[IR] Add disjoint flag for Or instructions. (#72583)

This flag indicates that every bit is known to be zero in at least one
of the inputs. This allows the Or to be treated as an Add since there is
no possibility of a carry from any bit.

If the flag is present and this property does not hold, the result is
poison.

This makes it easier to reverse the InstCombine transform that turns Add
into Or.

This is inspired by a comment here
#71955 (comment)

Discourse thread
https://discourse.llvm.org/t/rfc-add-or-disjoint-flag/75036

Author: topperc

llvm/docs/LangRef.rst

@@ -9981,6 +9981,7 @@ Syntax:
 ::
 
       <result> = or <ty> <op1>, <op2>   ; yields ty:result
+      <result> = or disjoint <ty> <op1>, <op2>   ; yields ty:result
 
 Overview:
 """""""""
@@ -10012,6 +10013,12 @@ The truth table used for the '``or``' instruction is:
 |   1 |   1 |   1 |
 +-----+-----+-----+
 
+``disjoint`` means that for each bit, that bit is zero in at least one of the
+inputs. This allows the Or to be treated as an Add since no carry can occur from
+any bit. If the disjoint keyword is present, the result value of the ``or`` is a
+:ref:`poison value <poisonvalues>` if both inputs have a one in the same bit
+position. For vectors, only the element containing the bit is poison.
+
 Example:
 """"""""
 

llvm/include/llvm/AsmParser/LLToken.h

@@ -109,6 +109,7 @@ enum Kind {
   kw_nuw,
   kw_nsw,
   kw_exact,
+  kw_disjoint,
   kw_inbounds,
   kw_nneg,
   kw_inrange,

llvm/include/llvm/Bitcode/LLVMBitCodes.h

@@ -512,6 +512,10 @@ enum PossiblyNonNegInstOptionalFlags { PNNI_NON_NEG = 0 };
 /// PossiblyExactOperator's SubclassOptionalData contents.
 enum PossiblyExactOperatorOptionalFlags { PEO_EXACT = 0 };
 
+/// PossiblyDisjointInstOptionalFlags - Flags for serializing
+/// PossiblyDisjointInst's SubclassOptionalData contents.
+enum PossiblyDisjointInstOptionalFlags { PDI_DISJOINT = 0 };
+
 /// Encoded AtomicOrdering values.
 enum AtomicOrderingCodes {
   ORDERING_NOTATOMIC = 0,

llvm/include/llvm/IR/InstrTypes.h

@@ -415,6 +415,29 @@ struct OperandTraits<BinaryOperator> :
 
 DEFINE_TRANSPARENT_OPERAND_ACCESSORS(BinaryOperator, Value)
 
+/// An or instruction, which can be marked as "disjoint", indicating that the
+/// inputs don't have a 1 in the same bit position. Meaning this instruction
+/// can also be treated as an add.
+class PossiblyDisjointInst : public BinaryOperator {
+public:
+  enum { IsDisjoint = (1 << 0) };
+
+  void setIsDisjoint(bool B) {
+    SubclassOptionalData =
+        (SubclassOptionalData & ~IsDisjoint) | (B * IsDisjoint);
+  }
+
+  bool isDisjoint() const { return SubclassOptionalData & IsDisjoint; }
+
+  static bool classof(const Instruction *I) {
+    return I->getOpcode() == Instruction::Or;
+  }
+
+  static bool classof(const Value *V) {
+    return isa<Instruction>(V) && classof(cast<Instruction>(V));
+  }
+};
+
 //===----------------------------------------------------------------------===//
 //                               CastInst Class
 //===----------------------------------------------------------------------===//

llvm/lib/AsmParser/LLLexer.cpp

@@ -564,6 +564,7 @@ lltok::Kind LLLexer::LexIdentifier() {
   KEYWORD(nuw);
   KEYWORD(nsw);
   KEYWORD(exact);
+  KEYWORD(disjoint);
   KEYWORD(inbounds);
   KEYWORD(nneg);
   KEYWORD(inrange);

llvm/lib/AsmParser/LLParser.cpp

@@ -6370,8 +6370,15 @@ int LLParser::parseInstruction(Instruction *&Inst, BasicBlock *BB,
   case lltok::kw_srem:
     return parseArithmetic(Inst, PFS, KeywordVal,
                            /*IsFP*/ false);
+  case lltok::kw_or: {
+    bool Disjoint = EatIfPresent(lltok::kw_disjoint);
+    if (parseLogical(Inst, PFS, KeywordVal))
+      return true;
+    if (Disjoint)
+      cast<PossiblyDisjointInst>(Inst)->setIsDisjoint(true);
+    return false;
+  }
   case lltok::kw_and:
-  case lltok::kw_or:
   case lltok::kw_xor:
     return parseLogical(Inst, PFS, KeywordVal);
   case lltok::kw_icmp:

