promote extract_vector_elt nodes to unpacking mov

Also update the test cases.

Author: Prince781

llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp

@@ -464,10 +464,14 @@ bool NVPTXDAGToDAGISel::tryUNPACK_VECTOR(SDNode *N) {
 bool NVPTXDAGToDAGISel::tryEXTRACT_VECTOR_ELEMENT(SDNode *N) {
   SDValue Vector = N->getOperand(0);
 
-  // We only care about 16x2 as it's the only real vector type we
-  // need to deal with.
+  // We only care about packed vector types: 16x2 and 32x2.
   MVT VT = Vector.getSimpleValueType();
-  if (!Isv2x16VT(VT))
+  unsigned NewOpcode = 0;
+  if (Isv2x16VT(VT))
+    NewOpcode = NVPTX::I32toV2I16;
+  else if (VT == MVT::v2f32)
+    NewOpcode = NVPTX::I64toV2F32;
+  else
     return false;
   // Find and record all uses of this vector that extract element 0 or 1.
   SmallVector<SDNode *, 4> E0, E1;
@@ -487,16 +491,19 @@ bool NVPTXDAGToDAGISel::tryEXTRACT_VECTOR_ELEMENT(SDNode *N) {
     }
   }
 
-  // There's no point scattering f16x2 if we only ever access one
+  // There's no point scattering f16x2 or f32x2 if we only ever access one
   // element of it.
   if (E0.empty() || E1.empty())
     return false;
 
-  // Merge (f16 extractelt(V, 0), f16 extractelt(V,1))
-  // into f16,f16 SplitF16x2(V)
+  // Merge:
+  //  (f16 extractelt(V, 0), f16 extractelt(V,1))
+  //  -> f16,f16 SplitF16x2(V)
+  //  (f32 extractelt(V, 0), f32 extractelt(V,1))
+  //  -> f32,f32 SplitF32x2(V)
   MVT EltVT = VT.getVectorElementType();
   SDNode *ScatterOp =
-      CurDAG->getMachineNode(NVPTX::I32toV2I16, SDLoc(N), EltVT, EltVT, Vector);
+      CurDAG->getMachineNode(NewOpcode, SDLoc(N), EltVT, EltVT, Vector);
   for (auto *Node : E0)
     ReplaceUses(SDValue(Node, 0), SDValue(ScatterOp, 0));
   for (auto *Node : E1)

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

@@ -5545,10 +5545,10 @@ static SDValue PerformEXTRACTCombine(SDNode *N,
       IsPTXVectorType(VectorVT.getSimpleVT()))
     return SDValue(); // Native vector loads already combine nicely w/
                       // extract_vector_elt.
-  // Don't mess with singletons or v2*16, v4i8 and v8i8 types, we already
+  // Don't mess with singletons or v2*16, v2f32, v4i8 and v8i8 types, we already
   // handle them OK.
   if (VectorVT.getVectorNumElements() == 1 || Isv2x16VT(VectorVT) ||
-      VectorVT == MVT::v4i8 || VectorVT == MVT::v8i8)
+      VectorVT == MVT::v2f32 || VectorVT == MVT::v4i8 || VectorVT == MVT::v8i8)
     return SDValue();
 
   // Don't mess with undef values as sra may be simplified to 0, not undef.

llvm/lib/Target/NVPTX/NVPTXInstrInfo.td

@@ -3028,6 +3028,9 @@ let hasSideEffects = false in {
   def I64toV2I32 : NVPTXInst<(outs Int32Regs:$d1, Int32Regs:$d2),
                              (ins Int64Regs:$s),
                              "mov.b64 \t{{$d1, $d2}}, $s;", []>;
+  def I64toV2F32 : NVPTXInst<(outs Float32Regs:$d1, Float32Regs:$d2),
+                             (ins Int64Regs:$s),
+                             "mov.b64 \t{{$d1, $d2}}, $s;", []>;
   def I128toV2I64: NVPTXInst<(outs Int64Regs:$d1, Int64Regs:$d2),
                               (ins Int128Regs:$s),
                               "mov.b128 \t{{$d1, $d2}}, $s;", []>;