[llvm] [llvm] Fix crash when complex deinterleaving operates on an unrolled loop (PR #129735)

[Sander de Smalen via llvm-commits]

================
@@ -2253,8 +2261,31 @@ void ComplexDeinterleavingGraph::processReductionSingle(
   auto *FinalReduction = ReductionInfo[Real].second;
   Builder.SetInsertPoint(&*FinalReduction->getParent()->getFirstInsertionPt());
 
-  auto *AddReduce = Builder.CreateAddReduce(OperationReplacement);
+  Value *Other;
+  bool EraseFinalReductionHere = false;
+  if (match(FinalReduction, m_c_Add(m_Specific(Real), m_Value(Other)))) {
----------------
sdesmalen-arm wrote:

Thanks @igogo-x86, I agree it makes more sense to feed the result from one `cdot` into the other, rather than changing the PHI node to be a wider type.

My understanding is that the Complex Deinterleaving pass was created because for certain operations (like reductions) where there is both a PHI+reduction for the imaginary and one PHI+reduction for the real part, it is better to keep the intermediate values interleaved, because the `cmla` instruction takes interleaved tuples as input and returns interleaved tuples as output. This avoids having to deinterleave values first and it also allows using specialised `cmla` instructions to do the complex MLA operation. The reduction PHI then contains a vector of `<(r, i), (r, i), ..>` tuples, which need de-interleaving only when doing the final reduction.
For the case of `cdot` instructions there is no need for this, because the result vector will always be deinterleaved (the `cdot` instruction returns either a widened real, or a widened imaginary result).

If that understanding is correct, then I don't really see a need to implement this optimization in the ComplexDeinterleave pass. This looks more like a DAGcombine of `partialreduce(mul(ext(deinterleave(a)), ext(deinterleave(b))) -> cdot(a, b, #0)` (with some variation of this pattern for other rotations). With the new ISD node `ISD::PARTIAL_REDUCE_[U|S]MLA` added by @JamesChesterman this should be even easier to identify.

Please let me know if I'm missing anything here though.

https://github.com/llvm/llvm-project/pull/129735