[X86] Change avx512 COMPRESS and EXPAND lowering to use a single masked node instead of expand/compress+select.

[craig.topper]

Summary

For compress, a select node doesn't semantically reflect the behavior of the instruction. The mask would have holes in it, but the resulting write is to contiguous elements at the bottom of the vector.

Furthermore, as far as the compressing and expanding is concerned the behavior is depended on the mask. You can't just have an expand/compress node that only reads the input vector. That node would have no meaning by itself.

This all only works because we pattern match the compress/expand+select back to the instruction. But conceivably an optimization of the select could break the pattern and leave something meaningless.

This patch modifies the expand and compress node to take the mask and passthru as additional inputs and gets rid of the select all together.

[RKSimon]

LGTM

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

static SDValue lowerVectorShuffleToEXPAND(const SDLoc &DL, MVT VT,
   SDValue MaskNode = DAG.getConstant(VEXPANDMask, DL, IntegerType);
   unsigned NumElts = VT.getVectorNumElements();
   assert((NumElts == 4 || NumElts == 8 || NumElts == 16) &&
          "Unexpected number of vector elements");
   SDValue VMask = getMaskNode(MaskNode, MVT::getVectorVT(MVT::i1, NumElts),
                               Subtarget, DAG, DL);
   SDValue ZeroVector = getZeroVector(VT, Subtarget, DAG, DL);
   SDValue ExpandedVector = IsLeftZeroSide ? V2 : V1;
-  return DAG.getSelect(DL, VT, VMask,
-                       DAG.getNode(X86ISD::EXPAND, DL, VT, ExpandedVector),
-                       ZeroVector);
+  return DAG.getNode(X86ISD::EXPAND, DL, VT, ExpandedVector, ZeroVector, VMask);
 }
 
 static bool matchVectorShuffleWithUNPCK(MVT VT, SDValue &V1, SDValue &V2,
                                         unsigned &UnpackOpcode, bool IsUnary,
                                         ArrayRef<int> TargetMask,
                                         const SDLoc &DL, SelectionDAG &DAG,
                                         const X86Subtarget &Subtarget) {
   int NumElts = VT.getVectorNumElements();
case VSHIFT:
                                  DAG);
     case COMPRESS_EXPAND_IN_REG: {
       SDValue Mask = Op.getOperand(3);
       SDValue DataToCompress = Op.getOperand(1);
       SDValue PassThru = Op.getOperand(2);
       if (isAllOnesConstant(Mask)) // return data as is
         return Op.getOperand(1);
 
-      return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
-                                              DataToCompress),
-                                  Mask, PassThru, Subtarget, DAG);
+      MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
+      Mask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
+
+      // Avoid false dependency.
+      if (PassThru.isUndef())
+        PassThru = DAG.getConstant(0, dl, VT);
+
+      return DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress, PassThru,
+                         Mask);
     }
     case FIXUPIMMS:
     case FIXUPIMMS_MASKZ:
     case FIXUPIMM:
     case FIXUPIMM_MASKZ:{
       SDValue Src1 = Op.getOperand(1);
       SDValue Src2 = Op.getOperand(2);
       SDValue Src3 = Op.getOperand(3);

llvm/trunk/lib/Target/X86/X86InstrAVX512.td

 //===----------------------------------------------------------------------===//
 // AVX-512 - COMPRESS and EXPAND
 //
 
 multiclass compress_by_vec_width_common<bits<8> opc, X86VectorVTInfo _,
                                  string OpcodeStr, X86FoldableSchedWrite sched> {
   defm rr : AVX512_maskable<opc, MRMDestReg, _, (outs _.RC:$dst),
               (ins _.RC:$src1), OpcodeStr, "$src1", "$src1",
-              (_.VT (X86compress _.RC:$src1))>, AVX5128IBase,
+              (null_frag)>, AVX5128IBase,
               Sched<[sched]>;
 
   let mayStore = 1, hasSideEffects = 0 in
   def mr : AVX5128I<opc, MRMDestMem, (outs),
               (ins _.MemOp:$dst, _.RC:$src),
               OpcodeStr # "\t{$src, $dst|$dst, $src}",
               []>, EVEX_CD8<_.EltSize, CD8VT1>,
               Sched<[sched.Folded]>;
 
   def mrk : AVX5128I<opc, MRMDestMem, (outs),
               (ins _.MemOp:$dst, _.KRCWM:$mask, _.RC:$src),
               OpcodeStr # "\t{$src, $dst {${mask}}|$dst {${mask}}, $src}",
               []>,
               EVEX_K, EVEX_CD8<_.EltSize, CD8VT1>,
               Sched<[sched.Folded]>;
 }
 
 multiclass compress_by_vec_width_lowering<X86VectorVTInfo _, string Name> {
   def : Pat<(X86mCompressingStore (_.VT _.RC:$src), addr:$dst, _.KRCWM:$mask),
             (!cast<Instruction>(Name#_.ZSuffix##mrk)
                             addr:$dst, _.KRCWM:$mask, _.RC:$src)>;
+
+  def : Pat<(X86compress (_.VT _.RC:$src), _.RC:$src0, _.KRCWM:$mask),
+            (!cast<Instruction>(Name#_.ZSuffix##rrk)
+                            _.RC:$src0, _.KRCWM:$mask, _.RC:$src)>;
+  def : Pat<(X86compress (_.VT _.RC:$src), _.ImmAllZerosV, _.KRCWM:$mask),
+            (!cast<Instruction>(Name#_.ZSuffix##rrkz)
+                            _.KRCWM:$mask, _.RC:$src)>;
 }
 
 multiclass compress_by_elt_width<bits<8> opc, string OpcodeStr,
                                  X86FoldableSchedWrite sched,
                                  AVX512VLVectorVTInfo VTInfo,
                                  Predicate Pred = HasAVX512> {
   let Predicates = [Pred] in
   defm Z : compress_by_vec_width_common<opc, VTInfo.info512, OpcodeStr, sched>,

 defm VCOMPRESSPD : compress_by_elt_width <0x8A, "vcompresspd", WriteVarShuffle256,
                                           avx512vl_f64_info>, EVEX, VEX_W, NotMemoryFoldable;
 
