Patch to implement UMLAL/SMLAL instructions for the ARM architecture

This patch corrects the definition of umlal/smlal instructions and adds support
for matching them to the ARM dag combiner.

Bug 12213

Patch by Yin Ma!


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@161581 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 251 insertions, 17 deletions

diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index ee349a753f..7e8937587b 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -2747,6 +2747,38 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
                                     dl, MVT::i32, MVT::i32, Ops, 5);
     }
   }
+  case ARMISD::UMLAL:{
+    if (Subtarget->isThumb()) {
+      SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
+                        N->getOperand(3), getAL(CurDAG),
+                        CurDAG->getRegister(0, MVT::i32)};
+      return CurDAG->getMachineNode(ARM::t2UMLAL, dl, MVT::i32, MVT::i32, Ops, 6);
+    }else{
+      SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
+                        N->getOperand(3), getAL(CurDAG),
+                        CurDAG->getRegister(0, MVT::i32),
+                        CurDAG->getRegister(0, MVT::i32) };
+      return CurDAG->getMachineNode(Subtarget->hasV6Ops() ?
+                                      ARM::UMLAL : ARM::UMLALv5,
+                                      dl, MVT::i32, MVT::i32, Ops, 7);
+    }
+  }
+  case ARMISD::SMLAL:{
+    if (Subtarget->isThumb()) {
+      SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
+                        N->getOperand(3), getAL(CurDAG),
+                        CurDAG->getRegister(0, MVT::i32)};
+      return CurDAG->getMachineNode(ARM::t2SMLAL, dl, MVT::i32, MVT::i32, Ops, 6);
+    }else{
+      SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
+                        N->getOperand(3), getAL(CurDAG),
+                        CurDAG->getRegister(0, MVT::i32),
+                        CurDAG->getRegister(0, MVT::i32) };
+      return CurDAG->getMachineNode(Subtarget->hasV6Ops() ?
+                                      ARM::SMLAL : ARM::SMLALv5,
+                                      dl, MVT::i32, MVT::i32, Ops, 7);
+    }
+  }
   case ISD::LOAD: {
     SDNode *ResNode = 0;
     if (Subtarget->isThumb() && Subtarget->hasThumb2())
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 7d8222927b..bb2116f774 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -571,6 +571,11 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
     }
   }
 
+  // ARM and Thumb2 support UMLAL/SMLAL.
+  if (!Subtarget->isThumb1Only())
+    setTargetDAGCombine(ISD::ADDC);
+
+
   computeRegisterProperties();
 
   // ARM does not have f32 extending load.
@@ -989,6 +994,8 @@ const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case ARMISD::VTBL2:         return "ARMISD::VTBL2";
   case ARMISD::VMULLs:        return "ARMISD::VMULLs";
   case ARMISD::VMULLu:        return "ARMISD::VMULLu";
+  case ARMISD::UMLAL:         return "ARMISD::UMLAL";
+  case ARMISD::SMLAL:         return "ARMISD::SMLAL";
   case ARMISD::BUILD_VECTOR:  return "ARMISD::BUILD_VECTOR";
   case ARMISD::FMAX:          return "ARMISD::FMAX";
   case ARMISD::FMIN:          return "ARMISD::FMIN";
@@ -7127,6 +7134,154 @@ static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
   return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
 }
 
