Teach dag combine to match halfword byteswap patterns.

1. (((x) & 0xFF00) >> 8) | (((x) & 0x00FF) << 8)
   => (bswap x) >> 16
2. ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0xff000000)>>8)|((x&0x00ff0000)<<8))
   => (rotl (bswap x) 16)

This allows us to eliminate most of the def : Pat patterns for ARM rev16
revsh instructions. It catches many more cases for ARM and x86.

rdar://9609108


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@133503 91177308-0d34-0410-b5e6-96231b3b80d8

6 files changed, 348 insertions, 67 deletions

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 4ac590af57..443fb3271d 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -238,6 +238,9 @@ namespace {
     SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
     SDValue BuildSDIV(SDNode *N);
     SDValue BuildUDIV(SDNode *N);
+    SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
+                               bool DemandHighBits = true);
+    SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
     SDNode *MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL);
     SDValue ReduceLoadWidth(SDNode *N);
     SDValue ReduceLoadOpStoreWidth(SDNode *N);
@@ -2512,6 +2515,244 @@ SDValue DAGCombiner::visitAND(SDNode *N) {
   return SDValue();
 }
 
+/// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
+///
+SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
+                                        bool DemandHighBits) {
+  if (!LegalOperations)
+    return SDValue();
+
+  EVT VT = N->getValueType(0);
+  if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
+    return SDValue();
+  if (!TLI.isOperationLegal(ISD::BSWAP, VT))
+    return SDValue();
+
+  // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
+  bool LookPassAnd0 = false;
+  bool LookPassAnd1 = false;
+  if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
+      std::swap(N0, N1);
+  if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
+      std::swap(N0, N1);
+  if (N0.getOpcode() == ISD::AND) {
+    if (!N0.getNode()->hasOneUse())
+      return SDValue();
+    ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+    if (!N01C || N01C->getZExtValue() != 0xFF00)
+      return SDValue();
+    N0 = N0.getOperand(0);
+    LookPassAnd0 = true;
+  }
+
+  if (N1.getOpcode() == ISD::AND) {
+    if (!N1.getNode()->hasOneUse())
+      return SDValue();
+    ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
+    if (!N11C || N11C->getZExtValue() != 0xFF)
+      return SDValue();
+    N1 = N1.getOperand(0);
+    LookPassAnd1 = true;
+  }
+
+  if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
+    std::swap(N0, N1);
+  if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
+    return SDValue();
+  if (!N0.getNode()->hasOneUse() ||
+      !N1.getNode()->hasOneUse())
+    return SDValue();
+
+  ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+  ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
+  if (!N01C || !N11C)
+    return SDValue();
+  if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
+    return SDValue();
+
+  // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
+  SDValue N00 = N0->getOperand(0);
+  if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
+    if (!N00.getNode()->hasOneUse())
+      return SDValue();
+    ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
+    if (!N001C || N001C->getZExtValue() != 0xFF)
+      return SDValue();
+    N00 = N00.getOperand(0);
+    LookPassAnd0 = true;
+  }
+
+  SDValue N10 = N1->getOperand(0);
+  if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
+    if (!N10.getNode()->hasOneUse())
+      return SDValue();
+    ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
+    if (!N101C || N101C->getZExtValue() != 0xFF00)
+      return SDValue();
+    N10 = N10.getOperand(0);
+    LookPassAnd1 = true;
+  }
+
+  if (N00 != N10)
+    return SDValue();
+
+  // Make sure everything beyond the low halfword is zero since the SRL 16
+  // will clear the top bits.
+  unsigned OpSizeInBits = VT.getSizeInBits();
+  if (DemandHighBits && OpSizeInBits > 16 &&
+      (!LookPassAnd0 || !LookPassAnd1) &&
+      !DAG.MaskedValueIsZero(N10, APInt::getHighBitsSet(OpSizeInBits, 16)))
+    return SDValue();
+  
+  SDValue Res = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT, N00);
+  if (OpSizeInBits > 16)
+    Res = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Res,
+                      DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
+  return Res;
+}
+
+/// isBSwapHWordElement - Return true if the specified node is an element
+/// that makes up a 32-bit packed halfword byteswap. i.e.
+/// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
+static bool isBSwapHWordElement(SDValue N, SmallVector<SDNode*,4> &Parts) {
+  if (!N.getNode()->hasOneUse())
+    return false;
+
+  unsigned Opc = N.getOpcode();
+  if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
+    return false;
+
+  ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
+  if (!N1C)
+    return false;
+
+  unsigned Num;
+  switch (N1C->getZExtValue()) {
+  default:
+    return false;
+  case 0xFF:       Num = 0; break;
+  case 0xFF00:     Num = 1; break;
+  case 0xFF0000:   Num = 2; break;
+  case 0xFF000000: Num = 3; break;
