[TargetLowering][RISCV] Introduce isSExtCheaperThanZExt hook and implement for RISC-V

[asb]

Summary

DAGTypeLegalizer::PromoteSetCCOperands currently prefers to zero-extend operands when it is able to do so. For some targets this is more expensive than a sign-extension, which is also a valid choice. Introduce the isSExtCheaperThanZExt hook and use it in the new SExtOrZExtPromotedInteger helper. On RISC-V, we prefer sign-extension for FromTy == MVT::i32 and ToTy == MVT::i64, as it can be performed using a single instruction.

[asb]

Update to attache the correct diff this time.

[lewis-revill]

lib/Target/RISCV/RISCVISelLowering.cpp

274

Should

if (!Subtarget.is64Bit())
    return false;

be added here? i64 shouldn't make it this far on 32 bit but is it better to be explicit?

[efriedma]

Makes sense.

lib/CodeGen/SelectionDAG/LegalizeTypes.h

284

This probably needs a doc-comment explaining that for a given subtarget and type, the choice of either sign-extension or zero-extension is consistent.

[asb]

Update to address review comments (thanks!).

[efriedma]

LGTM

include/llvm/CodeGen/TargetLowering.h

public:
   virtual bool isZExtFree(Type *FromTy, Type *ToTy) const {
     return false;
   }
 
   virtual bool isZExtFree(EVT FromTy, EVT ToTy) const {
     return false;
   }
 
+  /// Return true if sign-extension from FromTy to ToTy is cheaper than
+  /// zero-extension.
+  virtual bool isSExtCheaperThanZExt(EVT FromTy, EVT ToTy) const {
+    return false;
+  }
+
   /// Return true if the target supplies and combines to a paired load
   /// two loaded values of type LoadedType next to each other in memory.
   /// RequiredAlignment gives the minimal alignment constraints that must be met
   /// to be able to select this paired load.
   ///
   /// This information is *not* used to generate actual paired loads, but it is
   /// used to generate a sequence of loads that is easier to combine into a
   /// paired load.

lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

bool DAGTypeLegalizer::PromoteIntegerOperand(SDNode *N, unsigned OpNo) {
   ReplaceValueWith(SDValue(N, 0), Res);
   return false;
 }
 
 /// PromoteSetCCOperands - Promote the operands of a comparison.  This code is
 /// shared among BR_CC, SELECT_CC, and SETCC handlers.
 void DAGTypeLegalizer::PromoteSetCCOperands(SDValue &NewLHS,SDValue &NewRHS,
                                             ISD::CondCode CCCode) {
-  // We have to insert explicit sign or zero extends.  Note that we could
-  // insert sign extends for ALL conditions, but zero extend is cheaper on
-  // many machines (an AND instead of two shifts), so prefer it.
+  // We have to insert explicit sign or zero extends. Note that we could
+  // insert sign extends for ALL conditions. For those operations where either
+  // zero or sign extension would be valid, use SExtOrZExtPromotedInteger
+  // which will choose the cheapest for the target.
   switch (CCCode) {
   default: llvm_unreachable("Unknown integer comparison!");
   case ISD::SETEQ:
   case ISD::SETNE: {
     SDValue OpL = GetPromotedInteger(NewLHS);
     SDValue OpR = GetPromotedInteger(NewRHS);
 
     // We would prefer to promote the comparison operand with sign extension.
     // If the width of OpL/OpR excluding the duplicated sign bits is no greater
     // than the width of NewLHS/NewRH, we can avoid inserting real truncate
     // instruction, which is redundant eventually.
     unsigned OpLEffectiveBits =
         OpL.getScalarValueSizeInBits() - DAG.ComputeNumSignBits(OpL) + 1;
     unsigned OpREffectiveBits =
         OpR.getScalarValueSizeInBits() - DAG.ComputeNumSignBits(OpR) + 1;
     if (OpLEffectiveBits <= NewLHS.getScalarValueSizeInBits() &&
         OpREffectiveBits <= NewRHS.getScalarValueSizeInBits()) {
       NewLHS = OpL;
       NewRHS = OpR;
     } else {
-      NewLHS = ZExtPromotedInteger(NewLHS);
-      NewRHS = ZExtPromotedInteger(NewRHS);
+      NewLHS = SExtOrZExtPromotedInteger(NewLHS);
+      NewRHS = SExtOrZExtPromotedInteger(NewRHS);
     }
     break;
   }
   case ISD::SETUGE:
   case ISD::SETUGT:
   case ISD::SETULE:
   case ISD::SETULT:
-    // ALL of these operations will work if we either sign or zero extend
-    // the operands (including the unsigned comparisons!).  Zero extend is
-    // usually a simpler/cheaper operation, so prefer it.
-    NewLHS = ZExtPromotedInteger(NewLHS);
-    NewRHS = ZExtPromotedInteger(NewRHS);
+    NewLHS = SExtOrZExtPromotedInteger(NewLHS);
+    NewRHS = SExtOrZExtPromotedInteger(NewRHS);
     break;
   case ISD::SETGE:
   case ISD::SETGT:
   case ISD::SETLT:
   case ISD::SETLE:
     NewLHS = SExtPromotedInteger(NewLHS);
     NewRHS = SExtPromotedInteger(NewRHS);
     break;

lib/CodeGen/SelectionDAG/LegalizeTypes.h

private:
   /// Get a promoted operand and zero extend it to the final size.
   SDValue ZExtPromotedInteger(SDValue Op) {
     EVT OldVT = Op.getValueType();
     SDLoc dl(Op);
     Op = GetPromotedInteger(Op);
     return DAG.getZeroExtendInReg(Op, dl, OldVT.getScalarType());
   }
 
