[PowerPC] Fix a performance bug for PPC::XXPERMDI.

[jtony]

Summary

There are some VectorShuffle Nodes in SDAG which can be selected to XXPERMDI Instruction, this patch recognizes them and does the selection to improve the PPC performance.

[nemanjai]

lib/Target/PowerPC/PPCISelLowering.cpp

1669 ↗

(On Diff #99739)

Seems like a bit of code duplication with isWordShuffleMask(). Perhaps it would be more consistent to just implement something like isNByteElemShuffleMask(unsigned Width) which can be called for width 1, 2, 4, 8, 16 as needed.

1670 ↗

(On Diff #99739)

Nit: formatting. This is indented more than the rest of the function for some reason.

1671 ↗

(On Diff #99739)

I'm not a fan of loops with 2 iterations. It's probably clearer if this is just straight-line code.

1743 ↗

(On Diff #99739)

Some people really don't like boolean parameters :). I really don't think this function is needed - just do the computation inline.
Especially when I look at the call sites - it looks like there are 3 and you've already checked the endianness on 2 of them.

1748 ↗

(On Diff #99739)

Nit: indentation.

1749 ↗

(On Diff #99739)

Is there a reason for this check (i.e. vs. an assert)? It seems we should be safe to assert this is the case.

1758 ↗

(On Diff #99739)

I think maybe an assert is in order here... Perhaps:
assert((M0 | M1 < 4) && "A mask element out of bounds?")

1763 ↗

(On Diff #99739)

Isn't this just (M0 | M1) < 2? Namely, neither is greater than 1.

1771 ↗

(On Diff #99739)

This entire if statement makes my head hurt. At the very least, it should be documented. But I'm sure when you document it, it'll be obvious how it can be rewritten to be much simpler. Comment it with something along the lines of:

// If element 0 of the result comes from the first input (LE) or second input (BE)
// the inputs need to be swapped and elements adjusted accordingly.

I've suggested simplifications for the LE conditions, but you should be able to simplify the BE conditions accordingly.

1772 ↗

(On Diff #99739)

If we ignore the fact that the second half of this can never be satisfied (M1 can't be 0 and 1), I assume this should be just:
if (M0 > 1 && M1 < 2)

1774 ↗

(On Diff #99739)

Similarly, this one seems like it should just be:
if (M0 < 2 && M1 > 1)

[kbarton]

lib/Target/PowerPC/PPCISelLowering.cpp

1671 ↗

(On Diff #99739)

I agree - this should be just straight line code, with a short comment explaining the logic.

1747 ↗

(On Diff #99739)

Please add a comment explaining the semantics, and what what conditions the parameters will be modified.

1782 ↗

(On Diff #99739)

This is a bit hard to follow, but I don't think Swap is modified on this path. Is that intentional? If so, it needs to be documented clearly in the comments above.

1794 ↗

(On Diff #99739)

Same comment for Swap.

lib/Target/PowerPC/PPCISelLowering.h

457 ↗

(On Diff #99739)

Please add a comment above to be consistent with all of the other declarations here.

[jtony]

lib/Target/PowerPC/PPCISelLowering.cpp

1669 ↗

(On Diff #99739)

Good suggestion, I have added the isNByteElemShuffleMask function you mentioned here to replace the isWordShuffleMask and isDoubleWordShuffleMask function, but only for 2,4,8,16 bytes since there is no need to call this function to check it is 1 byte element shuffle mask, it is always true.

1671 ↗

(On Diff #99739)

This function have been refactored to isNByteElemShuffleMask mentioned by Nemanja above.

[jtony]

Address comments from Nemanja and Kit.

[echristo]

lib/Target/PowerPC/PPCISelLowering.cpp

1772 ↗

(On Diff #100300)

Since this is basically two functions concatenated on each other with a boolean parameter it seems reasonable to just remove the parameter and split it into two functions and dispatch in the caller.

Or do it as a followup with the rest of them, but either way it'd be good to get the endianness stuff cleaned up.

[nemanjai]

lib/Target/PowerPC/PPCISelLowering.cpp

1606 ↗

(On Diff #100300)

Is there a compelling reason to manually allocate memory for this over using efficient containers (something like SmallVector or similar)?

I'd much rather avoid manual memory management of this sort if we can. And I don't see anything in this function to indicate we can't.

[hfinkel]

lib/Target/PowerPC/PPCISelLowering.cpp

1606 ↗

(On Diff #100300)

We should definitely be using a SmallVector here, if we need to keep track of the size easily. Otherwise, given that Width is never greater than 16, we can just use:

unsigned MaskVal[16];

[jtony]

Address comments from Hal and Nemanja.

[jtony]

lib/Target/PowerPC/PPCISelLowering.cpp

1772 ↗

(On Diff #100300)

This can be done together with the same issue created for the xxsldwi patch.

[nemanjai]

In addition to the inline nit, I think it'd be more natural to use the v2i64 type for this PPCISD node (and the instruction) since the instruction inherently operates on doublewords. You should be able to just change the type you bitcast to in the C++ code as well as the type you match in the TblGen code.

Otherwise, LGTM.

lib/Target/PowerPC/PPCISelLowering.cpp

1761 ↗

(On Diff #100564)

A couple of nits that can just be addressed on the commit (no need for another revision):

• This appears to use doxygen-frendly comments with parameter tags, etc. I think doxygen will only use comments that start with three slashes (///).
• I find it confusing for the comment to jump into the implementation like this. A short sentence explaining what the function does, then some details as needed. For example:

/// Can node \p N be lowered to an XXPERMDI instruction? If so, set \p Swap
/// if the inputs to the instruction should be swapped and set \p DM to the
/// value for the immediate.

llvm/trunk/lib/Target/PowerPC/PPCISelLowering.h

enum NodeType : unsigned {
       /// XXINSERT - The PPC VSX insert instruction
       ///
       XXINSERT,
 
       /// VECSHL - The PPC VSX shift left instruction
       ///
       VECSHL,
 
+      /// XXPERMDI - The PPC XXPERMDI instruction
+      ///
+      XXPERMDI,
+
       /// The CMPB instruction (takes two operands of i32 or i64).
       CMPB,
 
       /// Hi/Lo - These represent the high and low 16-bit parts of a global
       /// address respectively.  These nodes have two operands, the first of
       /// which must be a TargetGlobalAddress, and the second of which must be a
       /// Constant.  Selected naively, these turn into 'lis G+C' and 'li G+C',
       /// though these are usually folded into other nodes.
namespace PPC {
     /// isVMRGEOShuffleMask - Return true if this is a shuffle mask suitable for
     /// a VMRGEW or VMRGOW instruction
     bool isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
                              unsigned ShuffleKind, SelectionDAG &DAG);
     /// isXXSLDWIShuffleMask - Return true if this is a shuffle mask suitable
     /// for a XXSLDWI instruction.
     bool isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
                               bool &Swap, bool IsLE);
+    /// isXXPERMDIShuffleMask - Return true if this is a shuffle mask suitable
+    /// for a XXPERMDI instruction.
+    bool isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
+                              bool &Swap, bool IsLE);
 
     /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the
     /// shift amount, otherwise return -1.
     int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
                             SelectionDAG &DAG);
 
     /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
     /// specifies a splat of a single element that is suitable for input to

llvm/trunk/lib/Target/PowerPC/PPCISelLowering.cpp

const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case PPCISD::FRE:             return "PPCISD::FRE";
   case PPCISD::FRSQRTE:         return "PPCISD::FRSQRTE";
   case PPCISD::STFIWX:          return "PPCISD::STFIWX";
   case PPCISD::VMADDFP:         return "PPCISD::VMADDFP";
   case PPCISD::VNMSUBFP:        return "PPCISD::VNMSUBFP";
   case PPCISD::VPERM:           return "PPCISD::VPERM";
   case PPCISD::XXSPLT:          return "PPCISD::XXSPLT";
   case PPCISD::XXINSERT:        return "PPCISD::XXINSERT";
+  case PPCISD::XXPERMDI:        return "PPCISD::XXPERMDI";
   case PPCISD::VECSHL:          return "PPCISD::VECSHL";
   case PPCISD::CMPB:            return "PPCISD::CMPB";
   case PPCISD::Hi:              return "PPCISD::Hi";
   case PPCISD::Lo:              return "PPCISD::Lo";
   case PPCISD::TOC_ENTRY:       return "PPCISD::TOC_ENTRY";
   case PPCISD::DYNALLOC:        return "PPCISD::DYNALLOC";
   case PPCISD::DYNAREAOFFSET:   return "PPCISD::DYNAREAOFFSET";
   case PPCISD::GlobalBaseReg:   return "PPCISD::GlobalBaseReg";
for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
     if (N->getMaskElt(i) < 0) continue;
     for (unsigned j = 0; j != EltSize; ++j)
       if (N->getMaskElt(i+j) != N->getMaskElt(j))
         return false;
   }
   return true;
 }
 
