[ARM] Extract shifts out of multiply-by-constant

Turning (op x (mul y k)) into (op x (lsl (mul y k>>n) n)) is beneficial when
we can do the lsl as a shifted operand and the resulting multiply constant is
simpler to generate.

Do this by doing the transformation when trying to select a shifted operand,
as that ensures that it actually turns out better (the alternative would be to
do it in PreprocessISelDAG, but we don't know for sure there if extracting the
shift would allow a shifted operand to be used).

Differential Revision: http://reviews.llvm.org/D12196

llvm-svn: 247569

1 files changed, 111 insertions, 19 deletions

diff --git a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 5843d89b578..68724281f30 100644
--- a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -271,6 +271,22 @@ private:
   // Get the alignment operand for a NEON VLD or VST instruction.
   SDValue GetVLDSTAlign(SDValue Align, SDLoc dl, unsigned NumVecs,
                         bool is64BitVector);
+
+  /// Returns the number of instructions required to materialize the given
+  /// constant in a register, or 3 if a literal pool load is needed.
+  unsigned ConstantMaterializationCost(unsigned Val) const;
+
+  /// Checks if N is a multiplication by a constant where we can extract out a
+  /// power of two from the constant so that it can be used in a shift, but only
+  /// if it simplifies the materialization of the constant. Returns true if it
+  /// is, and assigns to PowerOfTwo the power of two that should be extracted
+  /// out and to NewMulConst the new constant to be multiplied by.
+  bool canExtractShiftFromMul(const SDValue &N, unsigned MaxShift,
+                              unsigned &PowerOfTwo, SDValue &NewMulConst) const;
+
+  /// Replace N with M in CurDAG, in a way that also ensures that M gets
+  /// selected when N would have been selected.
+  void replaceDAGValue(const SDValue &N, SDValue M);
 };
 }
 
@@ -464,6 +480,61 @@ bool ARMDAGToDAGISel::isShifterOpProfitable(const SDValue &Shift,
          (ShAmt == 2 || (Subtarget->isSwift() && ShAmt == 1));
 }
 
+unsigned ARMDAGToDAGISel::ConstantMaterializationCost(unsigned Val) const {
+  if (Subtarget->isThumb()) {
+    if (Val <= 255) return 1;                               // MOV
+    if (Subtarget->hasV6T2Ops() && Val <= 0xffff) return 1; // MOVW
+    if (~Val <= 255) return 2;                              // MOV + MVN
+    if (ARM_AM::isThumbImmShiftedVal(Val)) return 2;        // MOV + LSL
+  } else {
+    if (ARM_AM::getSOImmVal(Val) != -1) return 1;           // MOV
+    if (ARM_AM::getSOImmVal(~Val) != -1) return 1;          // MVN
+    if (Subtarget->hasV6T2Ops() && Val <= 0xffff) return 1; // MOVW
+    if (ARM_AM::isSOImmTwoPartVal(Val)) return 2;           // two instrs
+  }
+  if (Subtarget->useMovt(*MF)) return 2; // MOVW + MOVT
+  return 3; // Literal pool load
+}
+
+bool ARMDAGToDAGISel::canExtractShiftFromMul(const SDValue &N,
+                                             unsigned MaxShift,
+                                             unsigned &PowerOfTwo,
+                                             SDValue &NewMulConst) const {
+  assert(N.getOpcode() == ISD::MUL);
+  assert(MaxShift > 0);
+
+  // If the multiply is used in more than one place then changing the constant
+  // will make other uses incorrect, so don't.
+  if (!N.hasOneUse()) return false;
+  // Check if the multiply is by a constant
+  ConstantSDNode *MulConst = dyn_cast<ConstantSDNode>(N.getOperand(1));
+  if (!MulConst) return false;
+  // If the constant is used in more than one place then modifying it will mean
+  // we need to materialize two constants instead of one, which is a bad idea.
+  if (!MulConst->hasOneUse()) return false;
+  unsigned MulConstVal = MulConst->getZExtValue();
+  if (MulConstVal == 0) return false;
+
+  // Find the largest power of 2 that MulConstVal is a multiple of
+  PowerOfTwo = MaxShift;
+  while ((MulConstVal % (1 << PowerOfTwo)) != 0) {
+    --PowerOfTwo;
+    if (PowerOfTwo == 0) return false;
+  }
+
+  // Only optimise if the new cost is better
+  unsigned NewMulConstVal = MulConstVal / (1 << PowerOfTwo);
+  NewMulConst = CurDAG->getConstant(NewMulConstVal, SDLoc(N), MVT::i32);
+  unsigned OldCost = ConstantMaterializationCost(MulConstVal);
+  unsigned NewCost = ConstantMaterializationCost(NewMulConstVal);
+  return NewCost < OldCost;
+}
+
+void ARMDAGToDAGISel::replaceDAGValue(const SDValue &N, SDValue M) {
+  CurDAG->RepositionNode(N.getNode(), M.getNode());
+  CurDAG->ReplaceAllUsesWith(N, M);
+}
+
 bool ARMDAGToDAGISel::SelectImmShifterOperand(SDValue N,
                                               SDValue &BaseReg,
                                               SDValue &Opc,
@@ -471,6 +542,21 @@ bool ARMDAGToDAGISel::SelectImmShifterOperand(SDValue N,
   if (DisableShifterOp)
     return false;
 
