[ARM] Invert CSEL predicates if the opposite is a simpler constant to materialise

[dmgreen]

Summary

This moves ConstantMaterializationCost into ARMBaseInstrInfo so that it can also be used in ISel Lowering, adding codesize values to the computed costs, to be able to compare either approximate instruction counts or codesize costs.

It also adds a HasLowerConstantMaterializationCost, which compares the ConstantMaterializationCost of two values, returning true if the first is smaller either in instruction count/codesize, or falling back to the other in the case that they are equal.

This is used in constant CSEL lowering to invert the predicate if the opposite is easier to materialise.

[dmgreen]

ping

[samparker]

And I thought that I had already move this into ARMISelLowering... LGTM.

llvm/lib/Target/ARM/ARMBaseInstrInfo.h

 
 void addUnpredicatedMveVpredNOp(MachineInstrBuilder &MIB);
 void addUnpredicatedMveVpredROp(MachineInstrBuilder &MIB, unsigned DestReg);
 
 void addPredicatedMveVpredNOp(MachineInstrBuilder &MIB, unsigned Cond);
 void addPredicatedMveVpredROp(MachineInstrBuilder &MIB, unsigned Cond,
                               unsigned Inactive);
 
+/// Returns the number of instructions required to materialize the given
+/// constant in a register, or 3 if a literal pool load is needed.
+/// If ForCodesize is specified, an approximate cost in bytes is returned.
+unsigned ConstantMaterializationCost(unsigned Val,
+                                     const ARMSubtarget *Subtarget,
+                                     bool ForCodesize = false);
+
+/// Returns true if Val1 has a lower Constant Materialization Cost than Val2.
+/// Uses the cost from ConstantMaterializationCost, first with ForCodesize as
+/// specified. If the scores are equal, return the comparison for !ForCodesize.
+bool HasLowerConstantMaterializationCost(unsigned Val1, unsigned Val2,
+                                         const ARMSubtarget *Subtarget,
+                                         bool ForCodesize = false);
+
 } // end namespace llvm
 
 #endif // LLVM_LIB_TARGET_ARM_ARMBASEINSTRINFO_H

llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp

if (Pred != ARMCC::AL || CmpMI->getOperand(1).getImm() != 0)
     return nullptr;
   if (!isARMLowRegister(Reg))
     return nullptr;
   if (registerDefinedBetween(Reg, CmpMI->getNextNode(), Br, TRI))
     return nullptr;
 
   return &*CmpMI;
 }
+
+unsigned llvm::ConstantMaterializationCost(unsigned Val,
+                                           const ARMSubtarget *Subtarget,
+                                           bool ForCodesize) {
+  if (Subtarget->isThumb()) {
+    if (Val <= 255) // MOV
+      return ForCodesize ? 2 : 1;
+    if (Subtarget->hasV6T2Ops() && (Val <= 0xffff ||                    // MOV
+                                    ARM_AM::getT2SOImmVal(Val) != -1 || // MOVW
+                                    ARM_AM::getT2SOImmVal(~Val) != -1)) // MVN
+      return ForCodesize ? 4 : 1;
+    if (Val <= 510) // MOV + ADDi8
+      return ForCodesize ? 4 : 2;
+    if (~Val <= 255) // MOV + MVN
+      return ForCodesize ? 4 : 2;
+    if (ARM_AM::isThumbImmShiftedVal(Val)) // MOV + LSL
+      return ForCodesize ? 4 : 2;
+  } else {
+    if (ARM_AM::getSOImmVal(Val) != -1) // MOV
+      return ForCodesize ? 4 : 1;
+    if (ARM_AM::getSOImmVal(~Val) != -1) // MVN
+      return ForCodesize ? 4 : 1;
+    if (Subtarget->hasV6T2Ops() && Val <= 0xffff) // MOVW
+      return ForCodesize ? 4 : 1;
+    if (ARM_AM::isSOImmTwoPartVal(Val)) // two instrs
+      return ForCodesize ? 8 : 2;
+  }
+  if (Subtarget->useMovt()) // MOVW + MOVT
+    return ForCodesize ? 8 : 2;
+  return ForCodesize ? 8 : 3; // Literal pool load
+}
+
+bool llvm::HasLowerConstantMaterializationCost(unsigned Val1, unsigned Val2,
+                                               const ARMSubtarget *Subtarget,
+                                               bool ForCodesize) {
+  // Check with ForCodesize
+  unsigned Cost1 = ConstantMaterializationCost(Val1, Subtarget, ForCodesize);
+  unsigned Cost2 = ConstantMaterializationCost(Val2, Subtarget, ForCodesize);
+  if (Cost1 < Cost2)
+    return true;
+  if (Cost1 > Cost2)
+    return false;
+
+  // If they are equal, try with !ForCodesize
+  return ConstantMaterializationCost(Val1, Subtarget, !ForCodesize) <
+         ConstantMaterializationCost(Val2, Subtarget, !ForCodesize);
+}

llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp

private:
   SDNode *createQuadSRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
   SDNode *createQuadDRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
   SDNode *createQuadQRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
 
   // Get the alignment operand for a NEON VLD or VST instruction.
   SDValue GetVLDSTAlign(SDValue Align, const 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;
 
if (!Subtarget->isLikeA9() && !Subtarget->isSwift())
     return true;
   if (Shift.hasOneUse())
     return true;
   // R << 2 is free.
   return ShOpcVal == ARM_AM::lsl &&
          (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 ||                                   // MOV
-         ARM_AM::getT2SOImmVal(Val) != -1 ||                // MOVW
-         ARM_AM::getT2SOImmVal(~Val) != -1))                // MVN
-      return 1;
-    if (Val <= 510) return 2;                               // MOV + ADDi8
-    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()) 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
bool ARMDAGToDAGISel::canExtractShiftFromMul(const SDValue &N,
   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);
+  unsigned OldCost = ConstantMaterializationCost(MulConstVal, Subtarget);
+  unsigned NewCost = ConstantMaterializationCost(NewMulConstVal, Subtarget);
   return NewCost < OldCost;
 }
 
 void ARMDAGToDAGISel::replaceDAGValue(const SDValue &N, SDValue M) {
   CurDAG->RepositionNode(N.getNode()->getIterator(), M.getNode());
   ReplaceUses(N, M);
 }
 
case ISD::XOR:
     // Select special operations if XOR node forms integer ABS pattern
     if (tryABSOp(N))
       return;
     // Other cases are autogenerated.
     break;
   case ISD::Constant: {
     unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
     // If we can't materialize the constant we need to use a literal pool
-    if (ConstantMaterializationCost(Val) > 2) {
+    if (ConstantMaterializationCost(Val, Subtarget) > 2) {
       SDValue CPIdx = CurDAG->getTargetConstantPool(
           ConstantInt::get(Type::getInt32Ty(*CurDAG->getContext()), Val),
           TLI->getPointerTy(CurDAG->getDataLayout()));
 
       SDNode *ResNode;
       if (Subtarget->isThumb()) {
         SDValue Ops[] = {
           CPIdx,
if (N1C && N1C->hasOneUse() && Subtarget->isThumb()) {
 
       // In Thumb2 mode, an AND can take a 12-bit immediate. If this
       // immediate can be negated and fit in the immediate operand of
       // a t2BIC, don't do any manual transform here as this can be
       // handled by the generic ISel machinery.
       bool PreferImmediateEncoding =
         Subtarget->hasThumb2() && (is_t2_so_imm(Imm) || is_t2_so_imm_not(Imm));
       if (!PreferImmediateEncoding &&
-          ConstantMaterializationCost(Imm) >
-              ConstantMaterializationCost(~Imm)) {
+          ConstantMaterializationCost(Imm, Subtarget) >
+              ConstantMaterializationCost(~Imm, Subtarget)) {
         // The current immediate costs more to materialize than a negated
         // immediate, so negate the immediate and use a BIC.
         SDValue NewImm =
           CurDAG->getConstant(~N1C->getZExtValue(), dl, MVT::i32);
         // If the new constant didn't exist before, reposition it in the topological
         // ordering so it is just before N. Otherwise, don't touch its location.
         if (NewImm->getNodeId() == -1)
           CurDAG->RepositionNode(N->getIterator(), NewImm.getNode());

