Efficient lowering of vector sdiv when the divisor is a splatted power of two constant.

PR 14848. The lowered sequence is based on the existing sequence the target-independent
DAG Combiner creates for the scalar case.

Patch by Zvi Rackover.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@171953 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 4b00b46e73621b5334501e930d0a7aba096a535f..f42884dd2e80deb0687b439fb1a08135184a389f 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1047,6 +1047,8 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
 
       setOperationAction(ISD::SRA,             MVT::v4i32, Custom);
     }
+    setOperationAction(ISD::SDIV,              MVT::v8i16, Custom);
+    setOperationAction(ISD::SDIV,              MVT::v4i32, Custom);
   }
 
   if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
@@ -1111,6 +1113,8 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
     setOperationAction(ISD::SRA,               MVT::v16i16, Custom);
     setOperationAction(ISD::SRA,               MVT::v32i8, Custom);
 
+    setOperationAction(ISD::SDIV,              MVT::v16i16, Custom);
+
     setOperationAction(ISD::SETCC,             MVT::v32i8, Custom);
     setOperationAction(ISD::SETCC,             MVT::v16i16, Custom);
     setOperationAction(ISD::SETCC,             MVT::v8i32, Custom);
@@ -1166,6 +1170,8 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
       setOperationAction(ISD::SHL,             MVT::v8i32, Legal);
 
       setOperationAction(ISD::SRA,             MVT::v8i32, Legal);
+
+      setOperationAction(ISD::SDIV,            MVT::v8i32, Custom);
     } else {
       setOperationAction(ISD::ADD,             MVT::v4i64, Custom);
       setOperationAction(ISD::ADD,             MVT::v8i32, Custom);
@@ -11377,6 +11383,49 @@ static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
   return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
 }
 
+SDValue X86TargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
+  EVT VT = Op.getValueType();
+  EVT EltTy = VT.getVectorElementType();
+  unsigned NumElts = VT.getVectorNumElements();
+  SDValue N0 = Op.getOperand(0);
+  DebugLoc dl = Op.getDebugLoc();
+
+  // Lower sdiv X, pow2-const.
+  BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(Op.getOperand(1));
+  if (!C)
+    return SDValue();
+
+  APInt SplatValue, SplatUndef;
+  unsigned MinSplatBits;
+  bool HasAnyUndefs;
+  if (!C->isConstantSplat(SplatValue, SplatUndef, MinSplatBits, HasAnyUndefs))
+    return SDValue();
+
+  if ((SplatValue != 0) &&
+      (SplatValue.isPowerOf2() || (-SplatValue).isPowerOf2())) {
+    unsigned lg2 = SplatValue.countTrailingZeros();
+    // Splat the sign bit.
+    SDValue Sz = DAG.getConstant(EltTy.getSizeInBits()-1, MVT::i32);
+    SDValue SGN = getTargetVShiftNode(X86ISD::VSRAI, dl, VT, N0, Sz, DAG);
+    // Add (N0 < 0) ? abs2 - 1 : 0;
+    SDValue Amt = DAG.getConstant(EltTy.getSizeInBits() - lg2, MVT::i32);
+    SDValue SRL = getTargetVShiftNode(X86ISD::VSRLI, dl, VT, SGN, Amt, DAG);
+    SDValue ADD = DAG.getNode(ISD::ADD, dl, VT, N0, SRL);
+    SDValue Lg2Amt = DAG.getConstant(lg2, MVT::i32);
+    SDValue SRA = getTargetVShiftNode(X86ISD::VSRAI, dl, VT, ADD, Lg2Amt, DAG);
+
+    // If we're dividing by a positive value, we're done.  Otherwise, we must
+    // negate the result.
+    if (SplatValue.isNonNegative())
+      return SRA;
+
+    SmallVector<SDValue, 16> V(NumElts, DAG.getConstant(0, EltTy));
+    SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], NumElts);
+    return DAG.getNode(ISD::SUB, dl, VT, Zero, SRA);
+  }
+  return SDValue();
+}
+
 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
 
