[llvm] [InstCombine] Fold patterns which uses <2N x iM> type for comparisons on <N x i2M> vector types (PR #184328)

[Simon Pilgrim via llvm-commits]

================
@@ -1292,6 +1292,219 @@ Value *InstCombinerImpl::foldUsingDistributiveLaws(BinaryOperator &I) {
   return SimplifySelectsFeedingBinaryOp(I, LHS, RHS);
 }
 
+// Folds patterns which uses comparisons on <2N x iM> type for a <N x i2M>
+// equality comparison.
+//
+// (A1, ..., AN) -> (A1Lower, A1Upper, ..., ANLower, ANUpper)
+// (B1, ..., BN) -> (B1Lower, B1Upper, ..., BNLower, BNUpper)
+// (Result1, ..., ResultN) -> (Result1, Result1, ..., ResultN, ResultN)
+//
+// where,
+//
+// ResultX = EqLowerX & EqUpperX
+// EqLowerX = AXLower == BXLower
+// EqUpperX = AXUpper == BXUpper
+//
+// Bitwise AND between the upper and lower parts can be achived by performing
+// the operation between the original and shuffled equality vector.
+static Value *foldVecCmpEqOnHalfElementSize(Instruction &I,
+                                            FixedVectorType *ResultVecType,
+                                            InstCombiner::BuilderTy &Builder) {
+  // Check pattern existance
+  Value *L, *R;
+  CmpPredicate Pred;
+  ArrayRef<int> Mask;
+
+  auto Equal = m_SExtOrSelf(m_ICmp(Pred, m_Value(L), m_Value(R)));
----------------
RKSimon wrote:

Merge predicate matching with m_c_SpecificICmp?
```suggestion
  auto Equal = m_SExtOrSelf(m_c_SpecificICmp(CmpInst::ICMP_EQ, m_Value(L), m_Value(R)));
```

https://github.com/llvm/llvm-project/pull/184328