[X86] Remove SDIVREM8_SEXT_HREG/UDIVREM8_ZEXT_HREG and their associated DAG combine and target bits support. Use a post isel peephole instead.

[craig.topper]

Summary

These nodes exist to overcome an isel problem where we can generate a zero extend of an AH register followed by an extract subreg, and another zero extend. The first zero extend exists to avoid a partial register update copying the AH register into the low 8-bits. The second zero extend exists if the user wanted the remainder zero extended.

To make this work we had a DAG combine to morph the DIVREM opcode to a special opcode that included the extend. But then we had to add the new node to computeKnownBits and computeNumSignBits to process the extension portion.

This patch instead removes all of that and adds a late peephole to detect the two extends.

[RKSimon]

lib/Target/X86/X86ISelDAGToDAG.cpp

886–887

This comment probably needs changing.....

[craig.topper]

Move comment.

[RKSimon]

LGTM

lib/Target/X86/X86ISelDAGToDAG.cpp

private:
     bool isMaskZeroExtended(SDNode *N) const;
     bool tryShiftAmountMod(SDNode *N);
 
     MachineSDNode *emitPCMPISTR(unsigned ROpc, unsigned MOpc, bool MayFoldLoad,
                                 const SDLoc &dl, MVT VT, SDNode *Node);
     MachineSDNode *emitPCMPESTR(unsigned ROpc, unsigned MOpc, bool MayFoldLoad,
                                 const SDLoc &dl, MVT VT, SDNode *Node,
                                 SDValue &InFlag);
+
+    bool tryOptimizeRem8Extend(SDNode *N);
   };
 }
 
 
 // Returns true if this masked compare can be implemented legally with this
 // type.
 static bool isLegalMaskCompare(SDNode *N, const X86Subtarget *Subtarget) {
   unsigned Opcode = N->getOpcode();
for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
 
     // Now that we did that, the node is dead.  Increment the iterator to the
     // next node to process, then delete N.
     ++I;
     CurDAG->DeleteNode(N);
   }
 }
 
+// Look for a redundant movzx/movsx that can occur after an 8-bit divrem.
+bool X86DAGToDAGISel::tryOptimizeRem8Extend(SDNode *N) {
+  unsigned Opc = N->getMachineOpcode();
+  if (Opc != X86::MOVZX32rr8 && Opc != X86::MOVSX32rr8 &&
+      Opc != X86::MOVSX64rr8)
+    return false;
+
+  SDValue N0 = N->getOperand(0);
+
+  // We need to be extracting the lower bit of an extend.
+  if (!N0.isMachineOpcode() ||
+      N0.getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG ||
+      !N0.hasOneUse() || N0.getConstantOperandVal(1) != X86::sub_8bit)
+    return false;
+
+  // We're looking for either a movsx or movzx to match the original opcode.
+  unsigned ExpectedOpc = Opc == X86::MOVZX32rr8 ? X86::MOVZX32rr8_NOREX
+                                                : X86::MOVSX32rr8_NOREX;
+  SDValue N00 = N0.getOperand(0);
+  if (!N00.isMachineOpcode() || N00.getMachineOpcode() != ExpectedOpc)
+    return false;
+
+  if (Opc == X86::MOVSX64rr8) {
+    // If we had a sign extend from 8 to 64 bits. We still need to go from 32
+    // to 64.
+    MachineSDNode *Extend = CurDAG->getMachineNode(X86::MOVSX64rr32, SDLoc(N),
+                                                   MVT::i64, N00);
+    ReplaceUses(N, Extend);
+  } else {
+    // Ok we can drop this extend and just use the original extend.
+    ReplaceUses(N, N00.getNode());
+  }
+
+  return true;
+}
 
 void X86DAGToDAGISel::PostprocessISelDAG() {
   // Skip peepholes at -O0.
   if (TM.getOptLevel() == CodeGenOpt::None)
     return;
 
-  // Attempt to remove vectors moves that were inserted to zero upper bits.
-
   SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end();
 
