[IR] Add samesign flag to icmp instruction (#111419)

Inspired by
https://discourse.llvm.org/t/rfc-signedness-independent-icmps/81423

Author: elhewaty

llvm/docs/LangRef.rst

@@ -12331,6 +12331,7 @@ Syntax:
 ::
 
       <result> = icmp <cond> <ty> <op1>, <op2>   ; yields i1 or <N x i1>:result
+      <result> = icmp samesign <cond> <ty> <op1>, <op2>   ; yields i1 or <N x i1>:result
 
 Overview:
 """""""""
@@ -12400,6 +12401,9 @@ If the operands are integer vectors, then they are compared element by
 element. The result is an ``i1`` vector with the same number of elements
 as the values being compared. Otherwise, the result is an ``i1``.
 
+If the ``samesign`` keyword is present and the operands are not of the
+same sign then the result is a :ref:`poison value <poisonvalues>`.
+
 Example:
 """"""""
 

llvm/include/llvm/AsmParser/LLToken.h

@@ -114,6 +114,7 @@ enum Kind {
   kw_disjoint,
   kw_inbounds,
   kw_nneg,
+  kw_samesign,
   kw_inrange,
   kw_addrspace,
   kw_section,

llvm/include/llvm/Bitcode/LLVMBitCodes.h

@@ -540,6 +540,10 @@ enum GetElementPtrOptionalFlags {
   GEP_NUW = 2,
 };
 
+/// ICmpOptionalFlags - Flags for serializing
+/// ICmpOptionalFlags's SubclassOptionalData contents.
+enum ICmpOptionalFlags { ICMP_SAME_SIGN = 0 };
+
 /// Encoded AtomicOrdering values.
 enum AtomicOrderingCodes {
   ORDERING_NOTATOMIC = 0,

llvm/include/llvm/IR/Instructions.h

@@ -1165,6 +1165,8 @@ class ICmpInst: public CmpInst {
            "Invalid operand types for ICmp instruction");
   }
 
+  enum { SameSign = (1 << 0) };
+
 protected:
   // Note: Instruction needs to be a friend here to call cloneImpl.
   friend class Instruction;
@@ -1224,6 +1226,15 @@ class ICmpInst: public CmpInst {
   /// Return the unsigned version of the predicate.
   static Predicate getUnsignedPredicate(Predicate pred);
 
+  void setSameSign(bool B = true) {
+    SubclassOptionalData = (SubclassOptionalData & ~SameSign) | (B * SameSign);
+  }
+
+  /// An icmp instruction, which can be marked as "samesign", indicating that
+  /// the two operands have the same sign. This means that we can convert
+  /// "slt" to "ult" and vice versa, which enables more optimizations.
+  bool hasSameSign() const { return SubclassOptionalData & SameSign; }
+
   /// Return true if this predicate is either EQ or NE.  This also
   /// tests for commutativity.
   static bool isEquality(Predicate P) {

llvm/include/llvm/Transforms/Utils/ScalarEvolutionExpander.h

@@ -47,6 +47,7 @@ struct PoisonFlags {
   unsigned Exact : 1;
   unsigned Disjoint : 1;
   unsigned NNeg : 1;
+  unsigned SameSign : 1;
   GEPNoWrapFlags GEPNW;
 
   PoisonFlags(const Instruction *I);

llvm/lib/AsmParser/LLLexer.cpp

@@ -583,6 +583,7 @@ lltok::Kind LLLexer::LexIdentifier() {
   KEYWORD(disjoint);
   KEYWORD(inbounds);
   KEYWORD(nneg);
+  KEYWORD(samesign);
   KEYWORD(inrange);
   KEYWORD(addrspace);
   KEYWORD(section);

llvm/lib/AsmParser/LLParser.cpp

@@ -6952,8 +6952,14 @@ int LLParser::parseInstruction(Instruction *&Inst, BasicBlock *BB,
   case lltok::kw_and:
   case lltok::kw_xor:
     return parseLogical(Inst, PFS, KeywordVal);
-  case lltok::kw_icmp:
-    return parseCompare(Inst, PFS, KeywordVal);
+  case lltok::kw_icmp: {
+    bool SameSign = EatIfPresent(lltok::kw_samesign);
+    if (parseCompare(Inst, PFS, KeywordVal))
+      return true;
+    if (SameSign)
+      cast<ICmpInst>(Inst)->setSameSign();
+    return false;
+  }
   case lltok::kw_fcmp: {
     FastMathFlags FMF = EatFastMathFlagsIfPresent();
     int Res = parseCompare(Inst, PFS, KeywordVal);

llvm/lib/Bitcode/Reader/BitcodeReader.cpp

@@ -5471,9 +5471,6 @@ Error BitcodeReader::parseFunctionBody(Function *F) {
       if (IsFP && Record.size() > OpNum+1)
         FMF = getDecodedFastMathFlags(Record[++OpNum]);
 
-      if (OpNum+1 != Record.size())
-        return error("Invalid record");
-
       if (IsFP) {
         if (!CmpInst::isFPPredicate(PredVal))
           return error("Invalid fcmp predicate");
@@ -5482,8 +5479,14 @@ Error BitcodeReader::parseFunctionBody(Function *F) {
         if (!CmpInst::isIntPredicate(PredVal))
           return error("Invalid icmp predicate");
         I = new ICmpInst(PredVal, LHS, RHS);
+        if (Record.size() > OpNum + 1 &&
+            (Record[++OpNum] & (1 << bitc::ICMP_SAME_SIGN)))
+          cast<ICmpInst>(I)->setSameSign();
       }
 
+      if (OpNum + 1 != Record.size())
+        return error("Invalid record");
+
       ResTypeID = getVirtualTypeID(I->getType()->getScalarType());
       if (LHS->getType()->isVectorTy())
         ResTypeID = getVirtualTypeID(I->getType(), ResTypeID);

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp

@@ -1718,6 +1718,9 @@ static uint64_t getOptimizationFlags(const Value *V) {
       Flags |= 1 << bitc::GEP_NUSW;
     if (GEP->hasNoUnsignedWrap())
       Flags |= 1 << bitc::GEP_NUW;
+  } else if (const auto *ICmp = dyn_cast<ICmpInst>(V)) {
+    if (ICmp->hasSameSign())
+      Flags |= 1 << bitc::ICMP_SAME_SIGN;
   }
 
   return Flags;

llvm/lib/IR/AsmWriter.cpp

@@ -1433,6 +1433,9 @@ static void WriteOptimizationInfo(raw_ostream &Out, const User *U) {
       Out << " nuw";
     if (TI->hasNoSignedWrap())
       Out << " nsw";
+  } else if (const auto *ICmp = dyn_cast<ICmpInst>(U)) {
+    if (ICmp->hasSameSign())
+      Out << " samesign";
   }
 }