   EVT VT = Op.getValueType();
@@ -12033,6 +12082,7 @@ SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::SUBE:               return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
   case ISD::ADD:                return LowerADD(Op, DAG);
   case ISD::SUB:                return LowerSUB(Op, DAG);
+  case ISD::SDIV:               return LowerSDIV(Op, DAG);
   }
 }
 

diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 16ce364cd5245572246e6df1ac3349a3976a9ce5..35b5abd791eeaa49b529b722b7d3bd9dd37bef06 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -841,6 +841,7 @@ namespace llvm {
     SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerShift(SDValue Op, SelectionDAG &DAG) const;
+    SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) const;
 
     SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
 

diff --git a/test/CodeGen/X86/vec_sdiv_to_shift.ll b/test/CodeGen/X86/vec_sdiv_to_shift.ll
new file mode 100644 (file)
index 0000000..693463b
--- /dev/null
+++ b/test/CodeGen/X86/vec_sdiv_to_shift.ll
@@ -0,0 +1,72 @@
+; RUN: llc < %s -mcpu=penryn -mattr=+avx2 | FileCheck %s
+
+
+define <8 x i16> @sdiv_vec8x16(<8 x i16> %var) {
+entry:
+; CHECK: sdiv_vec8x16
+; CHECK: psraw  $15
+; CHECK: vpsrlw  $11
+; CHECK: vpaddw
+; CHECK: vpsraw  $5
+; CHECK: ret
+  %0 = sdiv <8 x i16> %var, <i16 32, i16 32, i16 32, i16 32, i16 32, i16 32, i16 32, i16 32>
+  ret <8 x i16> %0
+}
+
+define <4 x i32> @sdiv_zero(<4 x i32> %var) {
+entry:
+; CHECK: sdiv_zero
+; CHECK-NOT sra
+; CHECK: ret
+  %0 = sdiv <4 x i32> %var, <i32 0, i32 0, i32 0, i32 0>
+  ret <4 x i32> %0
+}
+
+define <4 x i32> @sdiv_vec4x32(<4 x i32> %var) {
+entry:
+; CHECK: sdiv_vec4x32
+; CHECK: vpsrad $31
+; CHECK: vpsrld $28
+; CHECK: vpaddd
+; CHECK: vpsrad $4
+; CHECK: ret
+%0 = sdiv <4 x i32> %var, <i32 16, i32 16, i32 16, i32 16>
+ret <4 x i32> %0
+}
+
+define <4 x i32> @sdiv_negative(<4 x i32> %var) {
+entry:
+; CHECK: sdiv_negative
+; CHECK: vpsrad $31
+; CHECK: vpsrld $28
+; CHECK: vpaddd
+; CHECK: vpsrad $4
+; CHECK: vpsubd
+; CHECK: ret
+%0 = sdiv <4 x i32> %var, <i32 -16, i32 -16, i32 -16, i32 -16>
+ret <4 x i32> %0
+}
+
+define <8 x i32> @sdiv8x32(<8 x i32> %var) {
+entry:
+; CHECK: sdiv8x32
+; CHECK: vpsrad $31
+; CHECK: vpsrld $26
+; CHECK: vpaddd
+; CHECK: vpsrad $6
+; CHECK: ret
+%0 = sdiv <8 x i32> %var, <i32 64, i32 64, i32 64, i32 64, i32 64, i32 64, i32 64, i32 64>
+ret <8 x i32> %0
+}
+
+define <16 x i16> @sdiv16x16(<16 x i16> %var) {
+entry:
+; CHECK: sdiv16x16
+; CHECK: vpsraw  $15
+; CHECK: vpsrlw  $14
+; CHECK: vpaddw
+; CHECK: vpsraw  $2
+; CHECK: ret
+  %a0 = sdiv <16 x i16> %var, <i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4>
+  ret <16 x i16> %a0
+}