feat[datafusion]: support regex and LIKE coercion on Dict<Binary> types

And others that need an extra coercion step before string coercion.

Signed-off-by: Alfonso Subiotto Marques <alfonso.subiotto@polarsignals.com>

Author: asubiotto

datafusion/expr-common/src/type_coercion/binary.rs

@@ -1766,7 +1766,7 @@ fn binary_coercion(lhs_type: &DataType, rhs_type: &DataType) -> Option<DataType>
 pub fn like_coercion(lhs_type: &DataType, rhs_type: &DataType) -> Option<DataType> {
     string_coercion(lhs_type, rhs_type)
         .or_else(|| binary_to_string_coercion(lhs_type, rhs_type))
-        .or_else(|| dictionary_coercion(lhs_type, rhs_type, false, string_coercion))
+        .or_else(|| dictionary_coercion(lhs_type, rhs_type, false, like_coercion))
         .or_else(|| ree_coercion(lhs_type, rhs_type, false, like_coercion))
         .or_else(|| regex_null_coercion(lhs_type, rhs_type))
         .or_else(|| null_coercion(lhs_type, rhs_type))
@@ -1787,7 +1787,7 @@ fn regex_null_coercion(lhs_type: &DataType, rhs_type: &DataType) -> Option<DataT
 /// This is a union of string coercion rules, dictionary coercion rules, and REE coercion rules.
 pub fn regex_coercion(lhs_type: &DataType, rhs_type: &DataType) -> Option<DataType> {
     string_coercion(lhs_type, rhs_type)
-        .or_else(|| dictionary_coercion(lhs_type, rhs_type, false, string_coercion))
+        .or_else(|| dictionary_coercion(lhs_type, rhs_type, false, regex_coercion))
         .or_else(|| ree_coercion(lhs_type, rhs_type, false, regex_coercion))
         .or_else(|| regex_null_coercion(lhs_type, rhs_type))
 }

datafusion/sqllogictest/test_files/regexp/regexp_like.slt

@@ -277,6 +277,23 @@ drop table strings
 statement ok
 drop table dict_table
 
+# Dict value types that themselves need further coercion against the literal
+statement ok
+create table dict_inner as
+select arrow_cast(arrow_cast(c, 'Binary'), 'Dictionary(UInt32, Binary)') as bin_col,
+       arrow_cast(arrow_cast(c, 'Dictionary(UInt32, Utf8)'),
+                  'Dictionary(Int32, Dictionary(UInt32, Utf8))') as nested_col
+from (values ('foo'), ('bar')) as t(c);
+
+query BB
+select bin_col LIKE 'foo', nested_col ~ 'foo' from dict_inner;
+----
+true true
+false false
+
+statement ok
+drop table dict_inner
+
 # Ensure that regexp_like is rewritten to use the (more optimized) regex operators
 statement ok
 create table regexp_test as values