+static SDValue findMUL_LOHI(SDValue V) {
+  if (V->getOpcode() == ISD::UMUL_LOHI ||
+      V->getOpcode() == ISD::SMUL_LOHI)
+    return V;
+  return SDValue();
+}
+
+static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
+                                     TargetLowering::DAGCombinerInfo &DCI,
+                                     const ARMSubtarget *Subtarget) {
+
+  if (Subtarget->isThumb1Only()) return SDValue();
+
+  // Only perform the checks after legalize when the pattern is available.
+  if (DCI.isBeforeLegalize()) return SDValue();
+
+  // Look for multiply add opportunities.
+  // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
+  // each add nodes consumes a value from ISD::UMUL_LOHI and there is
+  // a glue link from the first add to the second add.
+  // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
+  // a S/UMLAL instruction.
+  //          loAdd   UMUL_LOHI
+  //            \    / :lo    \ :hi
+  //             \  /          \          [no multiline comment]
+  //              ADDC         |  hiAdd
+  //                 \ :glue  /  /
+  //                  \      /  /
+  //                    ADDE
+  //
+  assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
+  SDValue AddcOp0 = AddcNode->getOperand(0);
+  SDValue AddcOp1 = AddcNode->getOperand(1);
+
+  // Check if the two operands are from the same mul_lohi node.
+  if (AddcOp0.getNode() == AddcOp1.getNode())
+    return SDValue();
+
+  assert(AddcNode->getNumValues() == 2 &&
+         AddcNode->getValueType(0) == MVT::i32 &&
+         AddcNode->getValueType(1) == MVT::Glue &&
+         "Expect ADDC with two result values: i32, glue");
+
+  // Check that the ADDC adds the low result of the S/UMUL_LOHI.
+  if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
+      AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
+      AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
+      AddcOp1->getOpcode() != ISD::SMUL_LOHI)
+    return SDValue();
+
+  // Look for the glued ADDE.
+  SDNode* AddeNode = AddcNode->getGluedUser();
+  if (AddeNode == NULL)
+    return SDValue();
+
+  // Make sure it is really an ADDE.
+  if (AddeNode->getOpcode() != ISD::ADDE)
+    return SDValue();
+
+  assert(AddeNode->getNumOperands() == 3 &&
+         AddeNode->getOperand(2).getValueType() == MVT::Glue &&
+         "ADDE node has the wrong inputs");
+
+  // Check for the triangle shape.
+  SDValue AddeOp0 = AddeNode->getOperand(0);
+  SDValue AddeOp1 = AddeNode->getOperand(1);
+
+  // Make sure that the ADDE operands are not coming from the same node.
+  if (AddeOp0.getNode() == AddeOp1.getNode())
+    return SDValue();
+
+  // Find the MUL_LOHI node walking up ADDE's operands.
+  bool IsLeftOperandMUL = false;
+  SDValue MULOp = findMUL_LOHI(AddeOp0);
+  if (MULOp == SDValue())
+   MULOp = findMUL_LOHI(AddeOp1);
+  else
+    IsLeftOperandMUL = true;
+  if (MULOp == SDValue())
+     return SDValue();
+
+  // Figure out the right opcode.
+  unsigned Opc = MULOp->getOpcode();
+  unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
+
+  // Figure out the high and low input values to the MLAL node.
+  SDValue* HiMul = &MULOp;
+  SDValue* HiAdd = NULL;
+  SDValue* LoMul = NULL;
+  SDValue* LowAdd = NULL;
+
+  if (IsLeftOperandMUL)
+    HiAdd = &AddeOp1;
+  else
+    HiAdd = &AddeOp0;
+
+
+  if (AddcOp0->getOpcode() == Opc) {
+    LoMul = &AddcOp0;
+    LowAdd = &AddcOp1;
+  }
+  if (AddcOp1->getOpcode() == Opc) {
+    LoMul = &AddcOp1;
+    LowAdd = &AddcOp0;
+  }
+
+  if (LoMul == NULL)
+    return SDValue();
+
+  if (LoMul->getNode() != HiMul->getNode())
+    return SDValue();
+
+  // Create the merged node.
+  SelectionDAG &DAG = DCI.DAG;
+
+  // Build operand list.
+  SmallVector<SDValue, 8> Ops;
+  Ops.push_back(LoMul->getOperand(0));
+  Ops.push_back(LoMul->getOperand(1));
+  Ops.push_back(*LowAdd);
+  Ops.push_back(*HiAdd);
+
+  SDValue MLALNode =  DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
+                                 DAG.getVTList(MVT::i32, MVT::i32),
+                                 &Ops[0], Ops.size());
+
+  // Replace the ADDs' nodes uses by the MLA node's values.
+  SDValue HiMLALResult(MLALNode.getNode(), 1);
+  DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
+
+  SDValue LoMLALResult(MLALNode.getNode(), 0);
+  DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
+
+  // Return original node to notify the driver to stop replacing.
+  SDValue resNode(AddcNode, 0);
+  return resNode;
+}
+
+/// PerformADDCCombine - Target-specific dag combine transform from
+/// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
+static SDValue PerformADDCCombine(SDNode *N,
+                                 TargetLowering::DAGCombinerInfo &DCI,
+                                 const ARMSubtarget *Subtarget) {
+
+  return AddCombineTo64bitMLAL(N, DCI, Subtarget);
+
+}
+
 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
 /// operands N0 and N1.  This is a helper for PerformADDCombine that is
 /// called with the default operands, and if that fails, with commuted
@@ -8738,6 +8893,7 @@ SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
                                              DAGCombinerInfo &DCI) const {
   switch (N->getOpcode()) {
   default: break;
+  case ISD::ADDC:       return PerformADDCCombine(N, DCI, Subtarget);
   case ISD::ADD:        return PerformADDCombine(N, DCI, Subtarget);
   case ISD::SUB:        return PerformSUBCombine(N, DCI);
   case ISD::MUL:        return PerformMULCombine(N, DCI, Subtarget);
diff --git a/lib/Target/ARM/ARMISelLowering.h b/lib/Target/ARM/ARMISelLowering.h
index 213af20ebd..8faadfd831 100644
--- a/lib/Target/ARM/ARMISelLowering.h
+++ b/lib/Target/ARM/ARMISelLowering.h
@@ -176,6 +176,9 @@ namespace llvm {
       VMULLs,       // ...signed
       VMULLu,       // ...unsigned
 
