Fix getNode to allow a vector for the shift amount for shifts of vectors.

Fix the shift amount when unrolling a vector shift into scalar shifts.
Fix problem in getShuffleScalarElt where it assumes that the input of
a bit convert must be a vector.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@60740 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 44 insertions, 9 deletions

diff --git a/docs/LangRef.html b/docs/LangRef.html
index 74027ce..48b619a 100644
--- a/docs/LangRef.html
+++ b/docs/LangRef.html
@@ -2591,6 +2591,7 @@ equal to or larger than the number of bits in <tt>op1</tt>, the result is undefi
   &lt;result&gt; = shl i32 4, 2      <i>; yields {i32}: 16</i>
   &lt;result&gt; = shl i32 1, 10     <i>; yields {i32}: 1024</i>
   &lt;result&gt; = shl i32 1, 32     <i>; undefined</i>
+  &lt;result&gt; = shl &lt;2 x i32&gt; &lt; i32 1, i32 1&gt;, &lt; i32 1, i32 2&gt;   <i>; yields: result=&lt;2 x i32&gt; &lt; i32 2, i32 4&gt;</i>
 </pre>
 </div>
 <!-- _______________________________________________________________________ -->
@@ -2624,6 +2625,7 @@ the number of bits in <tt>op1</tt>, the result is undefined.</p>
   &lt;result&gt; = lshr i8  4, 3   <i>; yields {i8}:result = 0</i>
   &lt;result&gt; = lshr i8 -2, 1   <i>; yields {i8}:result = 0x7FFFFFFF </i>
   &lt;result&gt; = lshr i32 1, 32  <i>; undefined</i>
+  &lt;result&gt; = lshr &lt;2 x i32&gt; &lt; i32 -2, i32 4&gt;, &lt; i32 1, i32 2&gt;   <i>; yields: result=&lt;2 x i32&gt; &lt; i32 0x7FFFFFFF, i32 1&gt;</i>
 </pre>
 </div>
 
@@ -2659,6 +2661,7 @@ larger than the number of bits in <tt>op1</tt>, the result is undefined.
   &lt;result&gt; = ashr i8  4, 3   <i>; yields {i8}:result = 0</i>
   &lt;result&gt; = ashr i8 -2, 1   <i>; yields {i8}:result = -1</i>
   &lt;result&gt; = ashr i32 1, 32  <i>; undefined</i>
+  &lt;result&gt; = ashr &lt;2 x i32&gt; &lt; i32 -2, i32 4&gt;, &lt; i32 1, i32 3&gt;   <i>; yields: result=&lt;2 x i32&gt; &lt; i32 -1, i32 0&gt;</i>
 </pre>
 </div>
 
diff --git a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index d61d6ca..445d32f 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -297,6 +297,9 @@ private:
 
   SDValue ExpandEXTRACT_SUBVECTOR(SDValue Op);
   SDValue ExpandEXTRACT_VECTOR_ELT(SDValue Op);
+
+  // Returns the legalized (truncated or extended) shift amount.
+  SDValue LegalizeShiftAmount(SDValue ShiftAmt);
 };
 }
 
@@ -786,8 +789,19 @@ SDValue SelectionDAGLegalize::UnrollVectorOp(SDValue Op) {
         Operands[j] = Operand;
       }
     }
-    Scalars.push_back(DAG.getNode(Op.getOpcode(), EltVT,
-                                  &Operands[0], Operands.size()));
+
+    switch (Op.getOpcode()) {
+    default:
+      Scalars.push_back(DAG.getNode(Op.getOpcode(), EltVT,
+                                    &Operands[0], Operands.size()));
+      break;
+    case ISD::SHL:
+    case ISD::SRA:
+    case ISD::SRL:
+      Scalars.push_back(DAG.getNode(Op.getOpcode(), EltVT, Operands[0],
+                                    LegalizeShiftAmount(Operands[1])));
+      break;
+    }
   }
 
   return DAG.getNode(ISD::BUILD_VECTOR, VT, &Scalars[0], Scalars.size());
@@ -850,6 +864,17 @@ PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx) {
                      PseudoSourceValue::getFixedStack(SPFI), 0);
 }
 
+SDValue SelectionDAGLegalize::LegalizeShiftAmount(SDValue ShiftAmt) {
+  if (TLI.getShiftAmountTy().bitsLT(ShiftAmt.getValueType()))
+    return DAG.getNode(ISD::TRUNCATE, TLI.getShiftAmountTy(), ShiftAmt);
+
+  if (TLI.getShiftAmountTy().bitsGT(ShiftAmt.getValueType()))
+    return DAG.getNode(ISD::ANY_EXTEND, TLI.getShiftAmountTy(), ShiftAmt);
+
+  return ShiftAmt;
+}
+
+
 /// LegalizeOp - We know that the specified value has a legal type, and
 /// that its operands are legal.  Now ensure that the operation itself
 /// is legal, recursively ensuring that the operands' operations remain
@@ -3094,11 +3119,8 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) {
          Node->getOpcode() == ISD::SRL ||
          Node->getOpcode() == ISD::SRA) &&
         !Node->getValueType(0).isVector()) {
-      if (TLI.getShiftAmountTy().bitsLT(Tmp2.getValueType()))
-        Tmp2 = DAG.getNode(ISD::TRUNCATE, TLI.getShiftAmountTy(), Tmp2);
-      else if (TLI.getShiftAmountTy().bitsGT(Tmp2.getValueType()))
-        Tmp2 = DAG.getNode(ISD::ANY_EXTEND, TLI.getShiftAmountTy(), Tmp2);
-    }
+      Tmp2 = LegalizeShiftAmount(Tmp2);
+     }
 
     Result = DAG.UpdateNodeOperands(Result, Tmp1, Tmp2);
 
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 6fc2a67..34f0cca 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -2066,7 +2066,8 @@ SDValue SelectionDAG::getShuffleScalarElt(const SDNode *N, unsigned i) {
 
   if (V.getOpcode() == ISD::BIT_CONVERT) {
     V = V.getOperand(0);
-    if (V.getValueType().getVectorNumElements() != NumElems)
+    MVT VVT = V.getValueType();
+    if (!VVT.isVector() || VVT.getVectorNumElements() != NumElems)
       return SDValue();
   }
   if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
@@ -2418,7 +2419,8 @@ SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT,
            "Shift operators return type must be the same as their first arg");
     assert(VT.isInteger() && N2.getValueType().isInteger() &&
            "Shifts only work on integers");
-    assert(N2.getValueType() == TLI.getShiftAmountTy() &&
+    assert((N2.getValueType() == TLI.getShiftAmountTy() ||
+            (N2.getValueType().isVector() && N2.getValueType().isInteger())) &&
            "Wrong type for shift amount");
 
     // Always fold shifts of i1 values so the code generator doesn't need to
diff --git a/test/CodeGen/X86/vshift_split.ll b/test/CodeGen/X86/vshift_split.ll
new file mode 100644
index 0000000..8f485dd
--- /dev/null
+++ b/test/CodeGen/X86/vshift_split.ll
@@ -0,0 +1,8 @@
+; RUN: llvm-as < %s | llc
+
+; Example that requires splitting and expanding a vector shift.
+define <2 x i64> @update(<2 x i64> %val) nounwind readnone {
+entry:
+	%shr = lshr <2 x i64> %val, < i64 2, i64 2 >		; <<2 x i64>> [#uses=1]
+	ret <2 x i64> %shr
+}