[DAGCombiner] Transform (icmp eq/ne (and X,C0),(shift X,C1)) to use…

… rotate or to getter constants.

If `C0` is a mask and `C1` shifts out all the masked bits (to
essentially compare two subsets of `X`), we can arbitrarily re-order
shift as `srl` or `shl`.

If `C1` (shift amount) is a power of 2, we can replace the and+shift
with a rotate.

Otherwise, based on target preference we can arbitrarily swap `shl`
and `shl` in/out to get better constants.

On x86 we can use this re-ordering to:
    1) get better `and` constants for `C0` (zero extended moves or
       avoid imm64).
    2) covert `srl` to `shl` if `shl` will be implementable with `lea`
       or `add` (both of which can be preferable).

Proofs: https://alive2.llvm.org/ce/z/qzGM_w

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D152116

llvm/include/llvm/CodeGen/TargetLowering.h

@@ -828,6 +828,24 @@ class TargetLoweringBase {
     return N->getOpcode() == ISD::FDIV;
   }
 
+  // Given:
+  //    (icmp eq/ne (and X, C0), (shift X, C1))
+  // or
+  //    (icmp eq/ne X, (rotate X, CPow2))
+
+  // If C0 is a mask or shifted mask and the shift amt (C1) isolates the
+  // remaining bits (i.e something like `(x64 & UINT32_MAX) == (x64 >> 32)`)
+  // Do we prefer the shift to be shift-right, shift-left, or rotate.
+  // Note: Its only valid to convert the rotate version to the shift version iff
+  // the shift-amt (`C1`) is a power of 2 (including 0).
+  // If ShiftOpc (current Opcode) is returned, do nothing.
+  virtual unsigned preferedOpcodeForCmpEqPiecesOfOperand(
+      EVT VT, unsigned ShiftOpc, bool MayTransformRotate,
+      const APInt &ShiftOrRotateAmt,
+      const std::optional<APInt> &AndMask) const {
+    return ShiftOpc;
+  }
+
   /// These two forms are equivalent:
   ///   sub %y, (xor %x, -1)
   ///   add (add %x, 1), %y

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -12443,27 +12443,127 @@ SDValue DAGCombiner::visitSETCC(SDNode *N) {
 
   ISD::CondCode Cond = cast<CondCodeSDNode>(N->getOperand(2))->get();
   EVT VT = N->getValueType(0);
+  SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
 
-  SDValue Combined = SimplifySetCC(VT, N->getOperand(0), N->getOperand(1), Cond,
-                                   SDLoc(N), !PreferSetCC);
-
-  if (!Combined)
-    return SDValue();
+  SDValue Combined = SimplifySetCC(VT, N0, N1, Cond, SDLoc(N), !PreferSetCC);
 
-  // If we prefer to have a setcc, and we don't, we'll try our best to
-  // recreate one using rebuildSetCC.
-  if (PreferSetCC && Combined.getOpcode() != ISD::SETCC) {
-    SDValue NewSetCC = rebuildSetCC(Combined);
+  if (Combined) {
+    // If we prefer to have a setcc, and we don't, we'll try our best to
+    // recreate one using rebuildSetCC.
+    if (PreferSetCC && Combined.getOpcode() != ISD::SETCC) {
+      SDValue NewSetCC = rebuildSetCC(Combined);
 
-    // We don't have anything interesting to combine to.
-    if (NewSetCC.getNode() == N)
-      return SDValue();
+      // We don't have anything interesting to combine to.
+      if (NewSetCC.getNode() == N)
+        return SDValue();
 
-    if (NewSetCC)
-      return NewSetCC;
+      if (NewSetCC)
+        return NewSetCC;
+    }
+    return Combined;
   }
 
