[llvm] [VPlan] Do not reuse scalar NSW on reconstructed widened induction mul/add (PR #229350)

[Ramkumar Ramachandra via llvm-commits]

================
@@ -267,9 +267,19 @@ expandVPWidenIntOrFpInduction(VPWidenIntOrFpInductionRecipe *WidenIVR) {
   Instruction::BinaryOps AddOp;
   Instruction::BinaryOps MulOp;
   VPIRFlags Flags = *WidenIVR;
+  VPIRFlags AddFlags = Flags;
+  VPIRFlags MulFlags = Flags;
   if (ID.getKind() == InductionDescriptor::IK_IntInduction) {
     AddOp = Instruction::Add;
     MulOp = Instruction::Mul;
+    // Reconstructed initial addition can preserve NUW from the scalar
+    // recurrence since unsigned addition is monotonic, but cannot preserve
+    // NSW as intermediate vector lane offsets (e.g., lane * step) may wrap
+    // signed bounds even if scalar additions do not. Reconstructed
+    // multiplications cannot preserve NSW or NUW from scalar additions.
+    AddFlags = VPIRFlags(VPIRFlags::WrapFlagsTy(
+        Flags.getNoWrapFlagsOrNone().HasNUW, /*HasNSW=*/false));
----------------
artagnon wrote:

```suggestion
    AddFlags = VPIRFlags::WrapFlagsTy(
        Flags.getNoWrapFlagsOrNone().HasNUW, /*HasNSW=*/false);
```

I think it will automatically construct a VPIRFlags?

https://github.com/llvm/llvm-project/pull/229350