 // expand
 multiclass expand_by_vec_width<bits<8> opc, X86VectorVTInfo _,
                                  string OpcodeStr, X86FoldableSchedWrite sched> {
   defm rr : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
               (ins _.RC:$src1), OpcodeStr, "$src1", "$src1",
-              (_.VT (X86expand _.RC:$src1))>, AVX5128IBase,
+              (null_frag)>, AVX5128IBase,
               Sched<[sched]>;
 
   defm rm : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
               (ins _.MemOp:$src1), OpcodeStr, "$src1", "$src1",
-              (_.VT (X86expand (_.VT (bitconvert
-                                      (_.LdFrag addr:$src1)))))>,
+              (null_frag)>,
             AVX5128IBase, EVEX_CD8<_.EltSize, CD8VT1>,
             Sched<[sched.Folded, sched.ReadAfterFold]>;
 }
 
 multiclass expand_by_vec_width_lowering<X86VectorVTInfo _, string Name> {
 
   def : Pat<(_.VT (X86mExpandingLoad addr:$src, _.KRCWM:$mask, undef)),
             (!cast<Instruction>(Name#_.ZSuffix##rmkz)
                                         _.KRCWM:$mask, addr:$src)>;
 
   def : Pat<(_.VT (X86mExpandingLoad addr:$src, _.KRCWM:$mask, _.ImmAllZerosV)),
             (!cast<Instruction>(Name#_.ZSuffix##rmkz)
                                         _.KRCWM:$mask, addr:$src)>;
 
   def : Pat<(_.VT (X86mExpandingLoad addr:$src, _.KRCWM:$mask,
                                                (_.VT _.RC:$src0))),
             (!cast<Instruction>(Name#_.ZSuffix##rmk)
                             _.RC:$src0, _.KRCWM:$mask, addr:$src)>;
+
+  def : Pat<(X86expand (_.VT _.RC:$src), _.RC:$src0, _.KRCWM:$mask),
+            (!cast<Instruction>(Name#_.ZSuffix##rrk)
+                            _.RC:$src0, _.KRCWM:$mask, _.RC:$src)>;
+  def : Pat<(X86expand (_.VT _.RC:$src), _.ImmAllZerosV, _.KRCWM:$mask),
+            (!cast<Instruction>(Name#_.ZSuffix##rrkz)
+                            _.KRCWM:$mask, _.RC:$src)>;
 }
 
 multiclass expand_by_elt_width<bits<8> opc, string OpcodeStr,
                                X86FoldableSchedWrite sched,
                                AVX512VLVectorVTInfo VTInfo,
                                Predicate Pred = HasAVX512> {
   let Predicates = [Pred] in
   defm Z : expand_by_vec_width<opc, VTInfo.info512, OpcodeStr, sched>,

llvm/trunk/lib/Target/X86/X86InstrFragmentsSIMD.td

 def X86RndScales : SDNode<"X86ISD::VRNDSCALES", SDTFPBinOpImm>;
 def X86Reduces   : SDNode<"X86ISD::VREDUCES",   SDTFPBinOpImm>;
 def X86GetMants  : SDNode<"X86ISD::VGETMANTS",  SDTFPBinOpImm>;
 def X86RangesRnd    : SDNode<"X86ISD::VRANGES_RND",    SDTFPBinOpImmRound>;
 def X86RndScalesRnd : SDNode<"X86ISD::VRNDSCALES_RND", SDTFPBinOpImmRound>;
 def X86ReducesRnd   : SDNode<"X86ISD::VREDUCES_RND",   SDTFPBinOpImmRound>;
 def X86GetMantsRnd  : SDNode<"X86ISD::VGETMANTS_RND",  SDTFPBinOpImmRound>;
 
-def X86compress: SDNode<"X86ISD::COMPRESS", SDTypeProfile<1, 1,
-                              [SDTCisSameAs<0, 1>, SDTCisVec<1>]>, []>;
-def X86expand  : SDNode<"X86ISD::EXPAND", SDTypeProfile<1, 1,
-                              [SDTCisSameAs<0, 1>, SDTCisVec<1>]>, []>;
+def X86compress: SDNode<"X86ISD::COMPRESS", SDTypeProfile<1, 3,
+                              [SDTCisSameAs<0, 1>, SDTCisVec<1>,
+                               SDTCisSameAs<0, 2>, SDTCVecEltisVT<3, i1>,
+                               SDTCisSameNumEltsAs<0, 3>]>, []>;
+def X86expand  : SDNode<"X86ISD::EXPAND", SDTypeProfile<1, 3,
+                              [SDTCisSameAs<0, 1>, SDTCisVec<1>,
+                               SDTCisSameAs<0, 2>, SDTCVecEltisVT<3, i1>,
+                               SDTCisSameNumEltsAs<0, 3>]>, []>;
 
 // vpshufbitqmb
 def X86Vpshufbitqmb : SDNode<"X86ISD::VPSHUFBITQMB",
                              SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
                                                   SDTCisSameAs<1,2>,
                                                   SDTCVecEltisVT<0,i1>,
                                                   SDTCisSameNumEltsAs<0,1>]>>;