+  }
+
+  // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
+  SDValue N0 = N.getOperand(0);
+  if (Opc == ISD::AND) {
+    if (Num == 0 || Num == 2) {
+      // (x >> 8) & 0xff
+      // (x >> 8) & 0xff0000
+      if (N0.getOpcode() != ISD::SRL)
+        return false;
+      ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+      if (!C || C->getZExtValue() != 8)
+        return false;
+    } else {
+      // (x << 8) & 0xff00
+      // (x << 8) & 0xff000000
+      if (N0.getOpcode() != ISD::SHL)
+        return false;
+      ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+      if (!C || C->getZExtValue() != 8)
+        return false;
+    }
+  } else if (Opc == ISD::SHL) {
+    // (x & 0xff) << 8
+    // (x & 0xff0000) << 8
+    if (Num != 0 && Num != 2)
+      return false;
+    ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
+    if (!C || C->getZExtValue() != 8)
+      return false;
+  } else { // Opc == ISD::SRL
+    // (x & 0xff00) >> 8
+    // (x & 0xff000000) >> 8
+    if (Num != 1 && Num != 3)
+      return false;
+    ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
+    if (!C || C->getZExtValue() != 8)
+      return false;
+  }
+
+  if (Parts[Num])
+    return false;
+
+  Parts[Num] = N0.getOperand(0).getNode();
+  return true;
+}
+
+/// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
+/// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
+/// => (rotl (bswap x), 16)
+SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
+  if (!LegalOperations)
+    return SDValue();
+
+  EVT VT = N->getValueType(0);
+  if (VT != MVT::i32)
+    return SDValue();
+  if (!TLI.isOperationLegal(ISD::BSWAP, VT))
+    return SDValue();
+
+  SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
+  // Look for either
+  // (or (or (and), (and)), (or (and), (and)))
+  // (or (or (or (and), (and)), (and)), (and))
+  if (N0.getOpcode() != ISD::OR)
+    return SDValue();
+  SDValue N00 = N0.getOperand(0);
+  SDValue N01 = N0.getOperand(1);
+
+  if (N1.getOpcode() == ISD::OR) {
+    // (or (or (and), (and)), (or (and), (and)))
+    SDValue N000 = N00.getOperand(0);
+    if (!isBSwapHWordElement(N000, Parts))
+      return SDValue();
+
+    SDValue N001 = N00.getOperand(1);
+    if (!isBSwapHWordElement(N001, Parts))
+      return SDValue();
+    SDValue N010 = N01.getOperand(0);
+    if (!isBSwapHWordElement(N010, Parts))
+      return SDValue();
+    SDValue N011 = N01.getOperand(1);
+    if (!isBSwapHWordElement(N011, Parts))
+      return SDValue();
+  } else {
+    // (or (or (or (and), (and)), (and)), (and))
+    if (!isBSwapHWordElement(N1, Parts))
+      return SDValue();
+    if (!isBSwapHWordElement(N01, Parts))
+      return SDValue();
+    if (N00.getOpcode() != ISD::OR)
+      return SDValue();
+    SDValue N000 = N00.getOperand(0);
+    if (!isBSwapHWordElement(N000, Parts))
+      return SDValue();
+    SDValue N001 = N00.getOperand(1);
+    if (!isBSwapHWordElement(N001, Parts))
+      return SDValue();
+  }
+
+  // Make sure the parts are all coming from the same node.
+  if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
+    return SDValue();
+
+  SDValue BSwap = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT,
+                              SDValue(Parts[0],0));
+
+  // Result of the bswap should be rotated by 16. If it's not legal, than
+  // do  (x << 16) | (x >> 16).
+  SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
+  if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
+    return DAG.getNode(ISD::ROTL, N->getDebugLoc(), VT, BSwap, ShAmt);
+  else if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
+    return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, BSwap, ShAmt);
+  return DAG.getNode(ISD::OR, N->getDebugLoc(), VT,
+                     DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, BSwap, ShAmt),
+                     DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, BSwap, ShAmt));
+}
+
 SDValue DAGCombiner::visitOR(SDNode *N) {
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
@@ -2547,6 +2788,15 @@ SDValue DAGCombiner::visitOR(SDNode *N) {
   // fold (or x, c) -> c iff (x & ~c) == 0
   if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
     return N1;
+
+  // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
+  SDValue BSwap = MatchBSwapHWord(N, N0, N1);
+  if (BSwap.getNode() != 0)
+    return BSwap;
+  BSwap = MatchBSwapHWordLow(N, N0, N1);
+  if (BSwap.getNode() != 0)
+    return BSwap;
+
   // reassociate or
   SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
   if (ROR.getNode() != 0)
@@ -4606,6 +4856,16 @@ SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
     CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
     return SDValue(N, 0);   // Return N so it doesn't get rechecked!
   }
+
+  // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
+  if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
+    SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
+                                       N0.getOperand(1), false);
+    if (BSwap.getNode() != 0)
+      return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
+                         BSwap, N1);
+  }
+
   return SDValue();
 }
 