llvm/lib/Bitcode/Reader/BitcodeReader.cpp

@@ -4866,12 +4866,14 @@ Error BitcodeReader::parseFunctionBody(Function *F) {
                    Opc == Instruction::AShr) {
           if (Record[OpNum] & (1 << bitc::PEO_EXACT))
             cast<BinaryOperator>(I)->setIsExact(true);
+        } else if (Opc == Instruction::Or) {
+          if (Record[OpNum] & (1 << bitc::PDI_DISJOINT))
+            cast<PossiblyDisjointInst>(I)->setIsDisjoint(true);
         } else if (isa<FPMathOperator>(I)) {
           FastMathFlags FMF = getDecodedFastMathFlags(Record[OpNum]);
           if (FMF.any())
             I->setFastMathFlags(FMF);
         }
-
       }
       break;
     }

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp

@@ -1540,6 +1540,9 @@ static uint64_t getOptimizationFlags(const Value *V) {
   } else if (const auto *PEO = dyn_cast<PossiblyExactOperator>(V)) {
     if (PEO->isExact())
       Flags |= 1 << bitc::PEO_EXACT;
+  } else if (const auto *PDI = dyn_cast<PossiblyDisjointInst>(V)) {
+    if (PDI->isDisjoint())
+      Flags |= 1 << bitc::PDI_DISJOINT;
   } else if (const auto *FPMO = dyn_cast<FPMathOperator>(V)) {
     if (FPMO->hasAllowReassoc())
       Flags |= bitc::AllowReassoc;

llvm/lib/IR/AsmWriter.cpp

@@ -1355,6 +1355,10 @@ static void WriteOptimizationInfo(raw_ostream &Out, const User *U) {
                dyn_cast<PossiblyExactOperator>(U)) {
     if (Div->isExact())
       Out << " exact";
+  } else if (const PossiblyDisjointInst *PDI =
+                 dyn_cast<PossiblyDisjointInst>(U)) {
+    if (PDI->isDisjoint())
+      Out << " disjoint";
   } else if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
     if (GEP->isInBounds())
       Out << " inbounds";

llvm/lib/IR/Instruction.cpp

@@ -357,6 +357,10 @@ void Instruction::dropPoisonGeneratingFlags() {
     cast<PossiblyExactOperator>(this)->setIsExact(false);
     break;
 
+  case Instruction::Or:
+    cast<PossiblyDisjointInst>(this)->setIsDisjoint(false);
+    break;
+
   case Instruction::GetElementPtr:
     cast<GetElementPtrInst>(this)->setIsInBounds(false);
     break;
@@ -532,6 +536,10 @@ void Instruction::copyIRFlags(const Value *V, bool IncludeWrapFlags) {
     if (isa<PossiblyExactOperator>(this))
       setIsExact(PE->isExact());
 
+  if (auto *SrcPD = dyn_cast<PossiblyDisjointInst>(V))
+    if (auto *DestPD = dyn_cast<PossiblyDisjointInst>(this))
+      DestPD->setIsDisjoint(SrcPD->isDisjoint());
+
   // Copy the fast-math flags.
   if (auto *FP = dyn_cast<FPMathOperator>(V))
     if (isa<FPMathOperator>(this))
@@ -558,6 +566,10 @@ void Instruction::andIRFlags(const Value *V) {
     if (isa<PossiblyExactOperator>(this))
       setIsExact(isExact() && PE->isExact());
 
+  if (auto *SrcPD = dyn_cast<PossiblyDisjointInst>(V))
+    if (auto *DestPD = dyn_cast<PossiblyDisjointInst>(this))
+      DestPD->setIsDisjoint(DestPD->isDisjoint() && SrcPD->isDisjoint());
+
   if (auto *FP = dyn_cast<FPMathOperator>(V)) {
     if (isa<FPMathOperator>(this)) {
       FastMathFlags FM = getFastMathFlags();

llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

@@ -231,8 +231,11 @@ Value *InstCombinerImpl::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
     // If either the LHS or the RHS are One, the result is One.
     if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
         SimplifyDemandedBits(I, 0, DemandedMask & ~RHSKnown.One, LHSKnown,
-                             Depth + 1))
+                             Depth + 1)) {
+      // Disjoint flag may not longer hold.
+      I->dropPoisonGeneratingFlags();
       return I;
+    }
     assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?");
     assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?");
 

llvm/test/Assembler/flags.ll

@@ -256,3 +256,8 @@ define i64 @test_zext(i32 %a) {
   ret i64 %res
 }
 
+define i64 @test_or(i64 %a, i64 %b) {
+; CHECK: %res = or disjoint i64 %a, %b
+  %res = or disjoint i64 %a, %b
+  ret i64 %res
+}

llvm/test/Bitcode/compatibility.ll

@@ -1359,6 +1359,10 @@ define void @instructions.bitwise_binops(i8 %op1, i8 %op2) {
   xor i8 %op1, %op2
   ; CHECK: xor i8 %op1, %op2
 
