Recognize canonical forms of vector shuffles where the same vector is used for

both source operands.  In the canonical form, the 2nd operand is changed to an
undef and the shuffle mask is adjusted to only reference elements from the 1st
operand.  Radar 7434842.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@90417 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 106 insertions, 1 deletions

diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index c839fc6560..915b309a30 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -2528,6 +2528,25 @@ static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
   return true;
 }
 
+/// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
+/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
+/// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
+static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
+                                unsigned &WhichResult) {
+  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+  if (EltSz == 64)
+    return false;
+
+  unsigned NumElts = VT.getVectorNumElements();
+  WhichResult = (M[0] == 0 ? 0 : 1);
+  for (unsigned i = 0; i < NumElts; i += 2) {
+    if ((unsigned) M[i] != i + WhichResult ||
+        (unsigned) M[i+1] != i + WhichResult)
+      return false;
+  }
+  return true;
+}
+
 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
                        unsigned &WhichResult) {
   unsigned EltSz = VT.getVectorElementType().getSizeInBits();
@@ -2548,6 +2567,33 @@ static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
   return true;
 }
 
+/// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
+/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
+/// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
+static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
+                                unsigned &WhichResult) {
+  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+  if (EltSz == 64)
+    return false;
+
+  unsigned Half = VT.getVectorNumElements() / 2;
+  WhichResult = (M[0] == 0 ? 0 : 1);
+  for (unsigned j = 0; j != 2; ++j) {
+    unsigned Idx = WhichResult;
+    for (unsigned i = 0; i != Half; ++i) {
+      if ((unsigned) M[i + j * Half] != Idx)
+        return false;
+      Idx += 2;
+    }
+  }
+
+  // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
+  if (VT.is64BitVector() && EltSz == 32)
+    return false;
+
+  return true;
+}
+
 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
                        unsigned &WhichResult) {
   unsigned EltSz = VT.getVectorElementType().getSizeInBits();
@@ -2571,6 +2617,33 @@ static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
   return true;
 }
 
+/// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
+/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
+/// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
+static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
+                                unsigned &WhichResult) {
+  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+  if (EltSz == 64)
+    return false;
+
+  unsigned NumElts = VT.getVectorNumElements();
+  WhichResult = (M[0] == 0 ? 0 : 1);
+  unsigned Idx = WhichResult * NumElts / 2;
+  for (unsigned i = 0; i != NumElts; i += 2) {
+    if ((unsigned) M[i] != Idx ||
+        (unsigned) M[i+1] != Idx)
+      return false;
+    Idx += 1;
+  }
+
+  // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
+  if (VT.is64BitVector() && EltSz == 32)
+    return false;
+
+  return true;
+}
+
+
 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
   // Canonicalize all-zeros and all-ones vectors.
   ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
@@ -2683,7 +2756,10 @@ ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
           isVEXTMask(M, VT, ReverseVEXT, Imm) ||
           isVTRNMask(M, VT, WhichResult) ||
           isVUZPMask(M, VT, WhichResult) ||
-          isVZIPMask(M, VT, WhichResult));
+          isVZIPMask(M, VT, WhichResult) ||
+          isVTRN_v_undef_Mask(M, VT, WhichResult) ||
+          isVUZP_v_undef_Mask(M, VT, WhichResult) ||
+          isVZIP_v_undef_Mask(M, VT, WhichResult));
 }
 
 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
@@ -2815,6 +2891,16 @@ static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
     return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
                        V1, V2).getValue(WhichResult);
 
+  if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
+    return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
+                       V1, V1).getValue(WhichResult);
+  if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
+    return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
+                       V1, V1).getValue(WhichResult);
+  if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
+    return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
+                       V1, V1).getValue(WhichResult);
+
   // If the shuffle is not directly supported and it has 4 elements, use
   // the PerfectShuffle-generated table to synthesize it from other shuffles.
   if (VT.getVectorNumElements() == 4 &&
diff --git a/test/CodeGen/ARM/2009-12-02-vtrn-undef.ll b/test/CodeGen/ARM/2009-12-02-vtrn-undef.ll
new file mode 100644
index 0000000000..a737591bd9
--- /dev/null
+++ b/test/CodeGen/ARM/2009-12-02-vtrn-undef.ll
@@ -0,0 +1,19 @@
+; RUN: llc -mcpu=cortex-a8 < %s | FileCheck %s
+
+target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:32-f32:32:32-f64:32:32-v64:64:64-v128:128:128-a0:0:32-n32"
+target triple = "armv7-apple-darwin10"
+
+%struct.int16x8_t = type { <8 x i16> }
+%struct.int16x8x2_t = type { [2 x %struct.int16x8_t] }
+
+define arm_apcscc void @t(%struct.int16x8x2_t* noalias nocapture sret %agg.result, <8 x i16> %tmp.0, %struct.int16x8x2_t* nocapture %dst) nounwind {
+entry:
+;CHECK: vtrn.16
+  %0 = shufflevector <8 x i16> %tmp.0, <8 x i16> undef, <8 x i32> <i32 0, i32 0, i32 2, i32 2, i32 4, i32 4, i32 6, i32 6>
+  %1 = shufflevector <8 x i16> %tmp.0, <8 x i16> undef, <8 x i32> <i32 1, i32 1, i32 3, i32 3, i32 5, i32 5, i32 7, i32 7>
+  %agg.result1218.0 = getelementptr %struct.int16x8x2_t* %agg.result, i32 0, i32 0, i32 0, i32 0 ; <<8 x i16>*>
+  store <8 x i16> %0, <8 x i16>* %agg.result1218.0, align 16
+  %agg.result12.1.0 = getelementptr %struct.int16x8x2_t* %agg.result, i32 0, i32 0, i32 1, i32 0 ; <<8 x i16>*>
+  store <8 x i16> %1, <8 x i16>* %agg.result12.1.0, align 16
+  ret void
+}