-  return Combined;
+  // Optimize
+  //    1) (icmp eq/ne (and X, C0), (shift X, C1))
+  // or
+  //    2) (icmp eq/ne X, (rotate X, C1))
+  // If C0 is a mask or shifted mask and the shift amt (C1) isolates the
+  // remaining bits (i.e something like `(x64 & UINT32_MAX) == (x64 >> 32)`)
+  // Then:
+  // If C1 is a power of 2, then the rotate and shift+and versions are
+  // equivilent, so we can interchange them depending on target preference.
+  // Otherwise, if we have the shift+and version we can interchange srl/shl
+  // which inturn affects the constant C0. We can use this to get better
+  // constants again determined by target preference.
+  if (Cond == ISD::SETNE || Cond == ISD::SETEQ) {
+    auto IsAndWithShift = [](SDValue A, SDValue B) {
+      return A.getOpcode() == ISD::AND &&
+             (B.getOpcode() == ISD::SRL || B.getOpcode() == ISD::SHL) &&
+             A.getOperand(0) == B.getOperand(0);
+    };
+    auto IsRotateWithOp = [](SDValue A, SDValue B) {
+      return (B.getOpcode() == ISD::ROTL || B.getOpcode() == ISD::ROTR) &&
+             B.getOperand(0) == A;
+    };
+    SDValue AndOrOp = SDValue(), ShiftOrRotate = SDValue();
+    bool IsRotate = false;
+
+    // Find either shift+and or rotate pattern.
+    if (IsAndWithShift(N0, N1)) {
+      AndOrOp = N0;
+      ShiftOrRotate = N1;
+    } else if (IsAndWithShift(N1, N0)) {
+      AndOrOp = N1;
+      ShiftOrRotate = N0;
+    } else if (IsRotateWithOp(N0, N1)) {
+      IsRotate = true;
+      AndOrOp = N0;
+      ShiftOrRotate = N1;
+    } else if (IsRotateWithOp(N1, N0)) {
+      IsRotate = true;
+      AndOrOp = N1;
+      ShiftOrRotate = N0;
+    }
+
+    if (AndOrOp && ShiftOrRotate && ShiftOrRotate.hasOneUse() &&
+        (IsRotate || AndOrOp.hasOneUse())) {
+      EVT OpVT = N0.getValueType();
+      // Get constant shift/rotate amount and possibly mask (if its shift+and
+      // variant).
+      auto GetAPIntValue = [](SDValue Op) -> std::optional<APInt> {
+        ConstantSDNode *CNode = isConstOrConstSplat(Op, /*AllowUndefs*/ false,
+                                                    /*AllowTrunc*/ false);
+        if (CNode == nullptr)
+          return std::nullopt;
+        return CNode->getAPIntValue();
+      };
+      std::optional<APInt> AndCMask =
+          IsRotate ? std::nullopt : GetAPIntValue(AndOrOp.getOperand(1));
+      std::optional<APInt> ShiftCAmt =
+          GetAPIntValue(ShiftOrRotate.getOperand(1));
+      unsigned NumBits = OpVT.getScalarSizeInBits();
+
+      // We found constants.
+      if (ShiftCAmt && (IsRotate || AndCMask) && ShiftCAmt->ult(NumBits)) {
+        unsigned ShiftOpc = ShiftOrRotate.getOpcode();
+        // Check that the constants meet the constraints.
+        bool CanTransform =
+            IsRotate ||
+            (*ShiftCAmt == (~*AndCMask).popcount() && ShiftOpc == ISD::SHL
+                 ? (~*AndCMask).isMask()
+                 : AndCMask->isMask());
+
+        // See if target prefers another shift/rotate opcode.
+        unsigned NewShiftOpc = TLI.preferedOpcodeForCmpEqPiecesOfOperand(
+            OpVT, ShiftOpc, ShiftCAmt->isPowerOf2(), *ShiftCAmt, AndCMask);
+        // Transform is valid and we have a new preference.
+        if (CanTransform && NewShiftOpc != ShiftOpc) {
+          SDLoc DL(N);
+          SDValue NewShiftOrRotate =
+              DAG.getNode(NewShiftOpc, DL, OpVT, ShiftOrRotate.getOperand(0),
+                          ShiftOrRotate.getOperand(1));
+          SDValue NewAndOrOp = SDValue();
+
+          if (NewShiftOpc == ISD::SHL || NewShiftOpc == ISD::SRL) {
+            APInt NewMask =
+                NewShiftOpc == ISD::SHL
+                    ? APInt::getHighBitsSet(NumBits,
+                                            NumBits - ShiftCAmt->getZExtValue())
+                    : APInt::getLowBitsSet(NumBits,
+                                           NumBits - ShiftCAmt->getZExtValue());
+            NewAndOrOp =
+                DAG.getNode(ISD::AND, DL, OpVT, ShiftOrRotate.getOperand(0),
+                            DAG.getConstant(NewMask, DL, OpVT));
+          } else {
+            NewAndOrOp = ShiftOrRotate.getOperand(0);
+          }
+
+          return DAG.getSetCC(DL, VT, NewAndOrOp, NewShiftOrRotate, Cond);
+        }
+      }
+    }
+  }
+  return SDValue();
 }
 