@@ -5231,7 +5491,8 @@ SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
   // fold (sint_to_fp c1) -> c1fp
   if (N0C && OpVT != MVT::ppcf128 &&
       // ...but only if the target supports immediate floating-point values
-      (Level == llvm::Unrestricted || TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
+      (Level == llvm::Unrestricted ||
+       TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
     return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
 
   // If the input is a legal type, and SINT_TO_FP is not legal on this target,
@@ -5255,7 +5516,8 @@ SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
   // fold (uint_to_fp c1) -> c1fp
   if (N0C && OpVT != MVT::ppcf128 &&
       // ...but only if the target supports immediate floating-point values
-      (Level == llvm::Unrestricted || TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
+      (Level == llvm::Unrestricted ||
+       TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
     return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
 
   // If the input is a legal type, and UINT_TO_FP is not legal on this target,
diff --git a/lib/Target/ARM/ARMInstrInfo.td b/lib/Target/ARM/ARMInstrInfo.td
index e2bbcfb12c..5c013de238 100644
--- a/lib/Target/ARM/ARMInstrInfo.td
+++ b/lib/Target/ARM/ARMInstrInfo.td
@@ -3008,41 +3008,22 @@ def REV  : AMiscA1I<0b01101011, 0b0011, (outs GPR:$Rd), (ins GPR:$Rm),
               IIC_iUNAr, "rev", "\t$Rd, $Rm",
               [(set GPR:$Rd, (bswap GPR:$Rm))]>, Requires<[IsARM, HasV6]>;
 
+let AddedComplexity = 5 in
 def REV16 : AMiscA1I<0b01101011, 0b1011, (outs GPR:$Rd), (ins GPR:$Rm),
                IIC_iUNAr, "rev16", "\t$Rd, $Rm",
-               [(set GPR:$Rd,
-                   (or (and (srl GPR:$Rm, (i32 8)), 0xFF),
-                       (or (and (shl GPR:$Rm, (i32 8)), 0xFF00),
-                           (or (and (srl GPR:$Rm, (i32 8)), 0xFF0000),
-                               (and (shl GPR:$Rm, (i32 8)), 0xFF000000)))))]>,
+               [(set GPR:$Rd, (rotr (bswap GPR:$Rm), (i32 16)))]>,
                Requires<[IsARM, HasV6]>;
 
