[X86] Add support for folding (insert_subvector vec1, (extract_subvector vec2, idx1), idx1) -> (blendi vec2, vec1).

llvm-svn: 294112

1 files changed, 25 insertions, 2 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index f51920f5e6e..2090d32d08c 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -34121,8 +34121,7 @@ static SDValue combineInsertSubvector(SDNode *N, SelectionDAG &DAG,
   // We are still creating an INSERT_SUBVECTOR below with an undef node to
   // extend the subvector to the size of the result vector. Make sure that
   // we are not recursing on that node by checking for undef here.
-  if (IdxVal == 0 && OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
-      !Vec.isUndef()) {
+  if (IdxVal == 0 && OpVT.is256BitVector() && !Vec.isUndef()) {
     SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT,
                                  DAG.getUNDEF(OpVT), SubVec, N->getOperand(2));
 
@@ -34144,6 +34143,30 @@ static SDValue combineInsertSubvector(SDNode *N, SelectionDAG &DAG,
     return DAG.getBitcast(OpVT, Vec256);
   }
 
+  // If we're inserting into the upper half of a 256-bit vector with a vector
+  // that was extracted from the upper half of a 256-bit vector, we should
+  // use a blend instead.
+  if (SubVec.getOpcode() == ISD::EXTRACT_SUBVECTOR && OpVT.is256BitVector() &&
+      SubVec.getOperand(0).getSimpleValueType() == OpVT &&
+      Idx == SubVec.getOperand(1) && IdxVal == OpVT.getVectorNumElements()/2) {
+
+    // Integers must be cast to 32-bit because there is only vpblendd;
+    // vpblendw can't be used for this because it has a handicapped mask.
+    // If we don't have AVX2, then cast to float. Using a wrong domain blend
+    // is still more efficient than using the wrong domain vinsertf128 that
+    // will be created by InsertSubVector().
+    MVT CastVT = OpVT;
+    if (OpVT.isInteger())
+      CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
+
+    // The blend instruction, and therefore its mask, depend on the data type.
+    unsigned MaskVal = CastVT.getScalarSizeInBits() == 64 ? 0x0c : 0xf0;
+    SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
+    Vec = DAG.getNode(X86ISD::BLENDI, dl, CastVT, DAG.getBitcast(CastVT, Vec),
+                      DAG.getBitcast(CastVT, SubVec.getOperand(0)), Mask);
+    return DAG.getBitcast(OpVT, Vec);
+  }
+
   // Fold two 16-byte or 32-byte subvector loads into one 32-byte or 64-byte
   // load:
   // (insert_subvector (insert_subvector undef, (load16 addr), 0),