMaxValue(ShiftAmtTy.getSizeInBits()).ult(HalfBits)) {
-      // The type from TLI is too small to fit the shift amount we want.
-      // Override it with i32. The shift will have to be legalized.
-      ShiftAmtTy = MVT::i32;
-    }
     SDValue Shift = DAG.getConstant(HalfBits, dl, ShiftAmtTy);
     SDValue TH = DAG.getNode(ISD::SRL, dl, NVT, T, Shift);
     SDValue LLH = DAG.getNode(ISD::SRL, dl, NVT, LL, Shift);
     SDValue RLH = DAG.getNode(ISD::SRL, dl, NVT, RL, Shift);
 
     SDValue U = DAG.getNode(ISD::ADD, dl, NVT,
                             DAG.getNode(ISD::MUL, dl, NVT, LLH, RLL), TH);
     SDValue UL = DAG.getNode(ISD::AND, dl, NVT, U, Mask);
if (LegalOrCustom && TLI.shouldExpandShift(DAG, N)) {
     GetExpandedInteger(N->getOperand(0), LHSL, LHSH);
     EVT VT = LHSL.getValueType();
 
     // If the shift amount operand is coming from a vector legalization it may
     // have an illegal type.  Fix that first by casting the operand, otherwise
     // the new SHL_PARTS operation would need further legalization.
     SDValue ShiftOp = N->getOperand(1);
     EVT ShiftTy = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
-    assert(ShiftTy.getScalarSizeInBits() >=
-           Log2_32_Ceil(VT.getScalarSizeInBits()) &&
-           "ShiftAmountTy is too small to cover the range of this type!");
     if (ShiftOp.getValueType() != ShiftTy)
       ShiftOp = DAG.getZExtOrTrunc(ShiftOp, dl, ShiftTy);
 
     SDValue Ops[] = { LHSL, LHSH, ShiftOp };
     Lo = DAG.getNode(PartsOpc, dl, DAG.getVTList(VT, VT), Ops);
     Hi = Lo.getValue(1);
     return;
   }

llvm/lib/CodeGen/TargetLoweringBase.cpp

MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL,
   return MVT::getIntegerVT(DL.getPointerSizeInBits(0));
 }
 
 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL,
                                          bool LegalTypes) const {
   assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
   if (LHSTy.isVector())
     return LHSTy;
-  return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy)
-                    : getPointerTy(DL);
+  MVT ShiftVT =
+      LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) : getPointerTy(DL);
+  // If any possible shift value won't fit in the prefered type, just use
+  // something safe. Assume it will be legalized when the shift is expanded.
+  if (ShiftVT.getSizeInBits() < Log2_32_Ceil(LHSTy.getSizeInBits()))
+    ShiftVT = MVT::i32;
+  assert(ShiftVT.getSizeInBits() >= Log2_32_Ceil(LHSTy.getSizeInBits()) &&
+         "ShiftVT is still too small!");
+  return ShiftVT;
 }
 
 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
   assert(isTypeLegal(VT));
   switch (Op) {
   default:
     return false;
   case ISD::SDIV: