[X86] Try Folding icmp of v8i32 -> fcmp of v8f32 on AVX

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -23391,6 +23391,136 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget &Subtarget,
     }
   }
 
+  // We get bad codegen for v8i32 compares on avx targets (without avx2) so if
+  // possible convert to a v8f32 compare.
+  if (VTOp0 == MVT::v8i32 && Subtarget.hasAVX() && !Subtarget.hasAVX2()) {
+    std::optional<KnownBits> KnownOps[2];
+    // Check if an op is known to be in a certain range.
+    auto OpInRange = [&DAG, Op, &KnownOps](unsigned OpNo, bool CmpLT,
+                                           const APInt Bound) {
+      if (!KnownOps[OpNo].has_value())
+        KnownOps[OpNo] = DAG.computeKnownBits(Op.getOperand(OpNo));
+
+      if (KnownOps[OpNo]->isUnknown())
+        return false;
+
+      std::optional<bool> Res;
+      if (CmpLT)
+        Res = KnownBits::ult(*KnownOps[OpNo], KnownBits::makeConstant(Bound));
+      else
+        Res = KnownBits::ugt(*KnownOps[OpNo], KnownBits::makeConstant(Bound));
+      return Res.value_or(false);
+    };
+
+    bool OkayCvt = false;
+    bool OkayBitcast = false;
+
+    const fltSemantics &Sem = SelectionDAG::EVTToAPFloatSemantics(MVT::f32);
+
+    // For cvt up to 1 << (Significand Precision), (1 << 24 for ieee float)
+    const APInt MaxConvertableCvt =
+        APInt::getOneBitSet(32, APFloat::semanticsPrecision(Sem));
+    // For bitcast up to (and including) first inf representation (0x7f800000 +
+    // 1 for ieee float)
+    const APInt MaxConvertableBitcast =
+        APFloat::getInf(Sem).bitcastToAPInt() + 1;
+    // For bitcast we also exclude de-norm values. This is absolutely necessary
+    // for strict semantic correctness, but DAZ (de-norm as zero) will break if
+    // we don't have this check.
+    const APInt MinConvertableBitcast =
+        APFloat::getSmallestNormalized(Sem).bitcastToAPInt() - 1;
+
+    assert(
+        MaxConvertableBitcast.getBitWidth() == 32 &&
+        MaxConvertableCvt == (1U << 24) &&
+        MaxConvertableBitcast == 0x7f800001 &&
+        MinConvertableBitcast.isNonNegative() &&
+        MaxConvertableBitcast.sgt(MinConvertableBitcast) &&
+        "This transform has only been verified to IEEE Single Precision Float");
+
+    // For bitcast we need both lhs/op1 u< MaxConvertableBitcast
+    // NB: It might be worth it to enable to bitcast version for unsigned avx2
+    // comparisons as they typically require multiple instructions to lower
+    // (they don't fit `vpcmpeq`/`vpcmpgt` well).
+    if (OpInRange(1, /*CmpLT*/ true, MaxConvertableBitcast) &&
+        OpInRange(1, /*CmpLT*/ false, MinConvertableBitcast) &&
+        OpInRange(0, /*CmpLT*/ true, MaxConvertableBitcast) &&
+        OpInRange(0, /*CmpLT*/ false, MinConvertableBitcast)) {
+      OkayBitcast = true;
+    }
+    // We want to convert icmp -> fcmp using `sitofp` iff one of the converts
+    // will be constant folded.
+    else if ((DAG.isConstantValueOfAnyType(peekThroughBitcasts(Op1)) ||
+              DAG.isConstantValueOfAnyType(peekThroughBitcasts(Op0)))) {
+      if (isUnsignedIntSetCC(Cond)) {
+        // For cvt + unsigned compare we need both lhs/rhs >= 0 and either lhs
+        // or rhs < MaxConvertableCvt
+
+        if (OpInRange(1, /*CmpLT*/ true, APInt::getSignedMinValue(32)) &&
+            OpInRange(0, /*CmpLT*/ true, APInt::getSignedMinValue(32)) &&
+            (OpInRange(1, /*CmpLT*/ true, MaxConvertableCvt) ||
+             OpInRange(0, /*CmpLT*/ true, MaxConvertableCvt)))
+          OkayCvt = true;
+      } else {
+        // For cvt + signed compare we need  abs(lhs) or abs(rhs) <
+        // MaxConvertableCvt
+        if (OpInRange(1, /*CmpLT*/ true, MaxConvertableCvt) ||
+            OpInRange(1, /*CmpLT*/ false, -MaxConvertableCvt) ||
+            OpInRange(0, /*CmpLT*/ true, MaxConvertableCvt) ||
+            OpInRange(0, /*CmpLT*/ false, -MaxConvertableCvt))
+          OkayCvt = true;
+      }
+    }
+    // TODO: If we can't prove any of the ranges, we could unconditionally lower
+    // `(icmp eq lhs, rhs)` as `(icmp eq (int_to_fp (xor lhs, rhs)), zero)`
+    if (OkayBitcast || OkayCvt) {
+      switch (Cond) {
+      default:
+        llvm_unreachable("Unexpected SETCC condition");
+        // Get the new FP condition. Note for the unsigned conditions we have
+        // verified its okay to convert to the signed version.
