[llvm] [LV]: Teach LV to recursively (de)interleave. (PR #89018)

[Hassnaa Hamdi via llvm-commits]

================
@@ -2779,10 +2779,21 @@ static Value *interleaveVectors(IRBuilderBase &Builder, ArrayRef<Value *> Vals,
   // Scalable vectors cannot use arbitrary shufflevectors (only splats), so
   // must use intrinsics to interleave.
   if (VecTy->isScalableTy()) {
-    VectorType *WideVecTy = VectorType::getDoubleElementsVectorType(VecTy);
-    return Builder.CreateIntrinsic(WideVecTy, Intrinsic::vector_interleave2,
-                                   Vals,
-                                   /*FMFSource=*/nullptr, Name);
+    unsigned InterleaveFactor = Vals.size();
+    SmallVector<Value *> InterleavingValues(Vals);
+    // When interleaving, the number of values will be shrunk until we have the
+    // single final interleaved value.
+    VectorType *InterleaveTy =
+        cast<VectorType>(InterleavingValues[0]->getType());
+    for (unsigned Midpoint = Factor / 2; Midpoint > 0; Midpoint /= 2) {
----------------
hassnaaHamdi wrote:

Hi @Mel-Chen 
Thanks for looking at the patch.
The assert statement is already added before calling the interleaveVectors(..) function.

About your questions above:
1. yes I have landed a patch for adding support to the InterleaveAccessPass to support reading the (de)interleave tree pattern.
2. Going to add them.
3. If we have (de)interleave3 intrinsics, then we will have to do same logic for recursive (de)interleave3, and then the extra needed work will be representing the interleave factor by multiple of 2 and 3. so for the case of factor 6, we will do single iteration of (de)interleave2 then single iteration of (de)interleave3. The same logic will be applied for all factors that consist of multiples of 2 and 3 only.

https://github.com/llvm/llvm-project/pull/89018