-def : ARMV6Pat<(or (or (or (and (srl GPR:$Rm, (i32 8)), 0xFF0000),
-                           (and (shl GPR:$Rm, (i32 8)), 0xFF000000)),
-                       (and (srl GPR:$Rm, (i32 8)), 0xFF)),
-                   (and (shl GPR:$Rm, (i32 8)), 0xFF00)),
-               (REV16 GPR:$Rm)>;
-
+let AddedComplexity = 5 in
 def REVSH : AMiscA1I<0b01101111, 0b1011, (outs GPR:$Rd), (ins GPR:$Rm),
                IIC_iUNAr, "revsh", "\t$Rd, $Rm",
-               [(set GPR:$Rd,
-                  (sext_inreg
-                    (or (srl GPR:$Rm, (i32 8)),
-                        (shl GPR:$Rm, (i32 8))), i16))]>,
+               [(set GPR:$Rd, (sra (bswap GPR:$Rm), (i32 16)))]>,
                Requires<[IsARM, HasV6]>;
 
-def : ARMV6Pat<(sext_inreg (or (srl (and GPR:$Rm, 0xFF00), (i32 8)),
-                               (shl GPR:$Rm, (i32 8))), i16),
-               (REVSH GPR:$Rm)>;
-
 def : ARMV6Pat<(or (sra (shl GPR:$Rm, (i32 24)), (i32 16)),
                    (and (srl GPR:$Rm, (i32 8)), 0xFF)),
                (REVSH GPR:$Rm)>;
 
-// Need the AddedComplexity or else MOVs + REV would be chosen.
-let AddedComplexity = 5 in
-def : ARMV6Pat<(sra (bswap GPR:$Rm), (i32 16)), (REVSH GPR:$Rm)>;
-
 def lsl_shift_imm : SDNodeXForm<imm, [{
   unsigned Sh = ARM_AM::getSORegOpc(ARM_AM::lsl, N->getZExtValue());
   return CurDAG->getTargetConstant(Sh, MVT::i32);
diff --git a/lib/Target/ARM/ARMInstrThumb.td b/lib/Target/ARM/ARMInstrThumb.td
index 8430aa31e7..44fbc021d8 100644
--- a/lib/Target/ARM/ARMInstrThumb.td
+++ b/lib/Target/ARM/ARMInstrThumb.td
@@ -1176,31 +1176,16 @@ def tREV16 :                    // A8.6.135
   T1pIMiscEncode<{1,0,1,0,0,1,?}, (outs tGPR:$Rd), (ins tGPR:$Rm),
                  IIC_iUNAr,
                  "rev16", "\t$Rd, $Rm",
-             [(set tGPR:$Rd,
-                   (or (and (srl tGPR:$Rm, (i32 8)), 0xFF),
-                       (or (and (shl tGPR:$Rm, (i32 8)), 0xFF00),
-                           (or (and (srl tGPR:$Rm, (i32 8)), 0xFF0000),
-                               (and (shl tGPR:$Rm, (i32 8)), 0xFF000000)))))]>,
+             [(set tGPR:$Rd, (rotr (bswap tGPR:$Rm), (i32 16)))]>,
                 Requires<[IsThumb, IsThumb1Only, HasV6]>;
 
 def tREVSH :                    // A8.6.136
   T1pIMiscEncode<{1,0,1,0,1,1,?}, (outs tGPR:$Rd), (ins tGPR:$Rm),
                  IIC_iUNAr,
                  "revsh", "\t$Rd, $Rm",
-                 [(set tGPR:$Rd,
-                       (sext_inreg
-                         (or (srl tGPR:$Rm, (i32 8)),
-                             (shl tGPR:$Rm, (i32 8))), i16))]>,
+                 [(set tGPR:$Rd, (sra (bswap tGPR:$Rm), (i32 16)))]>,
                  Requires<[IsThumb, IsThumb1Only, HasV6]>;
 
