Fix PR10688. Add support for spliting 256-bit vector shifts when the

shift amount is variable

llvm-svn: 137885

2 files changed, 31 insertions, 11 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index b330a4f9809..d1c71a86a08 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -9449,17 +9449,26 @@ SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
                                      DAG, dl);
 
     // Recreate the shift amount vectors
-    SmallVector<SDValue, 4> Amt1Csts;
-    SmallVector<SDValue, 4> Amt2Csts;
-    for (int i = 0; i < NumElems/2; ++i)
-      Amt1Csts.push_back(Amt->getOperand(i));
-    for (int i = NumElems/2; i < NumElems; ++i)
-      Amt2Csts.push_back(Amt->getOperand(i));
-
-    SDValue Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
-                               &Amt1Csts[0], NumElems/2);
-    SDValue Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
-                               &Amt2Csts[0], NumElems/2);
+    SDValue Amt1, Amt2;
+    if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
+      // Constant shift amount
+      SmallVector<SDValue, 4> Amt1Csts;
+      SmallVector<SDValue, 4> Amt2Csts;
+      for (int i = 0; i < NumElems/2; ++i)
+        Amt1Csts.push_back(Amt->getOperand(i));
+      for (int i = NumElems/2; i < NumElems; ++i)
+        Amt2Csts.push_back(Amt->getOperand(i));
+
+      Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
+                                 &Amt1Csts[0], NumElems/2);
+      Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
+                                 &Amt2Csts[0], NumElems/2);
+    } else {
+      // Variable shift amount
+      Amt1 = Extract128BitVector(Amt, DAG.getConstant(0, MVT::i32), DAG, dl);
+      Amt2 = Extract128BitVector(Amt, DAG.getConstant(NumElems/2, MVT::i32),
+                                 DAG, dl);
+    }
 
     // Issue new vector shifts for the smaller types
     V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
diff --git a/llvm/test/CodeGen/X86/avx-shift.ll b/llvm/test/CodeGen/X86/avx-shift.ll
index 791194fc1c7..3ea39a2358e 100644
--- a/llvm/test/CodeGen/X86/avx-shift.ll
+++ b/llvm/test/CodeGen/X86/avx-shift.ll
@@ -62,3 +62,14 @@ define <16 x i16> @vshift07(<16 x i16> %a) nounwind readnone {
   ret <16 x i16> %s
 }
 
+;;; Support variable shifts
+; CHECK: _vshift08
+; CHECK: vextractf128 $1
+; CHECK: vpslld $23
+; CHECK: vextractf128 $1
+; CHECK: vpslld $23
+define <8 x i32> @vshift08(<8 x i32> %a) nounwind {
+  %bitop = shl <8 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>, %a
+  ret <8 x i32> %bitop
+}
+