+      case ISD::SETULT:
+      case ISD::SETLT:
+        Cond = ISD::SETOLT;
+        break;
+      case ISD::SETUGT:
+      case ISD::SETGT:
+        Cond = ISD::SETOGT;
+        break;
+      case ISD::SETULE:
+      case ISD::SETLE:
+        Cond = ISD::SETOLE;
+        break;
+      case ISD::SETUGE:
+      case ISD::SETGE:
+        Cond = ISD::SETOGE;
+        break;
+      case ISD::SETEQ:
+        Cond = ISD::SETOEQ;
+        break;
+      case ISD::SETNE:
+        Cond = ISD::SETONE;
+        break;
+      }
+
+      MVT FpVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
+      SDNodeFlags Flags;
+      Flags.setNoNaNs(true);
+      Flags.setNoInfs(true);
+      Flags.setNoSignedZeros(true);
+      if (OkayBitcast) {
+        Op0 = DAG.getBitcast(FpVT, Op0);
+        Op1 = DAG.getBitcast(FpVT, Op1);
+      } else {
+        Op0 = DAG.getNode(ISD::SINT_TO_FP, dl, FpVT, Op0);
+        Op1 = DAG.getNode(ISD::SINT_TO_FP, dl, FpVT, Op1);
+      }
+      Op0->setFlags(Flags);
+      Op1->setFlags(Flags);
+      return DAG.getSetCC(dl, VT, Op0, Op1, Cond);
+    }
+  }
+
   // Break 256-bit integer vector compare into smaller ones.
   if (VT.is256BitVector() && !Subtarget.hasInt256())
     return splitIntVSETCC(VT, Op0, Op1, Cond, DAG, dl);
@@ -41216,6 +41346,156 @@ static SDValue combineShuffle(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+// Simplify a decomposed (sext (setcc)). Assumes prior check that
+// bitwidth(sext)==bitwidth(setcc operands).
+static SDValue simplifySExtOfDecomposedSetCCImpl(
+    SelectionDAG &DAG, const SDLoc &DL, ISD::CondCode CC, SDValue Op0,
+    SDValue Op1, const APInt &OriginalDemandedBits,
+    const APInt &OriginalDemandedElts, bool AllowNOT, unsigned Depth) {
+  // Possible TODO: We could handle any power of two demanded bit + unsigned
+  // comparison. There are no x86 specific comparisons that are unsigned so its
+  // unneeded.
+  if (!OriginalDemandedBits.isSignMask())
+    return SDValue();
+
+  EVT OpVT = Op0.getValueType();
+  // We need need nofpclass(nan inf nzero) to handle floats.
+  auto hasOkayFPFlags = [](SDValue Op) {
+    return Op.getOpcode() == ISD::SINT_TO_FP ||
+           Op.getOpcode() == ISD::UINT_TO_FP ||
+           (Op->getFlags().hasNoNaNs() && Op->getFlags().hasNoInfs() &&
+            Op->getFlags().hasNoSignedZeros());
+  };
+
+  if (OpVT.isFloatingPoint() && !hasOkayFPFlags(Op0))
+    return SDValue();
+
+  auto ValsEq = [OpVT](const APInt &V0, APInt V1) -> bool {
+    if (OpVT.isFloatingPoint()) {
+      const fltSemantics &Sem = SelectionDAG::EVTToAPFloatSemantics(OpVT);
+      return V0.eq(APFloat(Sem, V1).bitcastToAPInt());
+    }
+    return V0.eq(V1);
+  };
+
+  // Assume we canonicalized constants to Op1. That isn't always true but we
+  // call this function twice with inverted CC/Operands so its fine either way.
