[llvm] [LV] Add support for partial alias masking with tail folding (PR #182457)

[Sander de Smalen via llvm-commits]

================
@@ -8682,6 +8747,49 @@ void LoopVectorizationPlanner::attachRuntimeChecks(
   }
 }
 
+void LoopVectorizationPlanner::materializeAliasMask(
+    VPlan &Plan, ArrayRef<PointerDiffInfo> DiffChecks, bool HasBranchWeights) {
+  assert(CM.foldTailByMasking() && "Expected tail folding to be enabled");
+  VPBasicBlock *ClampedVFCheck =
+      Plan.createVPBasicBlock("vector.clamped.vf.check");
+
+  VPValue *ClampedVF = VPlanTransforms::materializeAliasMask(
+      Plan, ClampedVFCheck,
+      *CM.Legal->getRuntimePointerChecking()->getDiffChecks());
+  VPBuilder Builder(ClampedVFCheck);
+  DebugLoc DL = DebugLoc::getCompilerGenerated();
+  Type *TCTy = VPTypeAnalysis(Plan).inferScalarType(Plan.getTripCount());
+
+  // Check the "ClampedVF" from the alias mask is not scalar.
+  VPValue *IsScalar =
+      Builder.createICmp(CmpInst::ICMP_ULE, ClampedVF,
+                         Plan.getConstantInt(TCTy, 1), DL, "vf.is.scalar");
+
+  VPValue *TripCount = Plan.getTripCount();
+  VPValue *MaxUIntTripCount =
+      Plan.getConstantInt(cast<IntegerType>(TCTy)->getMask());
+  VPValue *DistanceToMax = Builder.createSub(MaxUIntTripCount, TripCount);
+
+  // For tail-folding: Don't execute the vector loop if (UMax - n) < ClampedVF.
----------------
sdesmalen-arm wrote:

nit: maybe explain that this is required because the ClampVF is not guaranteed to be a power of 2.

https://github.com/llvm/llvm-project/pull/182457