[llvm] [VPlan] Delay adding canonical IV increment. (PR #82270)

[via llvm-commits]

================
@@ -7647,8 +7647,19 @@ DenseMap<const SCEV *, Value *> LoopVectorizationPlanner::executePlan(
        (!VectorizingEpilogue && !ExpandedSCEVs)) &&
       "expanded SCEVs to reuse can only be used during epilogue vectorization");
 
-  // TODO: Move to VPlan transform stage once the transition to the VPlan-based
-  // cost model is complete for better cost estimates.
+  bool IVUpdateMayOverflow =
+      !isIndvarOverflowCheckKnownFalse(&CM, BestVF, BestUF);
+  TailFoldingStyle Style = CM.getTailFoldingStyle(IVUpdateMayOverflow);
+  bool WithoutRuntimeCheck =
+      Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck;
+  // Use NUW for the induction increment if we proved that it won't overflow in
+  // the vector loop or when not folding the tail. In the latter case, we know
+  // that the canonical induction increment will not overflow as the vector trip
+  // count is >= increment and a multiple of the increment.
+  bool HasNUW = !IVUpdateMayOverflow || Style == TailFoldingStyle::None;
+  // TODO: Move transforms to VPlan transform stage once the transition to the
+  // VPlan-based cost model is complete for better cost estimates.
+  VPlanTransforms::convertCanonicalIV(BestVPlan, HasNUW, WithoutRuntimeCheck);
----------------
ayalz wrote:

Setting HasNUW here during VPlan execute is more accurate because of the BestUF --> IVUpdateMayOverflow --> tail-fold-style --> HasNUW dependence (i.e., along with "setUF() or "unrollByUF"), but is there a better way to set this earlier during VPlan planning, possibly splitting VF ranges - and UF ranges?

https://github.com/llvm/llvm-project/pull/82270