Index: include/llvm/Target/TargetSelectionDAG.td
===================================================================
--- include/llvm/Target/TargetSelectionDAG.td
+++ include/llvm/Target/TargetSelectionDAG.td
@@ -132,7 +132,7 @@
 def SDTFPTernaryOp : SDTypeProfile<1, 3, [  // fmadd, fnmsub, etc.
   SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisSameAs<0, 3>, SDTCisFP<0>
 ]>;
-def SDTIntUnaryOp : SDTypeProfile<1, 1, [   // ctlz
+def SDTIntUnaryOp : SDTypeProfile<1, 1, [   // ctlz, cttz
   SDTCisSameAs<0, 1>, SDTCisInt<0>
 ]>;
 def SDTIntExtendOp : SDTypeProfile<1, 1, [  // sext, zext, anyext
Index: lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -2780,11 +2780,22 @@
     // This trivially expands to CTTZ.
     return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
   case ISD::CTTZ: {
+    EVT VT = Op.getValueType();
+    unsigned len = VT.getSizeInBits();
+
+    if (TLI.isOperationLegalOrCustom(ISD::CTTZ_ZERO_UNDEF, VT)) {
+      EVT SetCCVT = getSetCCResultType(VT);
+      SDValue CTTZ = DAG.getNode(ISD::CTTZ_ZERO_UNDEF, dl, VT, Op);
+      SDValue Zero = DAG.getConstant(0, dl, VT);
+      SDValue SrcIsZero = DAG.getSetCC(dl, SetCCVT, Op, Zero, ISD::SETEQ);
+      return DAG.getNode(ISD::SELECT, dl, VT, SrcIsZero,
+                         DAG.getConstant(len, dl, VT), CTTZ);
+    }
+
     // for now, we use: { return popcount(~x & (x - 1)); }
     // unless the target has ctlz but not ctpop, in which case we use:
     // { return 32 - nlz(~x & (x-1)); }
     // Ref: "Hacker's Delight" by Henry Warren
-    EVT VT = Op.getValueType();
     SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
                                DAG.getNOT(dl, Op, VT),
                                DAG.getNode(ISD::SUB, dl, VT, Op,
Index: lib/Target/AMDGPU/AMDGPUISelLowering.h
===================================================================
--- lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -115,6 +115,7 @@
   SDValue LowerSDIVREM(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const;
+  SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) const;
   void LowerUDIVREM64(SDValue Op, SelectionDAG &DAG,
                                     SmallVectorImpl<SDValue> &Results) const;
   void analyzeFormalArgumentsCompute(CCState &State,
@@ -371,6 +372,8 @@
   BFM, // Insert a range of bits into a 32-bit word.
   FFBH_U32, // ctlz with -1 if input is zero.
   FFBH_I32,
+  FFBL_U32, // cttz with -1 if input is zero.
+  FFBL_I32,
   MUL_U24,
   MUL_I24,
   MULHI_U24,
Index: lib/Target/AMDGPU/AMDGPUISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -417,7 +417,7 @@
     setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Custom);
 
   if (Subtarget->hasFFBL())
-    setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Legal);
+    setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Custom);
 
   setOperationAction(ISD::CTLZ, MVT::i64, Custom);
   setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
@@ -1105,6 +1105,9 @@
   case ISD::FP_TO_FP16: return LowerFP_TO_FP16(Op, DAG);
   case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
   case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
+  case ISD::CTTZ: 
+  case ISD::CTTZ_ZERO_UNDEF:
+		return LowerCTTZ(Op, DAG);
   case ISD::CTLZ:
   case ISD::CTLZ_ZERO_UNDEF:
     return LowerCTLZ(Op, DAG);
@@ -2014,6 +2017,61 @@
   return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
 }
 
