[llvm] [AArch64][InstCombine] Fold zext cmpne umin boolean trees (PR #209234)

[Kerry McLaughlin via llvm-commits]

================
@@ -2205,6 +2205,69 @@ static std::optional<Instruction *> instCombineZExtSVECmpNE(InstCombiner &IC,
   return std::nullopt;
 }
 
+static std::optional<Instruction *> isZextedCmpneUMin(Value *V) {
+  Instruction *CmpNE;
+  if (!match(V, m_ZExt(m_Instruction(CmpNE))))
+    return std::nullopt;
+
+  Value *Pg;
+  Instruction *UMin;
+  if (!match(CmpNE, m_Intrinsic<Intrinsic::aarch64_sve_cmpne>(
+                        m_Value(Pg), m_Zero(), m_Instruction(UMin))) ||
+      !isAllActivePredicate(Pg) || V->getType() != UMin->getType())
+    return std::nullopt;
+
+  Value *UMinPg;
+  if (!match(UMin, m_Intrinsic<Intrinsic::aarch64_sve_umin_u>(
+                       m_Value(UMinPg), m_Value(), m_Value())) ||
+      !isAllActivePredicate(UMinPg))
+    return std::nullopt;
+
+  return UMin;
+}
+
+// InstCombine:
+// zext(cmpne(0, umin(zext(cmpne(0, umin(%a, %b)))
+//                    zext(cmpne(0, umin(%c, %d))))))
+// ->
+// zext(cmpne(0, umin(umin(%a, %b),
+//                    umin(%c, %d))))
+//
+// To remove redundant ZExt and CmpNE intrinsics.
+static std::optional<Instruction *>
+instCombineZExtCmpNEUMinTree(InstCombiner &IC, IntrinsicInst &II) {
+  Value *Pg = II.getOperand(0);
+  if (!isAllActivePredicate(Pg) || !II.hasOneUse())
+    return std::nullopt;
+
+  auto ZExt = cast<Instruction>(*II.user_begin());
+  auto _UMin = isZextedCmpneUMin(ZExt);
+  if (!_UMin)
+    return std::nullopt;
+  Instruction *UMin = *_UMin;
+  // Check for outer UMin single use to avoid generating an extra UMin
+  if (!UMin->hasOneUse())
+    return std::nullopt;
----------------
kmclaughlin-arm wrote:

It looks like that might be because results from similar functions are then returned by `instCombineIntrinsic` (which also returns `std::optional<Instruction *>`), and that's not the case here?

https://github.com/llvm/llvm-project/pull/209234