[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:

When I replace the `add` in this test, by a `sub`, the pass still crashes, so this is not sufficient.
Does it matter what the operation (the one outside the loop) actually is?

I would have expected something like this:
```
define <vscale x 4 x i32> @cdotp_i8_rot0(<vscale x 32 x i8> %a, <vscale x 32 x i8> %b) {
entry:
  br label %vector.body

vector.body:                                      ; preds = %vector.body, %entry                                                                           
  %vec.phi = phi <vscale x 4 x i32> [ zeroinitializer, %entry ], [ %partial.reduce.sub, %vector.body ]                                                     
  %a.deinterleaved = call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.vector.deinterleave2.v32i8(<vscale x 32 x i8> %a)                               
  %b.deinterleaved = call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.vector.deinterleave2.v32i8(<vscale x 32 x i8> %b)                               
  %a.real = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8> } %a.deinterleaved, 0                                                                    
  %a.imag = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8> } %a.deinterleaved, 1                                                                    
  %b.real = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8> } %b.deinterleaved, 0
  %b.imag = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8> } %b.deinterleaved, 1                                                                    
  %a.real.ext = sext <vscale x 16 x i8> %a.real to <vscale x 16 x i32>                                                                                     
  %a.imag.ext = sext <vscale x 16 x i8> %a.imag to <vscale x 16 x i32>                                                                                     
  %b.real.ext = sext <vscale x 16 x i8> %b.real to <vscale x 16 x i32>
  %b.imag.ext = sext <vscale x 16 x i8> %b.imag to <vscale x 16 x i32>
  %real.mul = mul <vscale x 16 x i32> %b.real.ext, %a.real.ext
  %real.mul.reduced = call <vscale x 4 x i32> @llvm.experimental.vector.partial.reduce.add.nxv4i32.nxv16i32(<vscale x 4 x i32> %vec.phi, <vscale x 16 x i32> %real.mul)
  %imag.mul = mul <vscale x 16 x i32> %b.imag.ext, %a.imag.ext
  %imag.mul.neg = sub <vscale x 16 x i32> zeroinitializer, %imag.mul
  %partial.reduce.sub = call <vscale x 4 x i32> @llvm.experimental.vector.partial.reduce.add.nxv4i32.nxv16i32(<vscale x 4 x i32> %real.mul.reduced, <vscale x 16 x i32> %imag.mul.neg)
  br i1 true, label %middle.block, label %vector.body

middle.block:                                     ; preds = %vector.body
  ret <vscale x 4 x i32> %partial.reduce.sub
}
```
to use `cdot` instructions as well, but this case also seems to crash. This suggests that the issue is not to do with unrolling, but rather with the user outside the loop being anything else than a reduction?

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