llvm/lib/Target/ARM/ARMISelLowering.cpp

if (CFVal && CTVal) {
       } else if (TVal == FVal + 1) {
         Opcode = ARMISD::CSINC;
         std::swap(TrueVal, FalseVal);
         std::swap(TVal, FVal);
         CC = ISD::getSetCCInverse(CC, true);
       }
 
       if (Opcode) {
-        // Drops F's value because we can get it by inverting/negating TVal.
-        FalseVal = TrueVal;
+        // If one of the constants is cheaper than another, materialise the
+        // cheaper one and let the csel generate the other.
+        if (Opcode != ARMISD::CSINC &&
+            HasLowerConstantMaterializationCost(FVal, TVal, Subtarget)) {
+          std::swap(TrueVal, FalseVal);
+          std::swap(TVal, FVal);
+          CC = ISD::getSetCCInverse(CC, true);
+        }
 
         // Attempt to use ZR checking TVal is 0, possibly inverting the
         // condition to get there.
         if (FVal == 0 && Opcode != ARMISD::CSINC) {
           std::swap(TrueVal, FalseVal);
           std::swap(TVal, FVal);
           CC = ISD::getSetCCInverse(CC, true);
         }
-        if (TVal == 0) {
+        if (TVal == 0)
           TrueVal = DAG.getRegister(ARM::ZR, MVT::i32);
-          FalseVal = DAG.getRegister(ARM::ZR, MVT::i32);
-        }
+
+        // Drops F's value because we can get it by inverting/negating TVal.
+        FalseVal = TrueVal;
 
         SDValue ARMcc;
         SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
         EVT VT = TrueVal.getValueType();
         return DAG.getNode(Opcode, dl, VT, TrueVal, FalseVal, ARMcc, Cmp);
       }
     }
   }

llvm/test/CodeGen/Thumb2/csel.ll

entry:
   %cmp = icmp sgt i32 %a, 45
   %spec.select = zext i1 %cmp to i32
   ret i32 %spec.select
 }
 
 define i32 @csinv_const_56(i32 %a) {
 ; CHECK-LABEL: csinv_const_56:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    mvn r1, #5
+; CHECK-NEXT:    movs r1, #5
 ; CHECK-NEXT:    cmp r0, #45
-; CHECK-NEXT:    csinv r0, r1, r1, gt
+; CHECK-NEXT:    csinv r0, r1, r1, le
 ; CHECK-NEXT:    bx lr
 entry:
   %cmp = icmp sgt i32 %a, 45
   %spec.select = select i1 %cmp, i32 -6, i32 5
   ret i32 %spec.select
 }
 
 define i32 @csinv_const_65(i32 %a) {
entry:
   %cmp = icmp sgt i32 %a, 45
   %spec.select = select i1 %cmp, i32 1, i32 -1
   ret i32 %spec.select
 }
 
 define i32 @csneg_const_r(i32 %a) {
 ; CHECK-LABEL: csneg_const_r:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    mov.w r1, #-1
+; CHECK-NEXT:    movs r1, #1
 ; CHECK-NEXT:    cmp r0, #45
-; CHECK-NEXT:    csneg r0, r1, r1, gt
+; CHECK-NEXT:    csneg r0, r1, r1, le
 ; CHECK-NEXT:    bx lr
 entry:
   %cmp = icmp sgt i32 %a, 45
   %spec.select = select i1 %cmp, i32 -1, i32 1
   ret i32 %spec.select
 }
 
 define i32 @csel_var(i32 %a, i32 %b, i32 %c) {