-  // Check that the mask is shuffling words
-static bool isWordShuffleMask(ShuffleVectorSDNode *N) {
-  for (unsigned i = 0; i < 4; ++i) {
-    unsigned B0 = N->getMaskElt(i*4);
-    unsigned B1 = N->getMaskElt(i*4+1);
-    unsigned B2 = N->getMaskElt(i*4+2);
-    unsigned B3 = N->getMaskElt(i*4+3);
-    if (B0 % 4)
+// Check that the mask is shuffling N byte elements.
+static bool isNByteElemShuffleMask(ShuffleVectorSDNode *N, unsigned Width) {
+  assert((Width == 2 || Width == 4 || Width == 8 || Width == 16) &&
+         "Unexpected element width.");
+
+  unsigned NumOfElem = 16 / Width;
+  unsigned MaskVal[16]; //  Width is never greater than 16
+  for (unsigned i = 0; i < NumOfElem; ++i) {
+    MaskVal[0] = N->getMaskElt(i * Width);
+    if (MaskVal[0] % Width) {
       return false;
-    if (B1 != B0+1 || B2 != B1+1 || B3 != B2+1)
+    }
+
+    for (unsigned int j = 1; j < Width; ++j) {
+      MaskVal[j] = N->getMaskElt(i * Width + j);
+      if (MaskVal[j] != MaskVal[j-1] + 1) {
         return false;
       }
+    }
+  }
 
   return true;
 }
 
 bool PPC::isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
                           unsigned &InsertAtByte, bool &Swap, bool IsLE) {
-  if (!isWordShuffleMask(N))
+  if (!isNByteElemShuffleMask(N, 4))
     return false;
 
   // Now we look at mask elements 0,4,8,12
   unsigned M0 = N->getMaskElt(0) / 4;
   unsigned M1 = N->getMaskElt(4) / 4;
   unsigned M2 = N->getMaskElt(8) / 4;
   unsigned M3 = N->getMaskElt(12) / 4;
   unsigned LittleEndianShifts[] = { 2, 1, 0, 3 };
bool PPC::isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
 
   return false;
 }
 
 bool PPC::isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
                                bool &Swap, bool IsLE) {
   assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8");
   // Ensure each byte index of the word is consecutive.
-  if (!isWordShuffleMask(N))
+  if (!isNByteElemShuffleMask(N, 4))
     return false;
 
   // Now we look at mask elements 0,4,8,12, which are the beginning of words.
   unsigned M0 = N->getMaskElt(0) / 4;
   unsigned M1 = N->getMaskElt(4) / 4;
   unsigned M2 = N->getMaskElt(8) / 4;
   unsigned M3 = N->getMaskElt(12) / 4;
 
if (M0 == 0 || M0 == 1 || M0 == 2 || M0 == 3) {
       Swap = true;
       ShiftElts = M0 - 4;
     }
 
