Generate a VTBL instruction instead of a series of loads and stores when we

can. As Nate pointed out, VTBL isn't super performant, but it *has* to be better
than this:

_shuf:
@ BB#0:       @ %entry
  push        {r4, r7, lr}
  add         r7, sp, #4
  sub         sp, #12
  mov         r4, sp
  bic         r4, r4, #7
  mov         sp, r4
  mov         r2, sp
  vmov        d16, r0, r1
  orr         r0, r2, #6
  orr         r3, r2, #7
  vst1.8      {d16[0]}, [r3]
  vst1.8      {d16[5]}, [r0]
  subs        r4, r7, #4
  orr         r0, r2, #5
  vst1.8      {d16[4]}, [r0]
  orr         r0, r2, #4
  vst1.8      {d16[4]}, [r0]
  orr         r0, r2, #3
  vst1.8      {d16[0]}, [r0]
  orr         r0, r2, #2
  vst1.8      {d16[2]}, [r0]
  orr         r0, r2, #1
  vst1.8      {d16[1]}, [r0]
  vst1.8      {d16[3]}, [r2]
  vldr.64     d16, [sp]
  vmov        r0, r1, d16
  mov         sp, r4
  pop         {r4, r7, pc}

The "illegal" testcase in vext.ll is no longer illegal.
<rdar://problem/9078775>


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@127630 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 67 insertions, 13 deletions

diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 8ef44d1..c7d847a 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -2842,6 +2842,35 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
     break;
   }
 
+  case ARMISD::VTBL1: {
+    DebugLoc dl = N->getDebugLoc();
+    EVT VT = N->getValueType(0);
+    SmallVector<SDValue, 6> Ops;
+
+    Ops.push_back(N->getOperand(0));
+    Ops.push_back(N->getOperand(1));
+    Ops.push_back(getAL(CurDAG));                    // Predicate
+    Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // Predicate Register
+    return CurDAG->getMachineNode(ARM::VTBL1, dl, VT, Ops.data(), Ops.size());
+  }
+  case ARMISD::VTBL2: {
+    DebugLoc dl = N->getDebugLoc();
+    EVT VT = N->getValueType(0);
+
+    // Form a REG_SEQUENCE to force register allocation.
+    SDValue V0 = N->getOperand(0);
+    SDValue V1 = N->getOperand(1);
+    SDValue RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
+
+    SmallVector<SDValue, 6> Ops;
+    Ops.push_back(RegSeq);
+    Ops.push_back(N->getOperand(2));
+    Ops.push_back(getAL(CurDAG));                    // Predicate
+    Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // Predicate Register
+    return CurDAG->getMachineNode(ARM::VTBL2Pseudo, dl, VT,
+                                  Ops.data(), Ops.size());
+  }
+
   case ISD::CONCAT_VECTORS:
     return SelectConcatVector(N);
   }
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 3a34051..665c823 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -852,6 +852,10 @@ const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case ARMISD::VZIP:          return "ARMISD::VZIP";
   case ARMISD::VUZP:          return "ARMISD::VUZP";
   case ARMISD::VTRN:          return "ARMISD::VTRN";
+  case ARMISD::VTBL1:         return "ARMISD::VTBL1";
+  case ARMISD::VTBL2:         return "ARMISD::VTBL2";
+  case ARMISD::VTBL3:         return "ARMISD::VTBL3";
+  case ARMISD::VTBL4:         return "ARMISD::VTBL4";
   case ARMISD::VMULLs:        return "ARMISD::VMULLs";
   case ARMISD::VMULLu:        return "ARMISD::VMULLu";
   case ARMISD::BUILD_VECTOR:  return "ARMISD::BUILD_VECTOR";
@@ -4055,6 +4059,29 @@ static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
   }
 }
 
+static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
+                                       SmallVectorImpl<int> &ShuffleMask,
+                                       SelectionDAG &DAG) {
+  // Check to see if we can use the VTBL instruction.
+  SDValue V1 = Op.getOperand(0);
+  SDValue V2 = Op.getOperand(1);
+  DebugLoc DL = Op.getDebugLoc();
+
+  SmallVector<SDValue, 8> VTBLMask;
+  for (SmallVectorImpl<int>::iterator
+         I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
+    VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
+
+  if (V2.getNode()->getOpcode() == ISD::UNDEF)
+    return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
+                       DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
+                                   &VTBLMask[0], 8));
+  else
+    return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2, 
+                       DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
+                                   &VTBLMask[0], 8));
+}
+
 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
   SDValue V1 = Op.getOperand(0);
   SDValue V2 = Op.getOperand(1);
@@ -4172,6 +4199,12 @@ static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
     return DAG.getNode(ISD::BITCAST, dl, VT, Val);
   }
 
+  if (VT == MVT::v8i8) {
+    SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
+    if (NewOp.getNode())
+      return NewOp;
+  }
+
   return SDValue();
 }
 
@@ -4534,7 +4567,7 @@ SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::GlobalAddress:
     return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
       LowerGlobalAddressELF(Op, DAG);
-  case ISD::GlobalTLSAddress:   return LowerGlobalTLSAddress(Op, DAG);
+  case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
   case ISD::SELECT:        return LowerSELECT(Op, DAG);
   case ISD::SELECT_CC:     return LowerSELECT_CC(Op, DAG);
   case ISD::BR_CC:         return LowerBR_CC(Op, DAG);
diff --git a/lib/Target/ARM/ARMISelLowering.h b/lib/Target/ARM/ARMISelLowering.h
index 0f56201..8eb4525 100644
--- a/lib/Target/ARM/ARMISelLowering.h
+++ b/lib/Target/ARM/ARMISelLowering.h
@@ -153,6 +153,10 @@ namespace llvm {
       VZIP,         // zip (interleave)
       VUZP,         // unzip (deinterleave)
       VTRN,         // transpose
+      VTBL1,        // 1-register shuffle with mask
+      VTBL2,        // 2-register shuffle with mask
+      VTBL3,        // 3-register shuffle with mask
+      VTBL4,        // 4-register shuffle with mask
 
       // Vector multiply long:
       VMULLs,       // ...signed
diff --git a/test/CodeGen/ARM/vext.ll b/test/CodeGen/ARM/vext.ll
index 55abefe..c8d9045 100644
--- a/test/CodeGen/ARM/vext.ll
+++ b/test/CodeGen/ARM/vext.ll
@@ -121,15 +121,3 @@ define <4 x i16> @test_largespan(<8 x i16>* %B) nounwind {
         %tmp2 = shufflevector <8 x i16> %tmp1, <8 x i16> undef, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
         ret <4 x i16> %tmp2
 }
-
-; The actual shuffle code only handles some cases, make sure we check
-; this rather than blindly emitting a VECTOR_SHUFFLE (infinite
-; lowering loop can result otherwise).
-define <8 x i8> @test_illegal(<16 x i8>* %A, <16 x i8>* %B) nounwind {
-;CHECK: test_illegal:
-;CHECK: vst1.8
-       %tmp1 = load <16 x i8>* %A
-       %tmp2 = load <16 x i8>* %B
-       %tmp3 = shufflevector <16 x i8> %tmp1, <16 x i8> %tmp2, <8 x i32> <i32 0, i32 7, i32 5, i32 25, i32 3, i32 2, i32 2, i32 26>
-       ret <8 x i8> %tmp3
-}