[llvm-commits] [llvm] r148225 - in /llvm/trunk: lib/CodeGen/SelectionDAG/TargetLowering.cpp lib/Target/X86/X86ISelLowering.cpp test/CodeGen/X86/blend-msb.ll

Eli Friedman eli.friedman at gmail.com
Tue Jan 17 15:22:32 PST 2012


On Sun, Jan 15, 2012 at 11:27 AM, Nadav Rotem <nadav.rotem at intel.com> wrote:
> Author: nadav
> Date: Sun Jan 15 13:27:55 2012
> New Revision: 148225
>
> URL: http://llvm.org/viewvc/llvm-project?rev=148225&view=rev
> Log:
> [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.

I'm a bit worried about having the semantics of ISD::VSELECT vary
depending on the target. Could you at least put a comment in
ISDOpcodes.h noting this variation?

-Eli

>
> Added:
>    llvm/trunk/test/CodeGen/X86/blend-msb.ll
> Modified:
>    llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
>    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
>
> Modified: llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp?rev=148225&r1=148224&r2=148225&view=diff
> ==============================================================================
> --- llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp (original)
> +++ llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp Sun Jan 15 13:27:55 2012
> @@ -1608,23 +1608,40 @@
>     }
>     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 @@
>     // 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;
>
> Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=148225&r1=148224&r2=148225&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
> +++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Sun Jan 15 13:27:55 2012
> @@ -12868,6 +12868,7 @@
>  /// 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 @@
>     }
>   }
>
> +  // 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 @@
>   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);
>
> Added: llvm/trunk/test/CodeGen/X86/blend-msb.ll
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/blend-msb.ll?rev=148225&view=auto
> ==============================================================================
> --- llvm/trunk/test/CodeGen/X86/blend-msb.ll (added)
> +++ llvm/trunk/test/CodeGen/X86/blend-msb.ll Sun Jan 15 13:27:55 2012
> @@ -0,0 +1,37 @@
> +; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=corei7 -promote-elements -mattr=+sse41 | FileCheck %s
> +
> +
> +; In this test we check that sign-extend of the mask bit is performed by
> +; shifting the needed bit to the MSB, and not using shl+sra.
> +
> +;CHECK: vsel_float
> +;CHECK: pslld
> +;CHECK-NEXT: blendvps
> +;CHECK: ret
> +define <4 x float> @vsel_float(<4 x float> %v1, <4 x float> %v2) {
> +  %vsel = select <4 x i1> <i1 true, i1 false, i1 false, i1 false>, <4 x float> %v1, <4 x float> %v2
> +  ret <4 x float> %vsel
> +}
> +
> +;CHECK: vsel_4xi8
> +;CHECK: pslld
> +;CHECK-NEXT: blendvps
> +;CHECK: ret
> +define <4 x i8> @vsel_4xi8(<4 x i8> %v1, <4 x i8> %v2) {
> +  %vsel = select <4 x i1> <i1 true, i1 false, i1 false, i1 false>, <4 x i8> %v1, <4 x i8> %v2
> +  ret <4 x i8> %vsel
> +}
> +
> +
> +; We do not have native support for v8i16 blends and we have to use the
> +; blendvb instruction or a sequence of NAND/OR/AND. Make sure that we do not r
> +; reduce the mask in this case.
> +;CHECK: vsel_8xi16
> +;CHECK: psllw
> +;CHECK: psraw
> +;CHECK: pblendvb
> +;CHECK: ret
> +define <8 x i16> @vsel_8xi16(<8 x i16> %v1, <8 x i16> %v2) {
> +  %vsel = select <8 x i1> <i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 false, i1 false>, <8 x i16> %v1, <8 x i16> %v2
> +  ret <8 x i16> %vsel
> +}
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits




More information about the llvm-commits mailing list