     return true;
   }
 }
 
+/// Can node \p N be lowered to an XXPERMDI instruction? If so, set \p Swap
+/// if the inputs to the instruction should be swapped and set \p DM to the
+/// value for the immediate.
+/// Specifically, set \p Swap to true only if \p N can be lowered to XXPERMDI
+/// AND element 0 of the result comes from the first input (LE) or second input
+/// (BE). Set \p DM to the calculated result (0-3) only if \p N can be lowered.
+/// \return true iff the given mask of shuffle node \p N is a XXPERMDI shuffle
+/// mask.
+bool PPC::isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &DM,
+                               bool &Swap, bool IsLE) {
+  assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8");
+
+  // Ensure each byte index of the double word is consecutive.
+  if (!isNByteElemShuffleMask(N, 8))
+    return false;
+
+  unsigned M0 = N->getMaskElt(0) / 8;
+  unsigned M1 = N->getMaskElt(8) / 8;
+  assert(((M0 | M1) < 4) && "A mask element out of bounds?");
+
+  // If both vector operands for the shuffle are the same vector, the mask will
+  // contain only elements from the first one and the second one will be undef.
+  if (N->getOperand(1).isUndef()) {
+    if ((M0 | M1) < 2) {
+      DM = IsLE ? (((~M1) & 1) << 1) + ((~M0) & 1) : (M0 << 1) + (M1 & 1);
+      Swap = false;
+      return true;
+    } else
+      return false;
+  }
+
+  if (IsLE) {
+    if (M0 > 1 && M1 < 2) {
+      Swap = false;
+    } else if (M0 < 2 && M1 > 1) {
+      M0 = (M0 + 2) % 4;
+      M1 = (M1 + 2) % 4;
+      Swap = true;
+    } else
+      return false;
+
+    // Note: if control flow comes here that means Swap is already set above
+    DM = (((~M1) & 1) << 1) + ((~M0) & 1);
+    return true;
+  } else { // BE
+    if (M0 < 2 && M1 > 1) {
+      Swap = false;
+    } else if (M0 > 1 && M1 < 2) {
+      M0 = (M0 + 2) % 4;
+      M1 = (M1 + 2) % 4;
+      Swap = true;
+    } else
+      return false;
+
+    // Note: if control flow comes here that means Swap is already set above
+    DM = (M0 << 1) + (M1 & 1);
+    return true;
+  }
+}
+
 
 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize,
                                 SelectionDAG &DAG) {
   ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
   assert(isSplatShuffleMask(SVOp, EltSize));
   if (DAG.getDataLayout().isLittleEndian())
if (Subtarget.hasVSX() &&
     SDValue Conv2 =
         DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V2.isUndef() ? V1 : V2);
 
     SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v4i32, Conv1, Conv2,
                               DAG.getConstant(ShiftElts, dl, MVT::i32));
     return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Shl);
   }
 
+  if (Subtarget.hasVSX() &&
+    PPC::isXXPERMDIShuffleMask(SVOp, ShiftElts, Swap, isLittleEndian)) {
+    if (Swap)
+      std::swap(V1, V2);
+    SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1);
+    SDValue Conv2 =
+        DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2.isUndef() ? V1 : V2);
+
+    SDValue PermDI = DAG.getNode(PPCISD::XXPERMDI, dl, MVT::v2i64, Conv1, Conv2,
+                              DAG.getConstant(ShiftElts, dl, MVT::i32));
+    return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, PermDI);
+  }
+
   if (Subtarget.hasVSX()) {
     if (V2.isUndef() && PPC::isSplatShuffleMask(SVOp, 4)) {
       int SplatIdx = PPC::getVSPLTImmediate(SVOp, 4, DAG);
 
       // If the source for the shuffle is a scalar_to_vector that came from a
       // 32-bit load, it will have used LXVWSX so we don't need to splat again.
       if (Subtarget.hasP9Vector() &&
           ((isLittleEndian && SplatIdx == 3) ||

llvm/trunk/lib/Target/PowerPC/PPCInstrInfo.td

 def SDT_PPCVecShift : SDTypeProfile<1, 3, [ SDTCisVec<0>,
   SDTCisVec<1>, SDTCisVec<2>, SDTCisPtrTy<3>
 ]>;
 
 def SDT_PPCVecInsert : SDTypeProfile<1, 3, [ SDTCisVec<0>,
   SDTCisVec<1>, SDTCisVec<2>, SDTCisInt<3>
 ]>;
 
