[llvm] [VPlan] Don't rely on region check in isUniformAfterVectorization. (PR #137883)

[via llvm-commits]

================
@@ -39,19 +39,43 @@ const SCEV *getSCEVExprForVPValue(VPValue *V, ScalarEvolution &SE);
 
 /// Returns true if \p VPV is uniform after vectorization.
 inline bool isUniformAfterVectorization(const VPValue *VPV) {
-  // A value defined outside the vector region must be uniform after
-  // vectorization inside a vector region.
-  if (VPV->isDefinedOutsideLoopRegions())
+  auto PreservesUniformity = [](unsigned Opcode) -> bool {
+    if (Instruction::isBinaryOp(Opcode) || Instruction::isCast(Opcode))
+      return true;
+    switch (Opcode) {
+    case Instruction::GetElementPtr:
+    case Instruction::ICmp:
+    case Instruction::FCmp:
+    case VPInstruction::Broadcast:
+    case VPInstruction::PtrAdd:
+      return true;
+    default:
+      return false;
+    }
+  };
+
+  // A live-in must be uniform across the scope of VPlan.
+  if (VPV->isLiveIn())
     return true;
-  if (auto *Rep = dyn_cast<VPReplicateRecipe>(VPV))
-    return Rep->isUniform();
+
+  if (auto *Rep = dyn_cast<VPReplicateRecipe>(VPV)) {
+    const VPRegionBlock *RegionOfR = Rep->getParent()->getParent();
+    if (RegionOfR && RegionOfR->isReplicator())
+      return false;
----------------
ayalz wrote:

Maybe worth a comment - it's ok functionally to sink uniform/after-vectorization recipes into replicate regions, we avoid doing so due to cost considerations. Here an early exit condition is introduced before checking operands recursively - another cost rather than functional decision - could do w/o the replicate region check?

Note that prior to introducing replicate regions, the recipes they are to include can (and should?) be identified by isPredicated().

https://github.com/llvm/llvm-project/pull/137883