[llvm] [IA]: Construct (de)interleave4 out of (de)interleave2 (PR #89276)

[Paul Walker via llvm-commits]

================
@@ -16637,40 +16691,79 @@ bool AArch64TargetLowering::lowerDeinterleaveIntrinsicToLoad(
         LdN = Builder.CreateCall(LdNFunc, {Pred, Address}, "ldN");
       else
         LdN = Builder.CreateCall(LdNFunc, Address, "ldN");
-
       Value *Idx =
           Builder.getInt64(I * LdTy->getElementCount().getKnownMinValue());
-      Left = Builder.CreateInsertVector(
-          VTy, Left, Builder.CreateExtractValue(LdN, 0), Idx);
-      Right = Builder.CreateInsertVector(
-          VTy, Right, Builder.CreateExtractValue(LdN, 1), Idx);
+      for (int J = 0; J < Factor; ++J) {
+        WideValues[J] = Builder.CreateInsertVector(
+            VTy, WideValues[J], Builder.CreateExtractValue(LdN, J), Idx);
+      }
+    }
+    if (Factor == 2)
+      Result = PoisonValue::get(StructType::get(VTy, VTy));
+    else
+      Result = PoisonValue::get(StructType::get(VTy, VTy, VTy, VTy));
+    // Construct the wide result out of the small results.
+    for (int J = 0; J < Factor; ++J) {
+      Result = Builder.CreateInsertValue(Result, WideValues[J], J);
     }
-
-    Result = PoisonValue::get(DI->getType());
-    Result = Builder.CreateInsertValue(Result, Left, 0);
-    Result = Builder.CreateInsertValue(Result, Right, 1);
   } else {
     if (UseScalable)
       Result = Builder.CreateCall(LdNFunc, {Pred, BaseAddr}, "ldN");
     else
       Result = Builder.CreateCall(LdNFunc, BaseAddr, "ldN");
   }
+  if (Factor > 2) {
+    for (unsigned I = 0; I < DeinterleavedValues.size(); I++) {
+      llvm::Value *CurrentExtract = DeinterleavedValues[I];
+      Value *NewExtract = Builder.CreateExtractValue(Result, I);
+      CurrentExtract->replaceAllUsesWith(NewExtract);
+      cast<Instruction>(CurrentExtract)->eraseFromParent();
----------------
paulwalker-arm wrote:

It 's possible for `DeinterleavedValues` to hold `Instruction*` rather than `Value*` then you'll not need such casts?

With that said, given the IR chain for for deinterleaving is non-trivial perhaps using `RecursivelyDeleteTriviallyDeadInstructions` at the end of the function would simplify things because you'll not need to track DeadInsts other than the top-level extracts you're replacing here.

https://github.com/llvm/llvm-project/pull/89276