+def SDT_PPCxxpermdi: SDTypeProfile<1, 3, [ SDTCisVec<0>,
+  SDTCisVec<1>, SDTCisVec<2>, SDTCisInt<3>
+]>;
+
 def SDT_PPCvcmp : SDTypeProfile<1, 3, [
   SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>, SDTCisVT<3, i32>
 ]>;
 
 def SDT_PPCcondbr : SDTypeProfile<0, 3, [
   SDTCisVT<0, i32>, SDTCisVT<2, OtherVT>
 ]>;
 
def PPCaddiTlsldLAddr : SDNode<"PPCISD::ADDI_TLSLD_L_ADDR",
                                  SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>,
                                  SDTCisSameAs<0, 3>, SDTCisInt<0> ]>>;
 def PPCaddisDtprelHA : SDNode<"PPCISD::ADDIS_DTPREL_HA", SDTIntBinOp>;
 def PPCaddiDtprelL   : SDNode<"PPCISD::ADDI_DTPREL_L", SDTIntBinOp>;
 
 def PPCvperm     : SDNode<"PPCISD::VPERM", SDT_PPCvperm, []>;
 def PPCxxsplt    : SDNode<"PPCISD::XXSPLT", SDT_PPCVecSplat, []>;
 def PPCxxinsert  : SDNode<"PPCISD::XXINSERT", SDT_PPCVecInsert, []>;
+def PPCxxpermdi  : SDNode<"PPCISD::XXPERMDI", SDT_PPCxxpermdi, []>;
 def PPCvecshl    : SDNode<"PPCISD::VECSHL", SDT_PPCVecShift, []>;
 
 def PPCqvfperm   : SDNode<"PPCISD::QVFPERM", SDT_PPCqvfperm, []>;
 def PPCqvgpci    : SDNode<"PPCISD::QVGPCI", SDT_PPCqvgpci, []>;
 def PPCqvaligni  : SDNode<"PPCISD::QVALIGNI", SDT_PPCqvaligni, []>;
 def PPCqvesplati : SDNode<"PPCISD::QVESPLATI", SDT_PPCqvesplati, []>;
 
 def PPCqbflt     : SDNode<"PPCISD::QBFLT", SDT_PPCqbflt, []>;

llvm/trunk/lib/Target/PowerPC/PPCInstrVSX.td

def XXMRGHW : XX3Form<60, 18,
                        (outs vsrc:$XT), (ins vsrc:$XA, vsrc:$XB),
                        "xxmrghw $XT, $XA, $XB", IIC_VecPerm, []>;
   def XXMRGLW : XX3Form<60, 50,
                        (outs vsrc:$XT), (ins vsrc:$XA, vsrc:$XB),
                        "xxmrglw $XT, $XA, $XB", IIC_VecPerm, []>;
 
   def XXPERMDI : XX3Form_2<60, 10,
                        (outs vsrc:$XT), (ins vsrc:$XA, vsrc:$XB, u2imm:$DM),
-                       "xxpermdi $XT, $XA, $XB, $DM", IIC_VecPerm, []>;
+                       "xxpermdi $XT, $XA, $XB, $DM", IIC_VecPerm,
+                       [(set v2i64:$XT, (PPCxxpermdi v2i64:$XA, v2i64:$XB,
+                         imm32SExt16:$DM))]>;
   let isCodeGenOnly = 1 in
   def XXPERMDIs : XX3Form_2s<60, 10, (outs vsrc:$XT), (ins vsfrc:$XA, u2imm:$DM),
                              "xxpermdi $XT, $XA, $XA, $DM", IIC_VecPerm, []>;
   def XXSEL : XX4Form<60, 3,
                       (outs vsrc:$XT), (ins vsrc:$XA, vsrc:$XB, vsrc:$XC),
                       "xxsel $XT, $XA, $XB, $XC", IIC_VecPerm, []>;
 
