[llvm-commits] VPBLENDD match in AVX2 - please review

[Nadav Rotem]

LGTM. Please commit.  I have two small comments:

+  //There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
+  unsigned NumLanes = (NumElems-1)/8 + 1; 
+  unsigned NumElemsInLane = NumElems / NumLanes;

Space before the "There".

+  // Convert i32 vectors to floating point if it is not AVX2.
+  // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
+  EVT BlendVT = VT;
+  if (!Subtarget->hasInt256() && (EltVT == MVT::i32 || EltVT == MVT::i64)) {
+    BlendVT = EVT::getVectorVT(*DAG.getContext(), 
+                              EVT::getFloatingPointVT(EltVT.getSizeInBits()), 
+                              NumElems);
+    V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
+    V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
+  }
+  
+  SDValue Ret =  DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
+                             DAG.getConstant(MaskValue, MVT::i32));

I think that this code is redundant.  Why not convert the blend operands to floats in all cases ?  If they are already floats then the SelectionDAG getNode method will not create an extra bit cast. 



On Dec 4, 2012, at 2:05 AM, "Demikhovsky, Elena" <elena.demikhovsky at intel.com> wrote:

> <blend.diff>