[SLP]Do not create copyable node, if parent node is non-schedulable and has a use in binop.

If the parent node is non-schedulable (only externally used instructions), and at least one instruction has multiple uses and used in the binop, such copyable node should be created. Otherwise, it may contain wrong def-use chain model, which cannot be effective detected.

Fixes #166035

Author: alexey-bataev

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -20975,6 +20975,27 @@ BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP,
   if (isa<PHINode>(S.getMainOp()) ||
       isVectorLikeInstWithConstOps(S.getMainOp()))
     return nullptr;
+  // If the parent node is non-schedulable and the current node is copyable, and
+  // any of parent instructions are used outside several basic blocks or in
+  // bin-op node - cancel scheduling, it may cause wrong def-use deps in
+  // analysis, leading to a crash.
+  // Non-scheduled nodes may not have related ScheduleData model, which may lead
+  // to a skipped dep analysis.
+  if (S.areInstructionsWithCopyableElements() && EI && EI.UserTE->hasState() &&
+      EI.UserTE->doesNotNeedToSchedule() &&
+      EI.UserTE->getOpcode() != Instruction::PHI &&
+      any_of(EI.UserTE->Scalars, [](Value *V) {
+        auto *I = dyn_cast<Instruction>(V);
+        if (!I || I->hasOneUser())
+          return false;
+        for (User *U : I->users()) {
+          auto *UI = cast<Instruction>(U);
+          if (isa<BinaryOperator>(UI))
+            return true;
+        }
+        return false;
+      }))
+    return std::nullopt;
   bool HasCopyables = S.areInstructionsWithCopyableElements();
   if (((!HasCopyables && doesNotNeedToSchedule(VL)) ||
        all_of(VL, [&](Value *V) { return S.isNonSchedulable(V); }))) {

llvm/test/Transforms/SLPVectorizer/X86/parent-non-schedule-multi-use-in-binop.ll

@@ -0,0 +1,40 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -passes=slp-vectorizer -S --mtriple=x86_64-unknown-linux-gnu < %s | FileCheck %s
+
+@a = common global [100 x i64] zeroinitializer, align 64
+
+define void @test() {
+; CHECK-LABEL: define void @test() {
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    [[TMP0:%.*]] = load <2 x i64>, ptr getelementptr inbounds nuw (i8, ptr @a, i64 48), align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = add <2 x i64> [[TMP0]], splat (i64 1)
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <2 x i64> [[TMP0]], <2 x i64> [[TMP1]], <2 x i32> <i32 0, i32 3>
+; CHECK-NEXT:    [[TMP3:%.*]] = add <2 x i64> [[TMP2]], splat (i64 1)
+; CHECK-NEXT:    br i1 false, label %[[LOP_RHSCNT_I_PEEL:.*]], label %[[LAND_END_I_PEEL:.*]]
+; CHECK:       [[LOP_RHSCNT_I_PEEL]]:
+; CHECK-NEXT:    [[TMP4:%.*]] = or <2 x i64> [[TMP1]], <i64 1, i64 0>
+; CHECK-NEXT:    br label %[[LAND_END_I_PEEL]]
+; CHECK:       [[LAND_END_I_PEEL]]:
+; CHECK-NEXT:    [[TMP5:%.*]] = phi <2 x i64> [ [[TMP3]], %[[ENTRY]] ], [ [[TMP4]], %[[LOP_RHSCNT_I_PEEL]] ]
+; CHECK-NEXT:    store <2 x i64> [[TMP5]], ptr getelementptr inbounds nuw (i8, ptr @a, i64 48), align 8
+; CHECK-NEXT:    ret void
+;
+entry:
+  %.promoted104.i = load i64, ptr getelementptr inbounds nuw (i8, ptr @a, i64 56), align 8
+  %.promoted103.i = load i64, ptr getelementptr inbounds nuw (i8, ptr @a, i64 48), align 8
+  %0 = add i64 %.promoted104.i, 1
+  %1 = add i64 %.promoted103.i, 1
+  %2 = add i64 %0, 1
+  br i1 false, label %lop.rhscnt.i.peel, label %land.end.i.peel
+
+lop.rhscnt.i.peel:
+  %3 = or i64 %1, 1
+  br label %land.end.i.peel
+
+land.end.i.peel:
+  %4 = phi i64 [ %2, %entry ], [ %0, %lop.rhscnt.i.peel ]
+  %5 = phi i64 [ %1, %entry ], [ %3, %lop.rhscnt.i.peel ]
+  store i64 %5, ptr getelementptr inbounds nuw (i8, ptr @a, i64 48), align 8
+  store i64 %4, ptr getelementptr inbounds nuw (i8, ptr @a, i64 56), align 8
+  ret void
+}