   def XXSLDWI : XX3Form_2<60, 2,

llvm/trunk/test/CodeGen/PowerPC/vec_xxpermdi.ll

+; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr8 < %s | \
+; RUN:   FileCheck %s  -check-prefix=CHECK-LE
+; RUN: llc -verify-machineinstrs -mtriple=powerpc64-unknown-linux-gnu -mcpu=pwr8 < %s | \
+; RUN:   FileCheck %s -check-prefix=CHECK-BE
+
+; Possible LE ShuffleVector masks (Case 1):
+; ShuffleVector((vector double)a, (vector double)b, 3, 1)
+; ShuffleVector((vector double)a, (vector double)b, 2, 1)
+; ShuffleVector((vector double)a, (vector double)b, 3, 0)
+; ShuffleVector((vector double)a, (vector double)b, 2, 0)
+; which targets at:
+; xxpermdi a, b, 0
+; xxpermdi a, b, 1
+; xxpermdi a, b, 2
+; xxpermdi a, b, 3
+; Possible LE Swap ShuffleVector masks (Case 2):
+; ShuffleVector((vector double)a, (vector double)b, 1, 3)
+; ShuffleVector((vector double)a, (vector double)b, 0, 3)
+; ShuffleVector((vector double)a, (vector double)b, 1, 2)
+; ShuffleVector((vector double)a, (vector double)b, 0, 2)
+; which targets at:
+; xxpermdi b, a, 0
+; xxpermdi b, a, 1
+; xxpermdi b, a, 2
+; xxpermdi b, a, 3
+; Possible LE ShuffleVector masks when a == b, b is undef (Case 3):
+; ShuffleVector((vector double)a, (vector double)a, 1, 1)
+; ShuffleVector((vector double)a, (vector double)a, 0, 1)
+; ShuffleVector((vector double)a, (vector double)a, 1, 0)
+; ShuffleVector((vector double)a, (vector double)a, 0, 0)
+; which targets at:
+; xxpermdi a, a, 0
+; xxpermdi a, a, 1
+; xxpermdi a, a, 2
+; xxpermdi a, a, 3
+
+; Possible BE ShuffleVector masks (Case 4):
+; ShuffleVector((vector double)a, (vector double)b, 0, 2)
+; ShuffleVector((vector double)a, (vector double)b, 0, 3)
+; ShuffleVector((vector double)a, (vector double)b, 1, 2)
+; ShuffleVector((vector double)a, (vector double)b, 1, 3)
+; which targets at:
+; xxpermdi a, b, 0
+; xxpermdi a, b, 1
+; xxpermdi a, b, 2
+; xxpermdi a, b, 3
+; Possible BE Swap ShuffleVector masks (Case 5):
+; ShuffleVector((vector double)a, (vector double)b, 2, 0)
+; ShuffleVector((vector double)a, (vector double)b, 3, 0)
+; ShuffleVector((vector double)a, (vector double)b, 2, 1)
+; ShuffleVector((vector double)a, (vector double)b, 3, 1)
+; which targets at:
+; xxpermdi b, a, 0
+; xxpermdi b, a, 1
+; xxpermdi b, a, 2
+; xxpermdi b, a, 3
+; Possible BE ShuffleVector masks when a == b, b is undef (Case 6):
+; ShuffleVector((vector double)a, (vector double)a, 0, 0)
+; ShuffleVector((vector double)a, (vector double)a, 0, 1)
+; ShuffleVector((vector double)a, (vector double)a, 1, 0)
+; ShuffleVector((vector double)a, (vector double)a, 1, 1)
+; which targets at:
+; xxpermdi a, a, 0
+; xxpermdi a, a, 1
+; xxpermdi a, a, 2
+; xxpermdi a, a, 3
+
+define <2 x double> @test_le_vec_xxpermdi_v2f64_v2f64_0(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 3, i32 1>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_vec_xxpermdi_v2f64_v2f64_0
+; CHECK-LE: xxmrghd 34, 34, 35
+; CHECK-LE: blr
+}
+
+define <2 x double> @test_le_vec_xxpermdi_v2f64_v2f64_1(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 2, i32 1>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_vec_xxpermdi_v2f64_v2f64_1
+; CHECK-LE: xxpermdi 34, 34, 35, 1
+; CHECK-LE: blr
+}
+
+define <2 x double> @test_le_vec_xxpermdi_v2f64_v2f64_2(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 3, i32 0>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_vec_xxpermdi_v2f64_v2f64_2
+; CHECK-LE: xxpermdi 34, 34, 35, 2
+; CHECK-LE: blr