+  APInt Op1C;
+  unsigned ValWidth = OriginalDemandedBits.getBitWidth();
+  if (ISD::isConstantSplatVectorAllZeros(Op1.getNode())) {
+    Op1C = APInt::getZero(ValWidth);
+  } else if (ISD::isConstantSplatVectorAllOnes(Op1.getNode())) {
+    Op1C = APInt::getAllOnes(ValWidth);
+  } else if (auto *C = dyn_cast<ConstantFPSDNode>(Op1)) {
+    Op1C = C->getValueAPF().bitcastToAPInt();
+  } else if (auto *C = dyn_cast<ConstantSDNode>(Op1)) {
+    Op1C = C->getAPIntValue();
+  } else if (ISD::isConstantSplatVector(Op1.getNode(), Op1C)) {
+    // isConstantSplatVector sets `Op1C`.
+  } else {
+    return SDValue();
+  }
+
+  bool Not = false;
+  bool Okay = false;
+  assert(OriginalDemandedBits.getBitWidth() == Op1C.getBitWidth() &&
+         "Invalid constant operand");
+
+  switch (CC) {
+  case ISD::SETGE:
+  case ISD::SETOGE:
+    Not = true;
+    [[fallthrough]];
+  case ISD::SETLT:
+  case ISD::SETOLT:
+    // signbit(sext(x s< 0)) == signbit(x)
+    // signbit(sext(x s>= 0)) == signbit(~x)
+    Okay = ValsEq(Op1C, APInt::getZero(ValWidth));
+    // For float ops we need to ensure Op0 is de-norm. Otherwise DAZ can break
+    // this fold.
+    // NB: We only need de-norm check here, for the rest of the constants any
+    // relationship with a de-norm value and zero will be identical.
+    if (Okay && OpVT.isFloatingPoint()) {
+      // Values from integers are always normal.
+      if (Op0.getOpcode() == ISD::SINT_TO_FP ||
+          Op0.getOpcode() == ISD::UINT_TO_FP)
+        break;
+
+      // See if we can prove normal with known bits.
+      KnownBits Op0Known =
+          DAG.computeKnownBits(Op0, OriginalDemandedElts, Depth);
+      // Negative/positive doesn't matter.
+      Op0Known.One.clearSignBit();
+      Op0Known.Zero.clearSignBit();
+
+      // Get min normal value.
+      const fltSemantics &Sem = SelectionDAG::EVTToAPFloatSemantics(OpVT);
+      KnownBits MinNormal = KnownBits::makeConstant(
+          APFloat::getSmallestNormalized(Sem).bitcastToAPInt());
+      // Are we above de-norm range?
+      std::optional<bool> Op0Normal = KnownBits::uge(Op0Known, MinNormal);
+      Okay = Op0Normal.value_or(false);
+    }
+    break;
+  case ISD::SETGT:
+  case ISD::SETOGT:
+    Not = true;
+    [[fallthrough]];
+  case ISD::SETLE:
+  case ISD::SETOLE:
+    // signbit(sext(x s<= -1)) == signbit(x)
+    // signbit(sext(x s> -1)) == signbit(~x)
+    Okay = ValsEq(Op1C, APInt::getAllOnes(ValWidth));
+    break;
+  case ISD::SETULT:
+    Not = true;
+    [[fallthrough]];
+  case ISD::SETUGE:
+    // signbit(sext(x u>= SIGNED_MIN)) == signbit(x)
+    // signbit(sext(x u< SIGNED_MIN)) == signbit(~x)
+    Okay = !OpVT.isFloatingPoint() && ValsEq(Op1C, OriginalDemandedBits);
+    break;
+  case ISD::SETULE:
+    Not = true;
+    [[fallthrough]];
+  case ISD::SETUGT:
+    // signbit(sext(x u> SIGNED_MAX)) == signbit(x)
+    // signbit(sext(x u<= SIGNED_MAX)) == signbit(~x)
+    Okay = !OpVT.isFloatingPoint() && ValsEq(Op1C, OriginalDemandedBits - 1);
+    break;
+  default:
+    break;
+  }
+
+  Okay &= Not ? AllowNOT : true;
+  if (!Okay)
+    return SDValue();
+
+  if (!Not)
+    return Op0;
+
+  if (!OpVT.isFloatingPoint())
+    return DAG.getNOT(DL, Op0, OpVT);
+
+  // Possible TODO: We could use `fneg` to do not.