+      UMLAL,        // 64bit Unsigned Accumulate Multiply
+      SMLAL,        // 64bit Signed Accumulate Multiply
+
       // Operands of the standard BUILD_VECTOR node are not legalized, which
       // is fine if BUILD_VECTORs are always lowered to shuffles or other
       // operations, but for ARM some BUILD_VECTORs are legal as-is and their
diff --git a/lib/Target/ARM/ARMInstrInfo.td b/lib/Target/ARM/ARMInstrInfo.td
index 6340a58f1a..52be43b020 100644
--- a/lib/Target/ARM/ARMInstrInfo.td
+++ b/lib/Target/ARM/ARMInstrInfo.td
@@ -83,6 +83,13 @@ def SDTBinaryArithWithFlagsInOut : SDTypeProfile<2, 3,
                                              SDTCisInt<0>,
                                              SDTCisVT<1, i32>,
                                              SDTCisVT<4, i32>]>;
+
+def SDT_ARM64bitmlal : SDTypeProfile<2,4, [ SDTCisVT<0, i32>, SDTCisVT<1, i32>,
+                                        SDTCisVT<2, i32>, SDTCisVT<3, i32>,
+                                        SDTCisVT<4, i32>, SDTCisVT<5, i32> ] >;
+def ARMUmlal         : SDNode<"ARMISD::UMLAL", SDT_ARM64bitmlal>;
+def ARMSmlal         : SDNode<"ARMISD::SMLAL", SDT_ARM64bitmlal>;
+
 // Node definitions.
 def ARMWrapper       : SDNode<"ARMISD::Wrapper",     SDTIntUnaryOp>;
 def ARMWrapperDYN    : SDNode<"ARMISD::WrapperDYN",  SDTIntUnaryOp>;
@@ -3396,6 +3403,18 @@ class AsMul1I64<bits<7> opcod, dag oops, dag iops, InstrItinClass itin,
   let Inst{11-8}  = Rm;
   let Inst{3-0}   = Rn;
 }
+class AsMla1I64<bits<7> opcod, dag oops, dag iops, InstrItinClass itin,
+             string opc, string asm, list<dag> pattern>
+  : AsMul1I<opcod, oops, iops, itin, opc, asm, pattern> {
+  bits<4> RdLo;
+  bits<4> RdHi;
+  bits<4> Rm;
+  bits<4> Rn;
+  let Inst{19-16} = RdHi;
+  let Inst{15-12} = RdLo;
+  let Inst{11-8}  = Rm;
+  let Inst{3-0}   = Rn;
+}
 