+}
+
+define <2 x double> @test_le_vec_xxpermdi_v2f64_v2f64_3(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 2, i32 0>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_vec_xxpermdi_v2f64_v2f64_3
+; CHECK-LE: xxmrgld 34, 34, 35
+; CHECK-LE: blr
+}
+
+define <2 x double> @test_le_swap_vec_xxpermdi_v2f64_v2f64_0(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 1, i32 3>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_swap_vec_xxpermdi_v2f64_v2f64_0
+; CHECK-LE: xxmrghd 34, 35, 34
+; CHECK-LE: blr
+}
+
+define <2 x double> @test_le_swap_vec_xxpermdi_v2f64_v2f64_1(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 0, i32 3>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_swap_vec_xxpermdi_v2f64_v2f64_1
+; CHECK-LE: xxpermdi 34, 35, 34, 1
+; CHECK-LE: blr
+}
+
+define <2 x double> @test_le_swap_vec_xxpermdi_v2f64_v2f64_2(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 1, i32 2>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_swap_vec_xxpermdi_v2f64_v2f64_2
+; CHECK-LE: xxpermdi 34, 35, 34, 2
+; CHECK-LE: blr
+}
+
+define <2 x double> @test_le_swap_vec_xxpermdi_v2f64_v2f64_3(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 0, i32 2>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_swap_vec_xxpermdi_v2f64_v2f64_3
+; CHECK-LE: xxmrgld 34, 35, 34
+; CHECK-LE: blr
+}
+
+define <2 x double> @test_le_vec_xxpermdi_v2f64_undef_0(<2 x double> %VA) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> undef, <2 x i32> <i32 1, i32 1>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_vec_xxpermdi_v2f64_undef_0
+; CHECK-LE: xxspltd 34, 34, 0
+; CHECK-LE: blr
+}
+
+define <2 x double> @test_le_vec_xxpermdi_v2f64_undef_1(<2 x double> %VA) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> undef, <2 x i32> <i32 0, i32 1>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_vec_xxpermdi_v2f64_undef_1
+; CHECK-LE: blr
+}
+
+define <2 x double> @test_le_vec_xxpermdi_v2f64_undef_2(<2 x double> %VA) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> undef, <2 x i32> <i32 1, i32 0>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_vec_xxpermdi_v2f64_undef_2
+; CHCECK-LE: xxswapd 34, 34
+}
+
+define <2 x double> @test_le_vec_xxpermdi_v2f64_undef_3(<2 x double> %VA) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> undef, <2 x i32> <i32 0, i32 0>
+      ret <2 x double> %0
+; CHECK-LE-LABEL: @test_le_vec_xxpermdi_v2f64_undef_3
+; CHECK-LE: xxspltd 34, 34, 1
+; CHECK-LE: blr
+}
+
+; Start testing BE
+define <2 x double> @test_be_vec_xxpermdi_v2f64_v2f64_0(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 0, i32 2>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_vec_xxpermdi_v2f64_v2f64_0
+; CHECK-BE: xxmrghd 34, 34, 35
+; CHECK-BE: blr
+}
+
+define <2 x double> @test_be_vec_xxpermdi_v2f64_v2f64_1(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 0, i32 3>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_vec_xxpermdi_v2f64_v2f64_1
+; CHECK-BE: xxpermdi 34, 34, 35, 1
+; CHECK-BE: blr
+}
+
+define <2 x double> @test_be_vec_xxpermdi_v2f64_v2f64_2(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 1, i32 2>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_vec_xxpermdi_v2f64_v2f64_2
+; CHECK-BE: xxpermdi 34, 34, 35, 2
+; CHECK-BE: blr
+}
+
+define <2 x double> @test_be_vec_xxpermdi_v2f64_v2f64_3(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 1, i32 3>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_vec_xxpermdi_v2f64_v2f64_3