+  bool MadeChange = false;
   while (Position != CurDAG->allnodes_begin()) {
     SDNode *N = &*--Position;
     // Skip dead nodes and any non-machine opcodes.
     if (N->use_empty() || !N->isMachineOpcode())
       continue;
 
+    if (tryOptimizeRem8Extend(N)) {
+      MadeChange = true;
+      continue;
+    }
+
+    // Attempt to remove vectors moves that were inserted to zero upper bits.
+
     if (N->getMachineOpcode() != TargetOpcode::SUBREG_TO_REG)
       continue;
 
     unsigned SubRegIdx = N->getConstantOperandVal(2);
     if (SubRegIdx != X86::sub_xmm && SubRegIdx != X86::sub_ymm)
       continue;
 
     SDValue Move = N->getOperand(1);
while (Position != CurDAG->allnodes_begin()) {
     if ((TSFlags & X86II::EncodingMask) != X86II::VEX &&
         (TSFlags & X86II::EncodingMask) != X86II::EVEX &&
         (TSFlags & X86II::EncodingMask) != X86II::XOP)
       continue;
 
     // Producing instruction is another vector instruction. We can drop the
     // move.
     CurDAG->UpdateNodeOperands(N, N->getOperand(0), In, N->getOperand(2));
-
-    // If the move is now dead, delete it.
-    if (Move.getNode()->use_empty())
-      CurDAG->RemoveDeadNode(Move.getNode());
+    MadeChange = true;
   }
+
+  if (MadeChange)
+    CurDAG->RemoveDeadNodes();
 }
 
 
 /// Emit any code that needs to be executed only in the main function.
 void X86DAGToDAGISel::emitSpecialCodeForMain() {
   if (Subtarget->isTargetCygMing()) {
     TargetLowering::ArgListTy Args;
     auto &DL = CurDAG->getDataLayout();
if (!SDValue(Node, 1).use_empty()) {
                  dbgs() << '\n');
     }
 
     CurDAG->RemoveDeadNode(Node);
     return;
   }
 
   case ISD::SDIVREM:
-  case ISD::UDIVREM:
-  case X86ISD::SDIVREM8_SEXT_HREG:
-  case X86ISD::UDIVREM8_ZEXT_HREG: {
+  case ISD::UDIVREM: {
     SDValue N0 = Node->getOperand(0);
     SDValue N1 = Node->getOperand(1);
 
     unsigned Opc, MOpc;
-    bool isSigned = (Opcode == ISD::SDIVREM ||
-                     Opcode == X86ISD::SDIVREM8_SEXT_HREG);
+    bool isSigned = Opcode == ISD::SDIVREM;
     if (!isSigned) {
       switch (NVT.SimpleTy) {
       default: llvm_unreachable("Unsupported VT!");
       case MVT::i8:  Opc = X86::DIV8r;  MOpc = X86::DIV8m;  break;
       case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
       case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
       case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
       }
if (HiReg == X86::AH && !SDValue(Node, 1).use_empty()) {
       unsigned AHExtOpcode =
           isSigned ? X86::MOVSX32rr8_NOREX : X86::MOVZX32rr8_NOREX;
 
       SDNode *RNode = CurDAG->getMachineNode(AHExtOpcode, dl, MVT::i32,
                                              MVT::Glue, AHCopy, InFlag);
       SDValue Result(RNode, 0);
       InFlag = SDValue(RNode, 1);
 
-      if (Opcode == X86ISD::UDIVREM8_ZEXT_HREG ||
-          Opcode == X86ISD::SDIVREM8_SEXT_HREG) {
-        assert(Node->getValueType(1) == MVT::i32 && "Unexpected result type!");
-      } else {
-        Result =
-            CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result);
-      }
+
       ReplaceUses(SDValue(Node, 1), Result);
       LLVM_DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG);
                  dbgs() << '\n');
     }
     // Copy the division (low) result, if it is needed.
     if (!SDValue(Node, 0).use_empty()) {
       SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
                                                 LoReg, NVT, InFlag);

lib/Target/X86/X86ISelLowering.h

enum NodeType : unsigned {
       BEXTR,
 
       // LOW, HI, FLAGS = umul LHS, RHS.
       UMUL,
 
       // 8-bit SMUL/UMUL - AX, FLAGS = smul8/umul8 AL, RHS.
       SMUL8, UMUL8,
 
-      // 8-bit divrem that zero-extend the high result (AH).
-      UDIVREM8_ZEXT_HREG,
-      SDIVREM8_SEXT_HREG,
-
       // X86-specific multiply by immediate.
       MUL_IMM,
 
