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

[Florian Hahn via llvm-commits]

================
@@ -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.
----------------
fhahn wrote:

> Should this be taken care of by VPRecipeWithIRFlags::dropPoisonGeneratingFlags(), i.e., have it replace the opcode of a disjoint Or with Add?

Hmmm, it may work if there's no need to directly delete the old recipe, i.e. it will get cleaned up. Let me check.

> Should code based on such SCEV expressions be generated (expanded(?)) directly?
> 
> (This case is reminiscent of dropping assumptions based on conditions subject to predication.)

At the moment, SCEV is only used for analysis of existing IR instructions in this case; switching use SCEV expansion for pointer expressions would probably introduce some other potential issues.

https://github.com/llvm/llvm-project/pull/83821