+; CHECK-BE: xxmrgld 34, 34, 35
+; CHECK-BE: blr
+}
+
+define <2 x double> @test_be_swap_vec_xxpermdi_v2f64_v2f64_0(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 2, i32 0>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_swap_vec_xxpermdi_v2f64_v2f64_0
+; CHECK-BE: xxmrghd 34, 35, 34
+; CHECK-BE: blr
+}
+
+define <2 x double> @test_be_swap_vec_xxpermdi_v2f64_v2f64_1(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 2, i32 1>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_swap_vec_xxpermdi_v2f64_v2f64_1
+; CHECK-BE: xxpermdi 34, 35, 34, 1
+; CHECK-BE: blr
+}
+
+define <2 x double> @test_be_swap_vec_xxpermdi_v2f64_v2f64_2(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 3, i32 0>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_swap_vec_xxpermdi_v2f64_v2f64_2
+; CHECK-BE: xxpermdi 34, 35, 34, 2
+; CHECK-BE: blr
+}
+
+define <2 x double> @test_be_swap_vec_xxpermdi_v2f64_v2f64_3(<2 x double> %VA, <2 x double> %VB) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> %VB,<2 x i32> <i32 3, i32 1>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_swap_vec_xxpermdi_v2f64_v2f64_3
+; CHECK-BE: xxmrgld 34, 35, 34
+; CHECK-BE: blr
+}
+
+define <2 x double> @test_be_vec_xxpermdi_v2f64_undef_0(<2 x double> %VA) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> undef, <2 x i32> <i32 0, i32 0>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_vec_xxpermdi_v2f64_undef_0
+; CHECK-BE: xxspltd 34, 34, 0
+; CHECK-BE: blr
+}
+
+define <2 x double> @test_be_vec_xxpermdi_v2f64_undef_1(<2 x double> %VA) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> undef, <2 x i32> <i32 0, i32 1>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_vec_xxpermdi_v2f64_undef_1
+; CHECK-BE: blr
+}
+
+define <2 x double> @test_be_vec_xxpermdi_v2f64_undef_2(<2 x double> %VA) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> undef, <2 x i32> <i32 1, i32 0>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_vec_xxpermdi_v2f64_undef_2
+; CHCECK-LE: xxswapd 34, 34
+}
+
+define <2 x double> @test_be_vec_xxpermdi_v2f64_undef_3(<2 x double> %VA) {
+     entry:
+      %0 = shufflevector <2 x double> %VA, <2 x double> undef, <2 x i32> <i32 1, i32 1>
+      ret <2 x double> %0
+; CHECK-BE-LABEL: @test_be_vec_xxpermdi_v2f64_undef_3
+; CHECK-BE: xxspltd 34, 34, 1
+; CHECK-BE: blr
+}
+
+; More test cases to test different types of vector inputs
+define <16 x i8> @test_be_vec_xxpermdi_v16i8_v16i8(<16 x i8> %VA, <16 x i8> %VB) {
+     entry:
+      %0 = shufflevector <16 x i8> %VA, <16 x i8> %VB,<16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19>
+      ret <16 x i8> %0
+; CHECK-BE-LABEL: @test_be_vec_xxpermdi_v16i8_v16i8
+; CHECK-BE: xxpermdi 34, 34, 35, 1
+; CHECK-BE: blr
+}
+
+define <8 x i16> @test_le_swap_vec_xxpermdi_v8i16_v8i16(<8 x i16> %VA, <8 x i16> %VB) {
+     entry:
+      %0 = shufflevector <8 x i16> %VA, <8 x i16> %VB,<8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 12, i32 13, i32 14, i32 15>
+      ret <8 x i16> %0
+; CHECK-LE-LABEL: @test_le_swap_vec_xxpermdi_v8i16_v8i16
+; CHECK-LE: xxpermdi 34, 35, 34, 1
+; CHECK-LE: blr
+}
+
+define <4 x i32> @test_le_swap_vec_xxpermdi_v4i32_v4i32(<4 x i32> %VA, <4 x i32> %VB) {
+     entry:
+      %0 = shufflevector <4 x i32> %VA, <4 x i32> %VB,<4 x i32> <i32 0, i32 1, i32 6, i32 7>
+      ret <4 x i32> %0
+; CHECK-LE-LABEL: @test_le_swap_vec_xxpermdi_v4i32_v4i32
+; CHECK-LE: xxpermdi 34, 35, 34, 1
+; CHECK-LE: blr
+}