[llvm] [VPlan] Directly unroll VectorEndPointerRecipe (PR #172372)

[Florian Hahn via llvm-commits]

================
@@ -2992,10 +2992,19 @@ static VPRecipeBase *optimizeMaskToEVL(VPValue *HeaderMask,
 
   /// Adjust any end pointers so that they point to the end of EVL lanes not VF.
   auto AdjustEndPtr = [&CurRecipe, &EVL](VPValue *EndPtr) {
-    auto *EVLEndPtr = cast<VPVectorEndPointerRecipe>(EndPtr)->clone();
-    EVLEndPtr->insertBefore(&CurRecipe);
-    EVLEndPtr->setOperand(1, &EVL);
-    return EVLEndPtr;
+    auto *VEPR = cast<VPVectorEndPointerRecipe>(EndPtr);
+    VPBuilder Builder(&CurRecipe);
+    return Builder.createVectorEndPointerRecipe(
+        VEPR->getOperand(0), VEPR->getSourceElementType(), VEPR->getStride(),
+        VEPR->getGEPNoWrapFlags(), &EVL, VEPR->getDebugLoc());
+  };
+
+  auto m_VecEndPtrVF = [&Plan](VPValue *&Addr, int64_t Stride) { // NOLINT
+    return m_VecEndPtr(
+        m_VPValue(Addr),
+        m_c_Mul(
+            m_SpecificSInt(Stride),
+            m_Sub(m_ZExtOrTruncOrSelf(m_Specific(&Plan->getVF())), m_One())));
   };
----------------
fhahn wrote:

this complicaion comes from the fact that we now always generate the explicit offse.

could we instead keep the vf operand, and just have add the expanded stride after unrolling to keep things simpler early on, similar to https://github.com/llvm/llvm-project/pull/170906?

 we would make sure to materialize the offset even without unrolling, but that can be done in convertToConcreteRecipes for example



https://github.com/llvm/llvm-project/pull/172372