AMDGPU: Pattern match ffbh pattern to instruction.

The hardware instruction's output on 0 is -1 rather than 32.
Eliminate a test and select to -1. This removes an extra instruction
from the compatability function with HSAIL's firstbit instruction.

llvm-svn: 257352

6 files changed, 92 insertions, 24 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 810b325b230..fafe58d65b1 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -282,7 +282,9 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM,
   setOperationAction(ISD::SMAX, MVT::i32, Legal);
   setOperationAction(ISD::UMAX, MVT::i32, Legal);
 
-  if (!Subtarget->hasFFBH())
+  if (Subtarget->hasFFBH())
+    setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Custom);
+  else
     setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
 
   if (!Subtarget->hasFFBL())
@@ -2170,9 +2172,11 @@ SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
 SDValue AMDGPUTargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
   SDLoc SL(Op);
   SDValue Src = Op.getOperand(0);
-  assert(Src.getValueType() == MVT::i64);
-
   bool ZeroUndef = Op.getOpcode() == ISD::CTLZ_ZERO_UNDEF;
+
+  if (ZeroUndef && Src.getValueType() == MVT::i32)
+    return DAG.getNode(AMDGPUISD::FFBH_U32, SL, MVT::i32, Src);
+
   SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
 
   const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
@@ -2507,6 +2511,79 @@ SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
   return DAG.getSExtOrTrunc(Mul, DL, VT);
 }
 
+static bool isNegativeOne(SDValue Val) {
+  if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val))
+    return C->isAllOnesValue();
+  return false;
+}
+
+static bool isCtlzOpc(unsigned Opc) {
+  return Opc == ISD::CTLZ || Opc == ISD::CTLZ_ZERO_UNDEF;
+}
+
+// The native instructions return -1 on 0 input. Optimize out a select that
+// produces -1 on 0.
+//
+// TODO: If zero is not undef, we could also do this if the output is compared
+// against the bitwidth.
+//
+// TODO: Should probably combine against FFBH_U32 instead of ctlz directly.
+SDValue AMDGPUTargetLowering::performCtlzCombine(SDLoc SL,
+                                                 SDValue Cond,
+                                                 SDValue LHS,
+                                                 SDValue RHS,
+                                                 DAGCombinerInfo &DCI) const {
+  ConstantSDNode *CmpRhs = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
+  if (!CmpRhs || !CmpRhs->isNullValue())
+    return SDValue();
+
+  SelectionDAG &DAG = DCI.DAG;
+  ISD::CondCode CCOpcode = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
+  SDValue CmpLHS = Cond.getOperand(0);
+
+  // select (setcc x, 0, eq), -1, (ctlz_zero_undef x) -> ffbh_u32 x
+  if (CCOpcode == ISD::SETEQ &&
+      isCtlzOpc(RHS.getOpcode()) &&
+      RHS.getOperand(0) == CmpLHS &&
+      isNegativeOne(LHS)) {
+    return DAG.getNode(AMDGPUISD::FFBH_U32, SL, MVT::i32, CmpLHS);
+  }
+
+  // select (setcc x, 0, ne), (ctlz_zero_undef x), -1 -> ffbh_u32 x
+  if (CCOpcode == ISD::SETNE &&
+      isCtlzOpc(LHS.getOpcode()) &&
+      LHS.getOperand(0) == CmpLHS &&
+      isNegativeOne(RHS)) {
+    return DAG.getNode(AMDGPUISD::FFBH_U32, SL, MVT::i32, CmpLHS);
+  }
+
+  return SDValue();
+}
+
+SDValue AMDGPUTargetLowering::performSelectCombine(SDNode *N,
+                                                   DAGCombinerInfo &DCI) const {
+  SDValue Cond = N->getOperand(0);
+  if (Cond.getOpcode() != ISD::SETCC)
+    return SDValue();
+
+  EVT VT = N->getValueType(0);
+  SDValue LHS = Cond.getOperand(0);
+  SDValue RHS = Cond.getOperand(1);
+  SDValue CC = Cond.getOperand(2);
+
+  SDValue True = N->getOperand(1);
+  SDValue False = N->getOperand(2);
+
+  if (VT == MVT::f32 && Cond.hasOneUse())
+    return CombineFMinMaxLegacy(SDLoc(N), VT, LHS, RHS, True, False, CC, DCI);
+
+  // There's no reason to not do this if the condition has other uses.
+  if (VT == MVT::i32)
+    return performCtlzCombine(SDLoc(N), Cond, True, False, DCI);
+
+  return SDValue();
+}
+
 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
                                                 DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
@@ -2531,23 +2608,8 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
     simplifyI24(N1, DCI);
     return SDValue();
   }
-  case ISD::SELECT: {
-    SDValue Cond = N->getOperand(0);
-    if (Cond.getOpcode() == ISD::SETCC && Cond.hasOneUse()) {
-      EVT VT = N->getValueType(0);
-      SDValue LHS = Cond.getOperand(0);
-      SDValue RHS = Cond.getOperand(1);
-      SDValue CC = Cond.getOperand(2);
-
-      SDValue True = N->getOperand(1);
-      SDValue False = N->getOperand(2);
-
-      if (VT == MVT::f32)
-        return CombineFMinMaxLegacy(DL, VT, LHS, RHS, True, False, CC, DCI);
-    }
-
-    break;
-  }
+  case ISD::SELECT:
+    return performSelectCombine(N, DCI);
   case AMDGPUISD::BFE_I32:
   case AMDGPUISD::BFE_U32: {
     assert(!N->getValueType(0).isVector() &&
@@ -2759,6 +2821,7 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(BFE_I32)
   NODE_NAME_CASE(BFI)
   NODE_NAME_CASE(BFM)
+  NODE_NAME_CASE(FFBH_U32)
   NODE_NAME_CASE(MUL_U24)
   NODE_NAME_CASE(MUL_I24)
   NODE_NAME_CASE(MAD_U24)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index b53a9f611b4..952fd4cc502 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -69,6 +69,9 @@ private:
   SDValue performStoreCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performShlCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performMulCombine(SDNode *N, DAGCombinerInfo &DCI) const;
+  SDValue performCtlzCombine(SDLoc SL, SDValue Cond, SDValue LHS, SDValue RHS,
+                             DAGCombinerInfo &DCI) const;
+  SDValue performSelectCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 
 protected:
   static EVT getEquivalentMemType(LLVMContext &Context, EVT VT);
@@ -265,6 +268,7 @@ enum NodeType : unsigned {
   BFE_I32, // Extract range of bits with sign extension to 32-bits.
   BFI, // (src0 & src1) | (~src0 & src2)
   BFM, // Insert a range of bits into a 32-bit word.
+  FFBH_U32, // ctlz with -1 if input is zero.
   MUL_U24,
   MUL_I24,
   MAD_U24,
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td b/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
index 70e589c2842..b7a263e34e7 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
@@ -191,6 +191,8 @@ def AMDGPUbfe_i32 : SDNode<"AMDGPUISD::BFE_I32", AMDGPUDTIntTernaryOp>;
 def AMDGPUbfi : SDNode<"AMDGPUISD::BFI", AMDGPUDTIntTernaryOp>;
 def AMDGPUbfm : SDNode<"AMDGPUISD::BFM", SDTIntBinOp>;
 