       // Vector sign bit extraction.
       MOVMSK,
 
       // Vector bitwise comparisons.
       PTEST,

lib/Target/X86/X86ISelLowering.cpp

const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case X86ISD::ADD:                return "X86ISD::ADD";
   case X86ISD::SUB:                return "X86ISD::SUB";
   case X86ISD::ADC:                return "X86ISD::ADC";
   case X86ISD::SBB:                return "X86ISD::SBB";
   case X86ISD::SMUL:               return "X86ISD::SMUL";
   case X86ISD::UMUL:               return "X86ISD::UMUL";
   case X86ISD::SMUL8:              return "X86ISD::SMUL8";
   case X86ISD::UMUL8:              return "X86ISD::UMUL8";
-  case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
-  case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
   case X86ISD::INC:                return "X86ISD::INC";
   case X86ISD::DEC:                return "X86ISD::DEC";
   case X86ISD::OR:                 return "X86ISD::OR";
   case X86ISD::XOR:                return "X86ISD::XOR";
   case X86ISD::AND:                return "X86ISD::AND";
   case X86ISD::BEXTR:              return "X86ISD::BEXTR";
   case X86ISD::MUL_IMM:            return "X86ISD::MUL_IMM";
   case X86ISD::MOVMSK:             return "X86ISD::MOVMSK";
case X86ISD::CMOV: {
     KnownBits Known2;
     DAG.computeKnownBits(Op.getOperand(0), Known2, Depth+1);
 
     // Only known if known in both the LHS and RHS.
     Known.One &= Known2.One;
     Known.Zero &= Known2.Zero;
     break;
   }
-  case X86ISD::UDIVREM8_ZEXT_HREG:
-    // TODO: Support more than just the zero extended bits?
-    if (Op.getResNo() != 1)
-      break;
-    // The remainder is zero extended.
-    Known.Zero.setBitsFrom(8);
-    break;
   }
 
   // Handle target shuffles.
   // TODO - use resolveTargetShuffleInputs once we can limit recursive depth.
   if (isTargetShuffle(Opc)) {
     bool IsUnary;
     SmallVector<int, 64> Mask;
     SmallVector<SDValue, 2> Ops;
case X86ISD::VPCOMU:
     return VTBits;
 
   case X86ISD::CMOV: {
     unsigned Tmp0 = DAG.ComputeNumSignBits(Op.getOperand(0), Depth+1);
     if (Tmp0 == 1) return 1;  // Early out.
     unsigned Tmp1 = DAG.ComputeNumSignBits(Op.getOperand(1), Depth+1);
     return std::min(Tmp0, Tmp1);
   }
-  case X86ISD::SDIVREM8_SEXT_HREG:
-    // TODO: Support more than just the sign extended bits?
-    if (Op.getResNo() != 1)
-      break;
-    // The remainder is sign extended.
-    return VTBits - 7;
   }
 
   // Fallback case.
   return 1;
 }
 
 SDValue X86TargetLowering::unwrapAddress(SDValue N) const {
   if (N->getOpcode() == X86ISD::Wrapper || N->getOpcode() == X86ISD::WrapperRIP)
static SDValue promoteExtBeforeAdd(SDNode *Ext, SelectionDAG &DAG,
   // The wider add is guaranteed to not wrap because both operands are
   // sign-extended.
   SDNodeFlags Flags;
   Flags.setNoSignedWrap(NSW);
   Flags.setNoUnsignedWrap(NUW);
   return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewExt, NewConstant, Flags);
 }
 