+  ; disjoint
+  or disjoint i8 %op1, %op2
+  ; CHECK: or disjoint i8 %op1, %op2
+
   ret void
 }
 

llvm/test/Bitcode/flags.ll

@@ -18,6 +18,8 @@ second:                                           ; preds = %first
   %z = add i32 %a, 0                              ; <i32> [#uses=0]
   %hh = zext nneg i32 %a to i64
   %ll = zext i32 %s to i64
+  %jj = or disjoint i32 %a, 0
+  %oo = or i32 %a, 0
   unreachable
 
 first:                                            ; preds = %entry
@@ -28,5 +30,7 @@ first:                                            ; preds = %entry
   %zz = add i32 %a, 0                             ; <i32> [#uses=0]
   %kk = zext nneg i32 %a to i64
   %rr = zext i32 %ss to i64
+  %mm = or disjoint i32 %a, 0
+  %nn = or i32 %a, 0
   br label %second
 }

llvm/test/Transforms/InstCombine/freeze.ll

@@ -1127,6 +1127,17 @@ define i32 @freeze_zext_nneg(i8 %x) {
   ret i32 %fr
 }
 
+define i32 @propagate_drop_flags_or(i32 %arg) {
+; CHECK-LABEL: @propagate_drop_flags_or(
+; CHECK-NEXT:    [[ARG_FR:%.*]] = freeze i32 [[ARG:%.*]]
+; CHECK-NEXT:    [[V1:%.*]] = or i32 [[ARG_FR]], 2
+; CHECK-NEXT:    ret i32 [[V1]]
+;
+  %v1 = or disjoint i32 %arg, 2
+  %v1.fr = freeze i32 %v1
+  ret i32 %v1.fr
+}
+
 !0 = !{}
 !1 = !{i64 4}
 !2 = !{i32 0, i32 100}

llvm/test/Transforms/InstCombine/or.ll

@@ -1576,3 +1576,14 @@ define <4 x i1> @and_or_not_or_logical_vec(<4 x i32> %ap, <4 x i32> %bp) {
   %Z = or <4 x i1> %X, %Y
   ret <4 x i1> %Z
 }
+
+; Make sure SimplifyDemandedBits drops the disjoint flag.
+define i8 @drop_disjoint(i8 %x) {
+; CHECK-LABEL: @drop_disjoint(
+; CHECK-NEXT:    [[B:%.*]] = or i8 [[X:%.*]], 1
+; CHECK-NEXT:    ret i8 [[B]]
+;
+  %a = and i8 %x, -2
+  %b = or disjoint i8 %a, 1
+  ret i8 %b
+}

llvm/test/Transforms/SimplifyCFG/HoistCode.ll

@@ -124,3 +124,33 @@ F:
   %z2 = zext i8 %x to i32
   ret i32 %z2
 }
+
+define i32 @hoist_or_flags_preserve(i1 %C, i32 %x, i32 %y) {
+; CHECK-LABEL: @hoist_or_flags_preserve(
+; CHECK-NEXT:  common.ret:
+; CHECK-NEXT:    [[Z1:%.*]] = or disjoint i32 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i32 [[Z1]]
+;
+  br i1 %C, label %T, label %F
+T:
+  %z1 = or disjoint i32 %x, %y
+  ret i32 %z1
+F:
+  %z2 = or disjoint i32 %x, %y
+  ret i32 %z2
+}
+
+define i32 @hoist_or_flags_drop(i1 %C, i32 %x, i32 %y) {
+; CHECK-LABEL: @hoist_or_flags_drop(
+; CHECK-NEXT:  common.ret:
+; CHECK-NEXT:    [[Z1:%.*]] = or i32 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i32 [[Z1]]
+;
+  br i1 %C, label %T, label %F
+T:
+  %z1 = or i32 %x, %y
+  ret i32 %z1
+F:
+  %z2 = or disjoint i32 %x, %y
+  ret i32 %z2
+}