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

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.

Differential Revision: https://reviews.llvm.org/D52978

llvm-svn: 347977
diff --git a/llvm/include/llvm/CodeGen/TargetLowering.h b/llvm/include/llvm/CodeGen/TargetLowering.h
index 8a971f6..6d034e8 100644
--- a/llvm/include/llvm/CodeGen/TargetLowering.h
+++ b/llvm/include/llvm/CodeGen/TargetLowering.h
@@ -2220,6 +2220,12 @@
     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
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
index 2b1df01..b55a732 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
@@ -1063,9 +1063,10 @@
 /// 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:
@@ -1086,8 +1087,8 @@
       NewLHS = OpL;
       NewRHS = OpR;
     } else {
-      NewLHS = ZExtPromotedInteger(NewLHS);
-      NewRHS = ZExtPromotedInteger(NewRHS);
+      NewLHS = SExtOrZExtPromotedInteger(NewLHS);
+      NewRHS = SExtOrZExtPromotedInteger(NewRHS);
     }
     break;
   }
@@ -1095,11 +1096,8 @@
   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:
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h b/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h
index 8b7c57c..5bb5995 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h
@@ -281,6 +281,20 @@
     return DAG.getZeroExtendInReg(Op, dl, OldVT.getScalarType());
   }
 
+  // Get a promoted operand and sign or zero extend it to the final size
+  // (depending on TargetLoweringInfo::isSExtCheaperThanZExt). For a given
+  // subtarget and type, the choice of sign or zero-extension will be
+  // consistent.
+  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);
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index d8c63a4..15fd29e8 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -267,6 +267,10 @@
   return TargetLowering::isZExtFree(Val, VT2);
 }
 
+bool RISCVTargetLowering::isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const {
+  return Subtarget.is64Bit() && 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.
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.h b/llvm/lib/Target/RISCV/RISCVISelLowering.h
index a99e2c2..6970900 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.h
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.h
@@ -54,6 +54,7 @@
   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;
diff --git a/llvm/test/CodeGen/RISCV/alu32.ll b/llvm/test/CodeGen/RISCV/alu32.ll
index 255a528..3c995e6 100644
--- a/llvm/test/CodeGen/RISCV/alu32.ll
+++ b/llvm/test/CodeGen/RISCV/alu32.ll
@@ -51,8 +51,7 @@
 ;
 ; 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
@@ -213,10 +212,8 @@
 ;
 ; 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