+  // If N is a multiply-by-constant and it's profitable to extract a shift and
+  // use it in a shifted operand do so.
+  if (N.getOpcode() == ISD::MUL) {
+    unsigned PowerOfTwo = 0;
+    SDValue NewMulConst;
+    if (canExtractShiftFromMul(N, 31, PowerOfTwo, NewMulConst)) {
+      replaceDAGValue(N.getOperand(1), NewMulConst);
+      BaseReg = N;
+      Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ARM_AM::lsl,
+                                                          PowerOfTwo),
+                                      SDLoc(N), MVT::i32);
+      return true;
+    }
+  }
+
   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
 
   // Don't match base register only case. That is matched to a separate
@@ -655,6 +741,18 @@ bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset,
     }
   }
 
+  // If Offset is a multiply-by-constant and it's profitable to extract a shift
+  // and use it in a shifted operand do so.
+  if (Offset.getOpcode() == ISD::MUL) {
+    unsigned PowerOfTwo = 0;
+    SDValue NewMulConst;
+    if (canExtractShiftFromMul(Offset, 31, PowerOfTwo, NewMulConst)) {
+      replaceDAGValue(Offset.getOperand(1), NewMulConst);
+      ShAmt = PowerOfTwo;
+      ShOpcVal = ARM_AM::lsl;
+    }
+  }
+
   Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
                                   SDLoc(N), MVT::i32);
   return true;
@@ -1314,6 +1412,17 @@ bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue N,
     }
   }
 
+  // If OffReg is a multiply-by-constant and it's profitable to extract a shift
+  // and use it in a shifted operand do so.
+  if (OffReg.getOpcode() == ISD::MUL) {
+    unsigned PowerOfTwo = 0;
+    SDValue NewMulConst;
+    if (canExtractShiftFromMul(OffReg, 3, PowerOfTwo, NewMulConst)) {
+      replaceDAGValue(OffReg.getOperand(1), NewMulConst);
+      ShAmt = PowerOfTwo;
+    }
+  }
+
   ShImm = CurDAG->getTargetConstant(ShAmt, SDLoc(N), MVT::i32);
 
   return true;
@@ -2392,25 +2501,8 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
   }
   case ISD::Constant: {
     unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
-    bool UseCP = true;
-    if (Subtarget->useMovt(*MF))
-      // Thumb2-aware targets have the MOVT instruction, so all immediates can
-      // be done with MOV + MOVT, at worst.
-      UseCP = false;
-    else {
-      if (Subtarget->isThumb()) {
-        UseCP = (Val > 255 &&                                  // MOV
-                 ~Val > 255 &&                                 // MOV + MVN
-                 !ARM_AM::isThumbImmShiftedVal(Val) &&         // MOV + LSL
-                 !(Subtarget->hasV6T2Ops() && Val <= 0xffff)); // MOVW
-      } else
-        UseCP = (ARM_AM::getSOImmVal(Val) == -1 &&             // MOV
-                 ARM_AM::getSOImmVal(~Val) == -1 &&            // MVN
-                 !ARM_AM::isSOImmTwoPartVal(Val) &&            // two instrs.
-                 !(Subtarget->hasV6T2Ops() && Val <= 0xffff)); // MOVW
-    }
-
-    if (UseCP) {
+    // If we can't materialize the constant we need to use a literal pool
+    if (ConstantMaterializationCost(Val) > 2) {
       SDValue CPIdx = CurDAG->getTargetConstantPool(
           ConstantInt::get(Type::getInt32Ty(*CurDAG->getContext()), Val),
           TLI->getPointerTy(CurDAG->getDataLayout()));