+  return SDValue();
+}
+
+static SDValue simplifySExtOfDecomposedSetCC(SelectionDAG &DAG, const SDLoc &DL,
+                                             ISD::CondCode CC, SDValue Op0,
+                                             SDValue Op1,
+                                             const APInt &OriginalDemandedBits,
+                                             const APInt &OriginalDemandedElts,
+                                             bool AllowNOT, unsigned Depth) {
+  if (SDValue R = simplifySExtOfDecomposedSetCCImpl(
+          DAG, DL, CC, Op0, Op1, OriginalDemandedBits, OriginalDemandedElts,
+          AllowNOT, Depth))
+    return R;
+  return simplifySExtOfDecomposedSetCCImpl(
+      DAG, DL, ISD::getSetCCSwappedOperands(CC), Op1, Op0, OriginalDemandedBits,
+      OriginalDemandedElts, AllowNOT, Depth);
+}
+
 // Simplify variable target shuffle masks based on the demanded elements.
 // TODO: Handle DemandedBits in mask indices as well?
 bool X86TargetLowering::SimplifyDemandedVectorEltsForTargetShuffle(
@@ -42395,13 +42675,26 @@ bool X86TargetLowering::SimplifyDemandedBitsForTargetNode(
     }
     break;
   }
-  case X86ISD::PCMPGT:
-    // icmp sgt(0, R) == ashr(R, BitWidth-1).
-    // iff we only need the sign bit then we can use R directly.
-    if (OriginalDemandedBits.isSignMask() &&
-        ISD::isBuildVectorAllZeros(Op.getOperand(0).getNode()))
-      return TLO.CombineTo(Op, Op.getOperand(1));
+  case X86ISD::PCMPGT: {
+    SDLoc DL(Op);
+    if (SDValue R = simplifySExtOfDecomposedSetCC(
+            TLO.DAG, DL, ISD::SETGT, Op.getOperand(0), Op.getOperand(1),
+            OriginalDemandedBits, OriginalDemandedElts,
+            /*AllowNOT*/ true, Depth))
+      return TLO.CombineTo(Op, R);
+    break;
+  }
+  case X86ISD::CMPP: {
+    SDLoc DL(Op);
+    ISD::CondCode CC = X86::getCondForCMPPImm(
+        cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
+    if (SDValue R = simplifySExtOfDecomposedSetCC(
+            TLO.DAG, DL, CC, Op.getOperand(0), Op.getOperand(1),
+            OriginalDemandedBits, OriginalDemandedElts,
+            !(TLO.LegalOperations() && TLO.LegalTypes()), Depth))
+      return TLO.CombineTo(Op, R);
     break;
+  }
   case X86ISD::MOVMSK: {
     SDValue Src = Op.getOperand(0);
     MVT SrcVT = Src.getSimpleValueType();
@@ -42585,13 +42878,25 @@ SDValue X86TargetLowering::SimplifyMultipleUseDemandedBitsForTargetNode(
     if (DemandedBits.isSignMask())
       return Op.getOperand(0);
     break;
-  case X86ISD::PCMPGT:
-    // icmp sgt(0, R) == ashr(R, BitWidth-1).
-    // iff we only need the sign bit then we can use R directly.
-    if (DemandedBits.isSignMask() &&
-        ISD::isBuildVectorAllZeros(Op.getOperand(0).getNode()))
-      return Op.getOperand(1);
+  case X86ISD::PCMPGT: {
+    SDLoc DL(Op);
+    if (SDValue R = simplifySExtOfDecomposedSetCC(
+            DAG, DL, ISD::SETGT, Op.getOperand(0), Op.getOperand(1),
+            DemandedBits, DemandedElts, /*AllowNOT*/ false, Depth))
+      return R;
+    break;
+  }
+  case X86ISD::CMPP: {
+    SDLoc DL(Op);
+    ISD::CondCode CC = X86::getCondForCMPPImm(
+        cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
+    if (SDValue R = simplifySExtOfDecomposedSetCC(DAG, DL, CC, Op.getOperand(0),
+                                                  Op.getOperand(1),
+                                                  DemandedBits, DemandedElts,
+                                                  /*AllowNOT*/ false, Depth))
+      return R;
     break;
+  }
   case X86ISD::BLENDV: {
     // BLENDV: Cond (MSB) ? LHS : RHS
     SDValue Cond = Op.getOperand(0);
@@ -48267,7 +48572,7 @@ static SDValue combineAndShuffleNot(SDNode *N, SelectionDAG &DAG,
 
   // We do not split for SSE at all, but we need to split vectors for AVX1 and
   // AVX2.
-  if (!Subtarget.useAVX512Regs() && VT.is512BitVector() && 
+  if (!Subtarget.useAVX512Regs() && VT.is512BitVector() &&
       TLI.isTypeLegal(VT.getHalfNumVectorElementsVT(*DAG.getContext()))) {
     SDValue LoX, HiX;
     std::tie(LoX, HiX) = splitVector(X, DAG, DL);