 // FIXME: The v5 pseudos are only necessary for the additional Constraint
 //        property. Remove them when it's possible to add those properties
@@ -3478,14 +3497,14 @@ def UMULLv5 : ARMPseudoExpand<(outs GPR:$RdLo, GPR:$RdHi),
 }
 
 // Multiply + accumulate
-def SMLAL : AsMul1I64<0b0000111, (outs GPR:$RdLo, GPR:$RdHi),
-                               (ins GPR:$Rn, GPR:$Rm), IIC_iMAC64,
+def SMLAL : AsMla1I64<0b0000111, (outs GPR:$RdLo, GPR:$RdHi),
+                        (ins GPR:$Rn, GPR:$Rm, GPR:$RLo, GPR:$RHi), IIC_iMAC64,
                     "smlal", "\t$RdLo, $RdHi, $Rn, $Rm", []>,
-                    Requires<[IsARM, HasV6]>;
-def UMLAL : AsMul1I64<0b0000101, (outs GPR:$RdLo, GPR:$RdHi),
-                               (ins GPR:$Rn, GPR:$Rm), IIC_iMAC64,
+         RegConstraint<"$RLo = $RdLo, $RHi = $RdHi">, Requires<[IsARM, HasV6]>;
+def UMLAL : AsMla1I64<0b0000101, (outs GPR:$RdLo, GPR:$RdHi),
+                        (ins GPR:$Rn, GPR:$Rm, GPR:$RLo, GPR:$RHi), IIC_iMAC64,
                     "umlal", "\t$RdLo, $RdHi, $Rn, $Rm", []>,
-                    Requires<[IsARM, HasV6]>;
+         RegConstraint<"$RLo = $RdLo, $RHi = $RdHi">, Requires<[IsARM, HasV6]>;
 
 def UMAAL : AMul1I <0b0000010, (outs GPR:$RdLo, GPR:$RdHi),
                                (ins GPR:$Rn, GPR:$Rm), IIC_iMAC64,
@@ -3501,17 +3520,22 @@ def UMAAL : AMul1I <0b0000010, (outs GPR:$RdLo, GPR:$RdHi),
   let Inst{3-0}   = Rn;
 }
 