-def : T1Pat<(sext_inreg (or (srl (and tGPR:$Rm, 0xFF00), (i32 8)),
-                            (shl tGPR:$Rm, (i32 8))), i16),
-            (tREVSH tGPR:$Rm)>,
-      Requires<[IsThumb, IsThumb1Only, HasV6]>;
-
-def : T1Pat<(sra (bswap tGPR:$Rm), (i32 16)), (tREVSH tGPR:$Rm)>,
-      Requires<[IsThumb, IsThumb1Only, HasV6]>;
-
 // Rotate right register
 def tROR :                      // A8.6.139
   T1sItDPEncode<0b0111, (outs tGPR:$Rdn), (ins tGPR:$Rn, tGPR:$Rm),
diff --git a/lib/Target/ARM/ARMInstrThumb2.td b/lib/Target/ARM/ARMInstrThumb2.td
index cd077a86e9..60fff05520 100644
--- a/lib/Target/ARM/ARMInstrThumb2.td
+++ b/lib/Target/ARM/ARMInstrThumb2.td
@@ -2587,35 +2587,16 @@ def t2REV : T2I_misc<0b01, 0b00, (outs rGPR:$Rd), (ins rGPR:$Rm), IIC_iUNAr,
 
 def t2REV16 : T2I_misc<0b01, 0b01, (outs rGPR:$Rd), (ins rGPR:$Rm), IIC_iUNAr,
                        "rev16", ".w\t$Rd, $Rm",
-                [(set rGPR:$Rd,
-                    (or (and (srl rGPR:$Rm, (i32 8)), 0xFF),
-                        (or (and (shl rGPR:$Rm, (i32 8)), 0xFF00),
-                            (or (and (srl rGPR:$Rm, (i32 8)), 0xFF0000),
-                               (and (shl rGPR:$Rm, (i32 8)), 0xFF000000)))))]>;
-
-def : T2Pat<(or (or (or (and (srl rGPR:$Rm, (i32 8)), 0xFF0000),
-                        (and (shl rGPR:$Rm, (i32 8)), 0xFF000000)),
-                    (and (srl rGPR:$Rm, (i32 8)), 0xFF)),
-                (and (shl rGPR:$Rm, (i32 8)), 0xFF00)),
-            (t2REV16 rGPR:$Rm)>;
+                [(set rGPR:$Rd, (rotr (bswap rGPR:$Rm), (i32 16)))]>;
 
 def t2REVSH : T2I_misc<0b01, 0b11, (outs rGPR:$Rd), (ins rGPR:$Rm), IIC_iUNAr,
                        "revsh", ".w\t$Rd, $Rm",
-                 [(set rGPR:$Rd,
-                    (sext_inreg
-                      (or (srl rGPR:$Rm, (i32 8)),
-                          (shl rGPR:$Rm, (i32 8))), i16))]>;
-
-def : T2Pat<(sext_inreg (or (srl (and rGPR:$Rm, 0xFF00), (i32 8)),
-                            (shl rGPR:$Rm, (i32 8))), i16),
-            (t2REVSH rGPR:$Rm)>;
+                 [(set rGPR:$Rd, (sra (bswap rGPR:$Rm), (i32 16)))]>;
 
 def : T2Pat<(or (sra (shl rGPR:$Rm, (i32 24)), (i32 16)),
-                   (and (srl rGPR:$Rm, (i32 8)), 0xFF)),
+                (and (srl rGPR:$Rm, (i32 8)), 0xFF)),
             (t2REVSH rGPR:$Rm)>;
 