+def AMDGPUffbh_u32 : SDNode<"AMDGPUISD::FFBH_U32", SDTIntUnaryOp>;
+
 // Signed and unsigned 24-bit mulitply.  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,
diff --git a/llvm/lib/Target/AMDGPU/EvergreenInstructions.td b/llvm/lib/Target/AMDGPU/EvergreenInstructions.td
index 779a14e95d2..2245f1417e5 100644
--- a/llvm/lib/Target/AMDGPU/EvergreenInstructions.td
+++ b/llvm/lib/Target/AMDGPU/EvergreenInstructions.td
@@ -349,7 +349,7 @@ def BCNT_INT : R600_1OP_Helper <0xAA, "BCNT_INT", ctpop, VecALU>;
 def ADDC_UINT : R600_2OP_Helper <0x52, "ADDC_UINT", AMDGPUcarry>;
 def SUBB_UINT : R600_2OP_Helper <0x53, "SUBB_UINT", AMDGPUborrow>;
 
-def FFBH_UINT : R600_1OP_Helper <0xAB, "FFBH_UINT", ctlz_zero_undef, 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>;
 
 let hasSideEffects = 1 in {
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index f7a9880752f..e85d78a69ee 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -2026,7 +2026,7 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
 
   case ISD::UINT_TO_FP: {
     return performUCharToFloatCombine(N, DCI);
-
+  }
   case ISD::FADD: {
     if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
       break;
@@ -2108,7 +2108,6 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
 
     break;
   }
-  }
   case ISD::LOAD:
   case ISD::STORE:
   case ISD::ATOMIC_LOAD:
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index b7df058b7c0..89692ab71f4 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -144,7 +144,7 @@ defm S_FF1_I32_B32 : SOP1_32 <sop1<0x13, 0x10>, "s_ff1_i32_b32",
 defm S_FF1_I32_B64 : SOP1_32_64 <sop1<0x14, 0x11>, "s_ff1_i32_b64", []>;
 
 defm S_FLBIT_I32_B32 : SOP1_32 <sop1<0x15, 0x12>, "s_flbit_i32_b32",
-  [(set i32:$dst, (ctlz_zero_undef i32:$src0))]
+  [(set i32:$dst, (AMDGPUffbh_u32 i32:$src0))]
 >;
 
 defm S_FLBIT_I32_B64 : SOP1_32_64 <sop1<0x16, 0x13>, "s_flbit_i32_b64", []>;