[llvm] [SLP]Prefer segmented/deinterleaved loads to strided and fix codegen (PR #135058)

[Gaëtan Bossu via llvm-commits]

================
@@ -17618,8 +17617,27 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
             MaskValues = replicateMask(MaskValues, VecTy->getNumElements());
           }
           Constant *MaskValue = ConstantVector::get(MaskValues);
-          NewLI = Builder.CreateMaskedLoad(LoadVecTy, PO, CommonAlignment,
-                                           MaskValue);
+          if (InterleaveFactor &&
+              TTI->hasActiveVectorLength(Instruction::Load, LoadVecTy,
+                                         CommonAlignment)) {
+            // FIXME: codegen currently recognizes only vp.load, not
+            // masked.load, as segmented (deinterleaved) loads.
----------------
gbossu wrote:

Is that a limitation for all codegen, or only some targets?

https://github.com/llvm/llvm-project/pull/135058