-def : T2Pat<(sra (bswap rGPR:$Rm), (i32 16)), (t2REVSH rGPR:$Rm)>;
-
 def t2PKHBT : T2ThreeReg<
             (outs rGPR:$Rd), (ins rGPR:$Rn, rGPR:$Rm, shift_imm:$sh),
                   IIC_iBITsi, "pkhbt", "\t$Rd, $Rn, $Rm$sh",
diff --git a/test/CodeGen/ARM/rev.ll b/test/CodeGen/ARM/rev.ll
index c210a55b66..ea44c28fb7 100644
--- a/test/CodeGen/ARM/rev.ll
+++ b/test/CodeGen/ARM/rev.ll
@@ -84,3 +84,43 @@ entry:
   %or10 = or i32 %or6, %shl
   ret i32 %or10
 }
+
+; rdar://9164521
+define i32 @test7(i32 %a) nounwind readnone {
+entry:
+; CHECK: test7
+; CHECK: rev r0, r0
+; CHECK: lsr r0, r0, #16
+  %and = lshr i32 %a, 8
+  %shr3 = and i32 %and, 255
+  %and2 = shl i32 %a, 8
+  %shl = and i32 %and2, 65280
+  %or = or i32 %shr3, %shl
+  ret i32 %or
+}
+
+define i32 @test8(i32 %a) nounwind readnone {
+entry:
+; CHECK: test8
+; CHECK: revsh r0, r0
+  %and = lshr i32 %a, 8
+  %shr4 = and i32 %and, 255
+  %and2 = shl i32 %a, 8
+  %or = or i32 %shr4, %and2
+  %sext = shl i32 %or, 16
+  %conv3 = ashr exact i32 %sext, 16
+  ret i32 %conv3
+}
+
+define zeroext i16 @test9(i16 zeroext %v) nounwind readnone {
+entry:
+; CHECK: test9
+; CHECK: rev r0, r0
+; CHECK: lsr r0, r0, #16
+  %conv = zext i16 %v to i32
+  %shr4 = lshr i32 %conv, 8
+  %shl = shl nuw nsw i32 %conv, 8
+  %or = or i32 %shr4, %shl
+  %conv3 = trunc i32 %or to i16
+  ret i16 %conv3
+}
diff --git a/test/CodeGen/X86/bswap.ll b/test/CodeGen/X86/bswap.ll
index 0a72c1c478..a7540aafa9 100644
--- a/test/CodeGen/X86/bswap.ll
+++ b/test/CodeGen/X86/bswap.ll
@@ -1,8 +1,6 @@
 ; bswap should be constant folded when it is passed a constant argument
 
-; RUN: llc < %s -march=x86 | \
-; RUN:   grep bswapl | count 3
-; RUN: llc < %s -march=x86 | grep rolw | count 1
+; RUN: llc < %s -march=x86 | FileCheck %s
 
 declare i16 @llvm.bswap.i16(i16)
 
@@ -11,17 +9,51 @@ declare i32 @llvm.bswap.i32(i32)
 declare i64 @llvm.bswap.i64(i64)
 
 define i16 @W(i16 %A) {
+; CHECK: W:
+; CHECK: rolw $8, %ax
         %Z = call i16 @llvm.bswap.i16( i16 %A )         ; <i16> [#uses=1]
         ret i16 %Z
 }
 
 define i32 @X(i32 %A) {
+; CHECK: X:
+; CHECK: bswapl %eax
         %Z = call i32 @llvm.bswap.i32( i32 %A )         ; <i32> [#uses=1]
         ret i32 %Z
 }
 
 define i64 @Y(i64 %A) {
+; CHECK: Y:
+; CHECK: bswapl %eax
+; CHECK: bswapl %edx
         %Z = call i64 @llvm.bswap.i64( i64 %A )         ; <i64> [#uses=1]
         ret i64 %Z
 }
 
+; rdar://9164521
+define i32 @test1(i32 %a) nounwind readnone {
+entry:
+; CHECK: test1
+; CHECK: bswapl %eax
+; CHECK: shrl $16, %eax
+  %and = lshr i32 %a, 8
+  %shr3 = and i32 %and, 255
+  %and2 = shl i32 %a, 8
+  %shl = and i32 %and2, 65280
+  %or = or i32 %shr3, %shl
+  ret i32 %or
+}
+
+define i32 @test2(i32 %a) nounwind readnone {
+entry:
+; CHECK: test2
+; CHECK: bswapl %eax
+; CHECK: sarl $16, %eax
+  %and = lshr i32 %a, 8
+  %shr4 = and i32 %and, 255
+  %and2 = shl i32 %a, 8
+  %or = or i32 %shr4, %and2
+  %sext = shl i32 %or, 16
+  %conv3 = ashr exact i32 %sext, 16
+  ret i32 %conv3
+}