 SDValue DAGCombiner::visitSETCCCARRY(SDNode *N) {

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -3257,6 +3257,72 @@ bool X86TargetLowering::
   return NewShiftOpcode == ISD::SHL;
 }
 
+unsigned X86TargetLowering::preferedOpcodeForCmpEqPiecesOfOperand(
+    EVT VT, unsigned ShiftOpc, bool MayTransformRotate,
+    const APInt &ShiftOrRotateAmt, const std::optional<APInt> &AndMask) const {
+  if (!VT.isInteger())
+    return ShiftOpc;
+
+  bool PreferRotate = false;
+  if (VT.isVector()) {
+    // For vectors, if we have rotate instruction support, then its definetly
+    // best. Otherwise its not clear what the best so just don't make changed.
+    PreferRotate = Subtarget.hasAVX512() && (VT.getScalarType() == MVT::i32 ||
+                                             VT.getScalarType() == MVT::i64);
+  } else {
+    // For scalar, if we have bmi prefer rotate for rorx. Otherwise prefer
+    // rotate unless we have a zext mask+shr.
+    PreferRotate = Subtarget.hasBMI2();
+    if (!PreferRotate) {
+      unsigned MaskBits =
+          VT.getScalarSizeInBits() - ShiftOrRotateAmt.getZExtValue();
+      PreferRotate = (MaskBits != 8) && (MaskBits != 16) && (MaskBits != 32);
+    }
+  }
+
+  if (ShiftOpc == ISD::SHL || ShiftOpc == ISD::SRL) {
+    assert(AndMask.has_value() && "Null andmask when querying about shift+and");
+
+    if (PreferRotate && MayTransformRotate)
+      return ISD::ROTL;
+
+    // If vector we don't really get much benefit swapping around constants.
+    // Maybe we could check if the DAG has the flipped node already in the
+    // future.
+    if (VT.isVector())
+      return ShiftOpc;
+
+    // See if the beneficial to swap shift type.
+    if (ShiftOpc == ISD::SHL) {
+      // If the current setup has imm64 mask, then inverse will have
+      // at least imm32 mask (or be zext i32 -> i64).
+      if (VT == MVT::i64)
+        return AndMask->getSignificantBits() > 32 ? ISD::SRL : ShiftOpc;
+
+      // We can only benefit if req at least 7-bit for the mask. We
+      // don't want to replace shl of 1,2,3 as they can be implemented
+      // with lea/add.
+      return ShiftOrRotateAmt.uge(7) ? ISD::SRL : ShiftOpc;
+    }
+
+    if (VT == MVT::i64)
+      // Keep exactly 32-bit imm64, this is zext i32 -> i64 which is
+      // extremely efficient.
+      return AndMask->getSignificantBits() > 33 ? ISD::SHL : ShiftOpc;
+
+    // Keep small shifts as shl so we can generate add/lea.
+    return ShiftOrRotateAmt.ult(7) ? ISD::SHL : ShiftOpc;
+  }
+
+  // We prefer rotate for vectors of if we won't get a zext mask with SRL
+  // (PreferRotate will be set in the latter case).
+  if (PreferRotate || VT.isVector())
+    return ShiftOpc;
+
+  // Non-vector type and we have a zext mask with SRL.
+  return ISD::SRL;
+}
+
 bool X86TargetLowering::preferScalarizeSplat(SDNode *N) const {
   return N->getOpcode() != ISD::FP_EXTEND;
 }

llvm/lib/Target/X86/X86ISelLowering.h

@@ -1138,6 +1138,11 @@ namespace llvm {
         unsigned OldShiftOpcode, unsigned NewShiftOpcode,
         SelectionDAG &DAG) const override;
 
+    unsigned preferedOpcodeForCmpEqPiecesOfOperand(
+        EVT VT, unsigned ShiftOpc, bool MayTransformRotate,
+        const APInt &ShiftOrRotateAmt,
+        const std::optional<APInt> &AndMask) const override;
+
     bool preferScalarizeSplat(SDNode *N) const override;
 
     bool shouldFoldConstantShiftPairToMask(const SDNode *N,