-let Constraints = "@earlyclobber $RdLo,@earlyclobber $RdHi" in {
+let Constraints = "$RLo = $RdLo,$RHi = $RdHi" in {
 def SMLALv5 : ARMPseudoExpand<(outs GPR:$RdLo, GPR:$RdHi),
-                              (ins GPR:$Rn, GPR:$Rm, pred:$p, cc_out:$s),
+                (ins GPR:$Rn, GPR:$Rm, GPR:$RLo, GPR:$RHi, pred:$p, cc_out:$s),
                               4, IIC_iMAC64, [],
-          (SMLAL GPR:$RdLo, GPR:$RdHi, GPR:$Rn, GPR:$Rm, pred:$p, cc_out:$s)>,
+             (SMLAL GPR:$RdLo, GPR:$RdHi, GPR:$Rn, GPR:$Rm, GPR:$RLo, GPR:$RHi,
+                           pred:$p, cc_out:$s)>,
                            Requires<[IsARM, NoV6]>;
 def UMLALv5 : ARMPseudoExpand<(outs GPR:$RdLo, GPR:$RdHi),
-                              (ins GPR:$Rn, GPR:$Rm, pred:$p, cc_out:$s),
+                (ins GPR:$Rn, GPR:$Rm, GPR:$RLo, GPR:$RHi, pred:$p, cc_out:$s),
                               4, IIC_iMAC64, [],
-          (UMLAL GPR:$RdLo, GPR:$RdHi, GPR:$Rn, GPR:$Rm, pred:$p, cc_out:$s)>,
+             (UMLAL GPR:$RdLo, GPR:$RdHi, GPR:$Rn, GPR:$Rm, GPR:$RLo, GPR:$RHi,
+                           pred:$p, cc_out:$s)>,
                            Requires<[IsARM, NoV6]>;
+}
+
+let Constraints = "@earlyclobber $RdLo,@earlyclobber $RdHi" in {
 def UMAALv5 : ARMPseudoExpand<(outs GPR:$RdLo, GPR:$RdHi),
                               (ins GPR:$Rn, GPR:$Rm, pred:$p),
                               4, IIC_iMAC64, [],
diff --git a/lib/Target/ARM/ARMInstrThumb2.td b/lib/Target/ARM/ARMInstrThumb2.td
index 307006f413..e600e8798a 100644
--- a/lib/Target/ARM/ARMInstrThumb2.td
+++ b/lib/Target/ARM/ARMInstrThumb2.td
@@ -523,6 +523,23 @@ class T2MulLong<bits<3> opc22_20, bits<4> opc7_4,
   let Inst{7-4}   = opc7_4;
   let Inst{3-0}   = Rm;
 }
+class T2MlaLong<bits<3> opc22_20, bits<4> opc7_4,
+                dag oops, dag iops, InstrItinClass itin,
+                string opc, string asm, list<dag> pattern>
+  : T2I<oops, iops, itin, opc, asm, pattern> {
+  bits<4> RdLo;
+  bits<4> RdHi;
+  bits<4> Rn;
+  bits<4> Rm;
+
+  let Inst{31-23} = 0b111110111;
+  let Inst{22-20} = opc22_20;
+  let Inst{19-16} = Rn;
+  let Inst{15-12} = RdLo;
+  let Inst{11-8}  = RdHi;
+  let Inst{7-4}   = opc7_4;
+  let Inst{3-0}   = Rm;
+}
 
 
 /// T2I_bin_irs - Defines a set of (op reg, {so_imm|r|so_reg}) patterns for a
@@ -2410,15 +2427,17 @@ def t2UMULL : T2MulLong<0b010, 0b0000,
 } // isCommutable
 
 // Multiply + accumulate
-def t2SMLAL : T2MulLong<0b100, 0b0000,
-                  (outs rGPR:$RdLo, rGPR:$RdHi),
+def t2SMLAL : T2MlaLong<0b100, 0b0000,
+                  (outs rGPR:$RdLo, rGPR:$RdHi, rGPR:$RLo, rGPR:$RHi),
                   (ins rGPR:$Rn, rGPR:$Rm), IIC_iMAC64,
-                  "smlal", "\t$RdLo, $RdHi, $Rn, $Rm", []>;
+                  "smlal", "\t$RdLo, $RdHi, $Rn, $Rm", []>,
+                  RegConstraint<"$RLo = $RdLo, $RHi = $RdHi">;
 
-def t2UMLAL : T2MulLong<0b110, 0b0000,
-                  (outs rGPR:$RdLo, rGPR:$RdHi),
+def t2UMLAL : T2MlaLong<0b110, 0b0000,
+                  (outs rGPR:$RdLo, rGPR:$RdHi, rGPR:$RLo, rGPR:$RHi),
                   (ins rGPR:$Rn, rGPR:$Rm), IIC_iMAC64,
-                  "umlal", "\t$RdLo, $RdHi, $Rn, $Rm", []>;
+                  "umlal", "\t$RdLo, $RdHi, $Rn, $Rm", []>,
+                  RegConstraint<"$RLo = $RdLo, $RHi = $RdHi">;
 
 def t2UMAAL : T2MulLong<0b110, 0b0110,
                   (outs rGPR:$RdLo, rGPR:$RdHi),