[VPlan] Replace disjoint or with add instead of dropping disjoint. (#83821)

Dropping disjoint from an OR may yield incorrect results, as some
analysis may have converted it to an Add implicitly (e.g. SCEV used for
dependence analysis). Instead, replace it with an equivalent Add.

This is possible as all users of the disjoint OR only access lanes where
the operands are disjoint or poison otherwise.

Note that replacing all disjoint ORs with ADDs instead of dropping the
flags is not strictly necessary. It is only needed for disjoint ORs that
SCEV treated as ADDs, but those are not tracked.

There are other places that may drop poison-generating flags; those
likely need similar treatment.

Fixes #81872


PR: #83821

Author: fhahn

llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h

@@ -68,6 +68,9 @@ class VPBuilder {
 public:
   VPBuilder() = default;
   VPBuilder(VPBasicBlock *InsertBB) { setInsertPoint(InsertBB); }
+  VPBuilder(VPRecipeBase *InsertPt) {
+    setInsertPoint(InsertPt->getParent(), InsertPt->getIterator());
+  }
 
   /// Clear the insertion point: created instructions will not be inserted into
   /// a block.

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -1127,6 +1127,12 @@ class VPRecipeWithIRFlags : public VPSingleDefRecipe {
     return WrapFlags.HasNSW;
   }
 
+  bool isDisjoint() const {
+    assert(OpType == OperationType::DisjointOp &&
+           "recipe cannot have a disjoing flag");
+    return DisjointFlags.IsDisjoint;
+  }
+
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
   void printFlags(raw_ostream &O) const;
 #endif

llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h

@@ -261,6 +261,11 @@ m_Mul(const Op0_t &Op0, const Op1_t &Op1) {
   return m_Binary<Instruction::Mul, Op0_t, Op1_t>(Op0, Op1);
 }
 
+template <typename Op0_t, typename Op1_t>
+inline AllBinaryRecipe_match<Op0_t, Op1_t, Instruction::Or>
+m_Or(const Op0_t &Op0, const Op1_t &Op1) {
+  return m_Binary<Instruction::Or, Op0_t, Op1_t>(Op0, Op1);
+}
 } // namespace VPlanPatternMatch
 } // namespace llvm
 

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -1216,6 +1216,23 @@ void VPlanTransforms::dropPoisonGeneratingRecipes(
       // load/store. If the underlying instruction has poison-generating flags,
       // drop them directly.
       if (auto *RecWithFlags = dyn_cast<VPRecipeWithIRFlags>(CurRec)) {
+        VPValue *A, *B;
+        using namespace llvm::VPlanPatternMatch;
+        // Dropping disjoint from an OR may yield incorrect results, as some
+        // analysis may have converted it to an Add implicitly (e.g. SCEV used
+        // for dependence analysis). Instead, replace it with an equivalent Add.
+        // This is possible as all users of the disjoint OR only access lanes
+        // where the operands are disjoint or poison otherwise.
+        if (match(RecWithFlags, m_Or(m_VPValue(A), m_VPValue(B))) &&
+            RecWithFlags->isDisjoint()) {
+          VPBuilder Builder(RecWithFlags);
+          VPInstruction *New = Builder.createOverflowingOp(
+              Instruction::Add, {A, B}, {false, false},
+              RecWithFlags->getDebugLoc());
+          RecWithFlags->replaceAllUsesWith(New);
+          RecWithFlags->eraseFromParent();
+          CurRec = New;
+        }
         RecWithFlags->dropPoisonGeneratingFlags();
       } else {
         Instruction *Instr = dyn_cast_or_null<Instruction>(

llvm/test/Transforms/LoopVectorize/X86/pr81872.ll

@@ -29,7 +29,7 @@ define void @test(ptr noundef align 8 dereferenceable_or_null(16) %arr) #0 {
 ; CHECK-NEXT:    [[TMP2:%.*]] = and <4 x i64> [[VEC_IND]], <i64 1, i64 1, i64 1, i64 1>
 ; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq <4 x i64> [[TMP2]], zeroinitializer
 ; CHECK-NEXT:    [[TMP4:%.*]] = select <4 x i1> [[TMP1]], <4 x i1> [[TMP3]], <4 x i1> zeroinitializer
-; CHECK-NEXT:    [[TMP5:%.*]] = or i64 [[TMP0]], 1
+; CHECK-NEXT:    [[TMP5:%.*]] = add i64 [[TMP0]], 1
 ; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr i64, ptr [[ARR]], i64 [[TMP5]]
 ; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr i64, ptr [[TMP6]], i32 0
 ; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr i64, ptr [[TMP7]], i32 -3