+  SDValue SExtOrZExtPromotedInteger(SDValue Op) {
+    EVT OldVT = Op.getValueType();
+    SDLoc DL(Op);
+    Op = GetPromotedInteger(Op);
+    if (TLI.isSExtCheaperThanZExt(OldVT, Op.getValueType()))
+      return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, Op.getValueType(), Op,
+                         DAG.getValueType(OldVT));
+    return DAG.getZeroExtendInReg(Op, DL, OldVT.getScalarType());
+  }
+
   // Integer Result Promotion.
   void PromoteIntegerResult(SDNode *N, unsigned ResNo);
   SDValue PromoteIntRes_MERGE_VALUES(SDNode *N, unsigned ResNo);
   SDValue PromoteIntRes_AssertSext(SDNode *N);
   SDValue PromoteIntRes_AssertZext(SDNode *N);
   SDValue PromoteIntRes_Atomic0(AtomicSDNode *N);
   SDValue PromoteIntRes_Atomic1(AtomicSDNode *N);
   SDValue PromoteIntRes_AtomicCmpSwap(AtomicSDNode *N, unsigned ResNo);

lib/Target/RISCV/RISCVISelLowering.h

public:
   bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
                              unsigned AS,
                              Instruction *I = nullptr) const override;
   bool isLegalICmpImmediate(int64_t Imm) const override;
   bool isLegalAddImmediate(int64_t Imm) const override;
   bool isTruncateFree(Type *SrcTy, Type *DstTy) const override;
   bool isTruncateFree(EVT SrcVT, EVT DstVT) const override;
   bool isZExtFree(SDValue Val, EVT VT2) const override;
+  bool isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const override;
 
   // Provide custom lowering hooks for some operations.
   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
 
   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
 
   // This method returns the name of a target specific DAG node.
   const char *getTargetNodeName(unsigned Opcode) const override;

lib/Target/RISCV/RISCVISelLowering.cpp

if ((MemVT == MVT::i8 || MemVT == MVT::i16 ||
         (LD->getExtensionType() == ISD::NON_EXTLOAD ||
          LD->getExtensionType() == ISD::ZEXTLOAD))
       return true;
   }
 
   return TargetLowering::isZExtFree(Val, VT2);
 }
 
+bool RISCVTargetLowering::isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const {
+  return SrcVT == MVT::i32 && DstVT == MVT::i64;
+}
+
 // Changes the condition code and swaps operands if necessary, so the SetCC
 // operation matches one of the comparisons supported directly in the RISC-V
 // ISA.
 static void normaliseSetCC(SDValue &LHS, SDValue &RHS, ISD::CondCode &CC) {
   switch (CC) {
   default:
     break;
   case ISD::SETGT:

test/CodeGen/RISCV/alu32.ll

 define i32 @sltiu(i32 %a) nounwind {
 ; RV32I-LABEL: sltiu:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    sltiu a0, a0, 3
 ; RV32I-NEXT:    ret
 ;
 ; RV64I-LABEL: sltiu:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    slli a0, a0, 32
-; RV64I-NEXT:    srli a0, a0, 32
+; RV64I-NEXT:    sext.w a0, a0
 ; RV64I-NEXT:    sltiu a0, a0, 3
 ; RV64I-NEXT:    ret
   %1 = icmp ult i32 %a, 3
   %2 = zext i1 %1 to i32
   ret i32 %2
 }
 
 define i32 @xori(i32 %a) nounwind {

 define i32 @sltu(i32 %a, i32 %b) nounwind {
 ; RV32I-LABEL: sltu:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    sltu a0, a0, a1
 ; RV32I-NEXT:    ret
 ;
 ; RV64I-LABEL: sltu:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    slli a1, a1, 32
-; RV64I-NEXT:    srli a1, a1, 32
-; RV64I-NEXT:    slli a0, a0, 32
-; RV64I-NEXT:    srli a0, a0, 32
+; RV64I-NEXT:    sext.w a1, a1
+; RV64I-NEXT:    sext.w a0, a0
 ; RV64I-NEXT:    sltu a0, a0, a1
 ; RV64I-NEXT:    ret
   %1 = icmp ult i32 %a, %b
   %2 = zext i1 %1 to i32
   ret i32 %2
 }
 
 define i32 @xor(i32 %a, i32 %b) nounwind {