[llvm] [InstCombine] Fold vector.reduce.op(vector.reverse(X)) -> vector.reduce.op(X) (PR #91743)

[David Sherwood via llvm-commits]

================
@@ -3394,8 +3436,31 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
                                 : 0;
     Value *Arg = II->getArgOperand(ArgIdx);
     Value *V;
+
+    if (!CanBeReassociated)
+      break;
+
+    if (match(Arg, m_VecReverse(m_Value(V)))) {
+      Value *Res;
+      switch (IID) {
+      case Intrinsic::vector_reduce_fadd:
+        Res = Builder.CreateFAddReduce(II->getArgOperand(0), V);
+        break;
+      case Intrinsic::vector_reduce_fmul:
+        Res = Builder.CreateFMulReduce(II->getArgOperand(0), V);
+        break;
+      case Intrinsic::vector_reduce_fmin:
+        Res = Builder.CreateFPMinReduce(V);
+        break;
+      case Intrinsic::vector_reduce_fmax:
+        Res = Builder.CreateFPMaxReduce(V);
+        break;
+      }
+      return replaceInstUsesWith(CI, Res);
+    }
+
----------------
david-arm wrote:

There is a problem with this approach. In the Intrinsic::vector_reduce_or case if I use the same approach as we do for fadd, etc. then what happens is on the first iteration we hit this code:

```
      replaceUse(II->getOperandUse(ArgIdx), V);
      return nullptr;
```

and then on the second iteration we still hit the existing code in Intrinsic::vector_reduce_or:

```
     if (!IsReverse && match(Arg, m_ZExtOrSExtOrSelf(m_Value(Vect)))) {
       ...
          return replaceInstUsesWith(CI, Res);
```

because on the 2nd iteration the reduce_or input is no longer a reverse intrinsic. InstCombine quite rightly gets unhappy with this error:

LLVM ERROR: Instruction Combining did not reach a fixpoint after 1 iterations

Essentially it looks like the only way I can stop the iterations is to return `replaceInstUsesWith` when we first spot the pattern (reduce.or(vector.reverse())) unless you have other suggestions that may help?

https://github.com/llvm/llvm-project/pull/91743