Simplified TRUNCATE operation that comes after SETCC. It is possible since SETCC result is 0 or -1.

Added a test.

llvm-svn: 171467

2 files changed, 35 insertions, 3 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index eca63f80ae0..9a553d61bfb 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -14661,12 +14661,29 @@ static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
   return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
 }
 
-/// PerformTruncateCombine - Converts truncate operation to
-/// a sequence of vector shuffle operations.
-/// It is possible when we truncate 256-bit vector to 128-bit vector
+/// PerformTruncateCombine - In some cases a sequence with "truncate"
+/// operation may be simplified.
 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
                                       TargetLowering::DAGCombinerInfo &DCI,
                                       const X86Subtarget *Subtarget)  {
+  EVT VT = N->getValueType(0);
+  if (DCI.isBeforeLegalize() || !VT.isVector())
+    return SDValue();
+
+  SDValue In = N->getOperand(0);
+  // Optimize the sequence setcc -> truncate
+  if (In.getOpcode() == ISD::SETCC) {
+    DebugLoc DL = N->getDebugLoc();
+    EVT InVT = In.getValueType();
+
+    // The vector element is all ones or all zero. Just take a half of it.
+    EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
+                                  InVT.getVectorNumElements()/2);
+    SDValue HalfVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, In,
+                                  DAG.getIntPtrConstant(0));
+    assert(HalfVT.getSizeInBits() == VT.getSizeInBits());
+    return DAG.getNode(ISD::BITCAST, DL, VT, HalfVec);
+  }
   return SDValue();
 }
 
diff --git a/llvm/test/CodeGen/X86/avx-trunc.ll b/llvm/test/CodeGen/X86/avx-trunc.ll
index d0077366444..aa186a05f21 100755
--- a/llvm/test/CodeGen/X86/avx-trunc.ll
+++ b/llvm/test/CodeGen/X86/avx-trunc.ll
@@ -13,3 +13,18 @@ define <8 x i16> @trunc_32_16(<8 x i32> %A) nounwind uwtable readnone ssp{
   ret <8 x i16>%B
 }
 
+define <8 x i16> @trunc_after_setcc(<8 x float> %a, <8 x float> %b, <8 x float> %c, <8 x float> %d) {
+; CHECK: trunc_after_setcc
+; CHECK: vcmpltps
+; CHECK-NOT: vextract
+; CHECK: vcmpltps
+; CHECK-NEXT: vandps
+; CHECK-NEXT: vandps
+; CHECK: ret
+  %res1 = fcmp olt <8 x float> %a, %b
+  %res2 = fcmp olt <8 x float> %c, %d
+  %andr = and <8 x i1>%res1, %res2
+  %ex = zext <8 x i1> %andr to <8 x i16>
+  ret <8 x i16>%ex
+}
+