[AVX] Optimize x86 VSELECT instructions using SimplifyDemandedBits.

We know that the blend instructions only use the MSB, so if the mask is
sign-extended then we can convert it into a SHL instruction. This is a
common pattern because the type-legalizer sign-extends the i1 type which
is used by the LLVM-IR for the condition.

Added a new optimization in SimplifyDemandedBits for SIGN_EXTEND_INREG -> SHL.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@148225 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 44 insertions, 6 deletions

diff --git a/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index bd930d8ef1..5574b21441 100644
--- a/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -1608,23 +1608,40 @@ bool TargetLowering::SimplifyDemandedBits(SDValue Op,
     }
     break;
   case ISD::SIGN_EXTEND_INREG: {
-    EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
+    EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
+
+    APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1);
+    // If we only care about the highest bit, don't bother shifting right.
+    if (MsbMask ==  DemandedMask) {
+      unsigned ShAmt = ExVT.getScalarType().getSizeInBits();
+      SDValue InOp = Op.getOperand(0);
+      EVT InVT = Op.getOperand(0).getValueType();
+      EVT ShTy = getShiftAmountTy(InVT);
+      // In this code we may handle vector types. We can't use the
+      // getShiftAmountTy API because it only works on scalars.
+      // We use the shift value type because we know that its an integer
+      // with enough bits.
+      SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt,
+                                             Op.getValueType());
+      return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
+                                            Op.getValueType(), InOp, ShiftAmt));
+    }
 
     // Sign extension.  Compute the demanded bits in the result that are not
     // present in the input.
     APInt NewBits =
       APInt::getHighBitsSet(BitWidth,
-                            BitWidth - EVT.getScalarType().getSizeInBits());
+                            BitWidth - ExVT.getScalarType().getSizeInBits());
 
     // If none of the extended bits are demanded, eliminate the sextinreg.
     if ((NewBits & NewMask) == 0)
       return TLO.CombineTo(Op, Op.getOperand(0));
 
     APInt InSignBit =
-      APInt::getSignBit(EVT.getScalarType().getSizeInBits()).zext(BitWidth);
+      APInt::getSignBit(ExVT.getScalarType().getSizeInBits()).zext(BitWidth);
     APInt InputDemandedBits =
       APInt::getLowBitsSet(BitWidth,
-                           EVT.getScalarType().getSizeInBits()) &
+                           ExVT.getScalarType().getSizeInBits()) &
       NewMask;
 
     // Since the sign extended bits are demanded, we know that the sign
@@ -1642,7 +1659,7 @@ bool TargetLowering::SimplifyDemandedBits(SDValue Op,
     // If the input sign bit is known zero, convert this into a zero extension.
     if (KnownZero.intersects(InSignBit))
       return TLO.CombineTo(Op,
-                           TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
+                          TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,ExVT));
 
     if (KnownOne.intersects(InSignBit)) {    // Input sign bit known set
       KnownOne |= NewBits;
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 68cd44116d..b5198c8510 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -12868,6 +12868,7 @@ static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
 /// nodes.
 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
+                                    TargetLowering::DAGCombinerInfo &DCI,
                                     const X86Subtarget *Subtarget) {
   DebugLoc DL = N->getDebugLoc();
   SDValue Cond = N->getOperand(0);
@@ -13144,6 +13145,26 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
     }
   }
 
+  // If we know that this node is legal then we know that it is going to be
+  // matched by one of the SSE/AVX BLEND instructions. These instructions only
+  // depend on the highest bit in each word. Try to use SimplifyDemandedBits
+  // to simplify previous instructions.
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
+      !DCI.isBeforeLegalize() &&
+      TLI.isOperationLegal(ISD::VSELECT, VT)) {
+    unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
+    assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
+    APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
+
+    APInt KnownZero, KnownOne;
+    TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
+                                          DCI.isBeforeLegalizeOps());
+    if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
+        TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne, TLO))
+      DCI.CommitTargetLoweringOpt(TLO);
+  }
+
   return SDValue();
 }
 
@@ -14609,7 +14630,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::EXTRACT_VECTOR_ELT:
     return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
   case ISD::VSELECT:
-  case ISD::SELECT:         return PerformSELECTCombine(N, DAG, Subtarget);
+  case ISD::SELECT:         return PerformSELECTCombine(N, DAG, DCI, Subtarget);
   case X86ISD::CMOV:        return PerformCMOVCombine(N, DAG, DCI);
   case ISD::ADD:            return PerformAddCombine(N, DAG, Subtarget);
   case ISD::SUB:            return PerformSubCombine(N, DAG, Subtarget);