-/// (i8,i32 {s/z}ext ({s/u}divrem (i8 x, i8 y)) ->
-/// (i8,i32 ({s/u}divrem_sext_hreg (i8 x, i8 y)
-/// This exposes the {s/z}ext to the sdivrem lowering, so that it directly
-/// extends from AH (which we otherwise need to do contortions to access).
-static SDValue getDivRem8(SDNode *N, SelectionDAG &DAG) {
-  SDValue N0 = N->getOperand(0);
-  auto OpcodeN = N->getOpcode();
-  auto OpcodeN0 = N0.getOpcode();
-  if (!((OpcodeN == ISD::SIGN_EXTEND && OpcodeN0 == ISD::SDIVREM) ||
-        (OpcodeN == ISD::ZERO_EXTEND && OpcodeN0 == ISD::UDIVREM)))
-    return SDValue();
-
-  EVT VT = N->getValueType(0);
-  EVT InVT = N0.getValueType();
-  if (N0.getResNo() != 1 || InVT != MVT::i8 ||
-      !(VT == MVT::i32 || VT == MVT::i64))
-    return SDValue();
-
-  SDVTList NodeTys = DAG.getVTList(MVT::i8, MVT::i32);
-  auto DivRemOpcode = OpcodeN0 == ISD::SDIVREM ? X86ISD::SDIVREM8_SEXT_HREG
-                                               : X86ISD::UDIVREM8_ZEXT_HREG;
-  SDValue R = DAG.getNode(DivRemOpcode, SDLoc(N), NodeTys, N0.getOperand(0),
-                          N0.getOperand(1));
-  DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
-  // If this was a 64-bit extend, complete it.
-  if (VT == MVT::i64)
-    return DAG.getNode(OpcodeN, SDLoc(N), VT, R.getValue(1));
-  return R.getValue(1);
-}
-
 // If we face {ANY,SIGN,ZERO}_EXTEND that is applied to a CMOV with constant
 // operands and the result of CMOV is not used anywhere else - promote CMOV
 // itself instead of promoting its result. This could be beneficial, because:
 //     1) X86TargetLowering::EmitLoweredSelect later can do merging of two
 //        (or more) pseudo-CMOVs only when they go one-after-another and
 //        getting rid of result extension code after CMOV will help that.
 //     2) Promotion of constant CMOV arguments is free, hence the
 //        {ANY,SIGN,ZERO}_EXTEND will just be deleted.

 static SDValue combineSext(SDNode *N, SelectionDAG &DAG,
                            TargetLowering::DAGCombinerInfo &DCI,
                            const X86Subtarget &Subtarget) {
   SDValue N0 = N->getOperand(0);
   EVT VT = N->getValueType(0);
   EVT InVT = N0.getValueType();
   SDLoc DL(N);
 
-  if (SDValue DivRem8 = getDivRem8(N, DAG))
-    return DivRem8;
-
   if (SDValue NewCMov = combineToExtendCMOV(N, DAG))
     return NewCMov;
 
   if (!DCI.isBeforeLegalizeOps())
     return SDValue();
 
   if (SDValue V = combineExtSetcc(N, DAG, Subtarget))
     return V;
static SDValue combineZext(SDNode *N, SelectionDAG &DAG,
 
   if (SDValue V = combineToExtendBoolVectorInReg(N, DAG, DCI, Subtarget))
     return V;
 
   if (VT.isVector())
     if (SDValue R = WidenMaskArithmetic(N, DAG, Subtarget))
       return R;
 
-  if (SDValue DivRem8 = getDivRem8(N, DAG))
-    return DivRem8;
-
   if (SDValue NewAdd = promoteExtBeforeAdd(N, DAG, Subtarget))
     return NewAdd;
 
   if (SDValue R = combineOrCmpEqZeroToCtlzSrl(N, DAG, DCI, Subtarget))
     return R;
 
   return SDValue();
 }