+
+SDValue AMDGPUTargetLowering:: LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
+  SDLoc SL(Op);
+  SDValue Src = Op.getOperand(0);
+  bool ZeroUndef = Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF;
+
+  if (ZeroUndef && Src.getValueType() == MVT::i32)
+    return DAG.getNode(AMDGPUISD::FFBL_U32, SL, MVT::i32, Src);
+
+  SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
+
+  const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
+  const SDValue One = DAG.getConstant(1, SL, MVT::i32);
+
+  SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
+  SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
+
+  EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(),
+                                   *DAG.getContext(), MVT::i32);
+
+  SDValue Hi0 = DAG.getSetCC(SL, SetCCVT, Hi, One, ISD::SETEQ);
+
+  SDValue CttzLo = DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SL, MVT::i32, Lo);
+  SDValue CttzHi = DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SL, MVT::i32, Hi);
+
+  const SDValue Bits32 = DAG.getConstant(32, SL, MVT::i32);
+  SDValue Add = DAG.getNode(ISD::ADD, SL, MVT::i32, CttzHi, Bits32);
+
+  // cttz(x) = lo_32(x) == 0 ? cttz(hi_32(x)) + 32 : cttz(lo_32(x))
+  SDValue NewCttz = DAG.getNode(ISD::SELECT, SL, MVT::i32, Hi0, Add, CttzLo);
+
+  if (!ZeroUndef) {
+    // Test if the full 64-bit input is zero.
+
+    // FIXME: DAG combines turn what should be an s_and_b64 into a v_or_b32,
+    // which we probably don't want.
+    SDValue Lo0 = DAG.getSetCC(SL, SetCCVT, Lo, Zero, ISD::SETEQ);
+    SDValue SrcIsZero = DAG.getNode(ISD::AND, SL, SetCCVT, Lo0, Hi0);
+
+    // TODO: If i64 setcc is half rate, it can result in 1 fewer instruction
+    // with the same cycles, otherwise it is slower.
+    // SDValue SrcIsZero = DAG.getSetCC(SL, SetCCVT, Src,
+    // DAG.getConstant(0, SL, MVT::i64), ISD::SETEQ);
+
+    const SDValue Bits32 = DAG.getConstant(64, SL, MVT::i32);
+
+    // The instruction returns -1 for 0 input, but the defined intrinsic
+    // behavior is to return the number of bits.
+    NewCttz = DAG.getNode(ISD::SELECT, SL, MVT::i32,
+                          SrcIsZero, Bits32, NewCttz);
+  }
+
+  return DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i64, NewCttz);   
+}
+
 SDValue AMDGPUTargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
   SDLoc SL(Op);
   SDValue Src = Op.getOperand(0);
@@ -3742,6 +3800,8 @@
   NODE_NAME_CASE(BFM)
   NODE_NAME_CASE(FFBH_U32)
   NODE_NAME_CASE(FFBH_I32)
+  NODE_NAME_CASE(FFBL_U32)
+  NODE_NAME_CASE(FFBL_I32)
   NODE_NAME_CASE(MUL_U24)
   NODE_NAME_CASE(MUL_I24)
   NODE_NAME_CASE(MULHI_U24)
Index: lib/Target/AMDGPU/AMDGPUInstrInfo.td
===================================================================
--- lib/Target/AMDGPU/AMDGPUInstrInfo.td
+++ lib/Target/AMDGPU/AMDGPUInstrInfo.td
@@ -298,6 +298,9 @@
 def AMDGPUffbh_u32 : SDNode<"AMDGPUISD::FFBH_U32", SDTIntUnaryOp>;
 def AMDGPUffbh_i32 : SDNode<"AMDGPUISD::FFBH_I32", SDTIntUnaryOp>;
 
+def AMDGPUffbl_u32 : SDNode<"AMDGPUISD::FFBL_U32", SDTIntUnaryOp>;
+def AMDGPUffbl_i32 : SDNode<"AMDGPUISD::FFBL_I32", SDTIntUnaryOp>;
+
 // Signed and unsigned 24-bit multiply. The highest 8-bits are ignore
 // when performing the mulitply. The result is a 32-bit value.
 def AMDGPUmul_u24 : SDNode<"AMDGPUISD::MUL_U24", SDTIntBinOp,
Index: lib/Target/AMDGPU/EvergreenInstructions.td
===================================================================
--- lib/Target/AMDGPU/EvergreenInstructions.td
+++ lib/Target/AMDGPU/EvergreenInstructions.td
@@ -442,7 +442,7 @@
 def FLT16_TO_FLT32 : R600_1OP_Helper <0xA3, "FLT16_TO_FLT32", f16_to_fp, VecALU>;
 def BCNT_INT : R600_1OP_Helper <0xAA, "BCNT_INT", ctpop, VecALU>;
 def FFBH_UINT : R600_1OP_Helper <0xAB, "FFBH_UINT", AMDGPUffbh_u32, VecALU>;
-def FFBL_INT : R600_1OP_Helper <0xAC, "FFBL_INT", cttz_zero_undef, VecALU>;
+def FFBL_INT : R600_1OP_Helper <0xAC, "FFBL_INT", AMDGPUffbl_u32, VecALU>;
 
 let hasSideEffects = 1 in {
   def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", [], VecALU>;
Index: lib/Target/AMDGPU/SOPInstructions.td
===================================================================
--- lib/Target/AMDGPU/SOPInstructions.td
+++ lib/Target/AMDGPU/SOPInstructions.td
@@ -159,10 +159,11 @@
 
 def S_FF0_I32_B32 : SOP1_32 <"s_ff0_i32_b32">;
 def S_FF0_I32_B64 : SOP1_32_64 <"s_ff0_i32_b64">;
+def S_FF1_I32_B64 : SOP1_32_64 <"s_ff1_i32_b64">;
+
 def S_FF1_I32_B32 : SOP1_32 <"s_ff1_i32_b32",
-  [(set i32:$sdst, (cttz_zero_undef i32:$src0))]
+  [(set i32:$sdst, (AMDGPUffbl_u32 i32:$src0))]
 >;
-def S_FF1_I32_B64 : SOP1_32_64 <"s_ff1_i32_b64">;
 
 def S_FLBIT_I32_B32 : SOP1_32 <"s_flbit_i32_b32",
   [(set i32:$sdst, (AMDGPUffbh_u32 i32:$src0))]
Index: test/CodeGen/AMDGPU/cttz_zero_undef.ll
===================================================================
--- test/CodeGen/AMDGPU/cttz_zero_undef.ll
+++ test/CodeGen/AMDGPU/cttz_zero_undef.ll
@@ -76,3 +76,28 @@
   store <4 x i32> %cttz, <4 x i32> addrspace(1)* %out, align 16
   ret void
 }
+
+; FUNC-LABEL: {{^}}s_cttz_zero_undef_i32_with_select:
+; SI: s_ff1_i32_b32
+; EG: MEM_RAT_CACHELESS STORE_RAW [[RESULT:T[0-9]+\.[XYZW]]]
+; EG: FFBL_INT {{\*? *}}[[RESULT]]
+define amdgpu_kernel void @s_cttz_zero_undef_i32_with_select(i32 addrspace(1)* noalias %out, i32 %val) nounwind {
+  %cttz = tail call i32 @llvm.cttz.i32(i32 %val, i1 true) nounwind readnone
+  %cttz_ret = icmp ne i32 %val, 0
+  %ret = select i1 %cttz_ret, i32 %cttz, i32 32
+  store i32 %cttz, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: {{^}}v_cttz_zero_undef_i32_with_select:
+; SI: v_ffbl_b32_e32
+; EG: MEM_RAT_CACHELESS STORE_RAW [[RESULT:T[0-9]+\.[XYZW]]]
+define amdgpu_kernel void @v_cttz_zero_undef_i32_with_select(i32 addrspace(1)* noalias %out, i32 addrspace(1)* nocapture readonly %arrayidx) nounwind {
+  %val = load i32, i32 addrspace(1)* %arrayidx, align 1
+  %cttz = tail call i32 @llvm.cttz.i32(i32 %val, i1 true) nounwind readnone
+  %cttz_ret = icmp ne i32 %val, 0
+  %ret = select i1 %cttz_ret, i32 %cttz, i32 32
+  store i32 %ret, i32 addrspace(1)* %out, align 4
+  ret void
+}
+