New IR -- WIP

[kasiafi]

See core/trino-main/src/main/java/io/trino/sql/newir/README.md for details.

[kasiafi]

example assembly printout

IR version = 1
%0 = query() : () -> "boolean" ({
    ^query
        %1 = table_scan() : () -> "multiset(row(""f_1"" varchar(25),""f_2"" bigint))" ()
            {table_handle = "{""catalogHandle"":""tpch:normal:default"",""connectorHandle"":{""@type"":""../../plugin/trino-tpch/pom.xml:io.trino.plugin.tpch.TpchTableHandle"",""schemaName"":""tiny"",""tableName"":""nation"",""scaleFactor"":0.01,""constraint"":{""columnDomains"":[]}},""transaction"":[""../../plugin/trino-tpch/pom.xml:io.trino.plugin.tpch.TpchTransactionHandle"",""INSTANCE""]}", column_handles = "[{""@type"":""../../plugin/trino-tpch/pom.xml:io.trino.plugin.tpch.TpchColumnHandle"",""columnName"":""name"",""type"":""varchar(25)""},{""@type"":""../../plugin/trino-tpch/pom.xml:io.trino.plugin.tpch.TpchColumnHandle"",""columnName"":""regionkey"",""type"":""bigint""}]", constraint = "{""columnDomains"":[]}", update_target = "false", use_connector_node_partitioning = "false"}
        %2 = filter(%1) : ("multiset(row(""f_1"" varchar(25),""f_2"" bigint))") -> "multiset(row(""f_1"" varchar(25),""f_2"" bigint))" ({
            ^predicate (%3 : "row(""f_1"" varchar(25),""f_2"" bigint)")
                %4 = field_selection(%3) : ("row(""f_1"" varchar(25),""f_2"" bigint)") -> "bigint" ()
                    {field_name = "f_2"}
                %5 = constant() : () -> "bigint" ()
                    {constant_result = "{""type"":""bigint"",""value"":2}"}
                %6 = comparison(%4, %5) : ("bigint", "bigint") -> "boolean" ()
                    {comparison_operator = "GREATER_THAN"}
                %7 = return(%6) : ("boolean") -> "boolean" ()
                    {ir.terminal = "true"}
            })
        %8 = project(%2) : ("multiset(row(""f_1"" varchar(25),""f_2"" bigint))") -> "multiset(row(""f_1"" varchar(25)))" ({
            ^assignments (%9 : "row(""f_1"" varchar(25),""f_2"" bigint)")
                %10 = field_selection(%9) : ("row(""f_1"" varchar(25),""f_2"" bigint)") -> "varchar(25)" ()
                    {field_name = "f_1"}
                %11 = call(%10) : ("varchar(25)") -> "varchar(25)" ()
                    {resolved_function = "{""signature"":{""name"":{""catalogName"":""system"",""schemaName"":""builtin"",""functionName"":""lower""},""returnType"":""varchar(25)"",""argumentTypes"":[""varchar(25)""]},""catalogHandle"":""system:normal:system"",""functionId"":""lower(varchar(x)):varchar(x)"",""functionKind"":""SCALAR"",""deterministic"":true,""functionNullability"":{""returnNullable"":false,""argumentNullable"":[false]},""typeDependencies"":{},""functionDependencies"":[]}"}
                %12 = row(%11) : ("varchar(25)") -> "row(varchar(25))" ()
                %13 = return(%12) : ("row(varchar(25))") -> "row(varchar(25))" ()
                    {ir.terminal = "true"}
            })
        %14 = output(%8) : ("multiset(row(""f_1"" varchar(25)))") -> "boolean" ({
            ^outputFieldSelector (%15 : "row(""f_1"" varchar(25))")
                %16 = field_selection(%15) : ("row(""f_1"" varchar(25))") -> "varchar(25)" ()
                    {field_name = "f_1"}
                %17 = row(%16) : ("varchar(25)") -> "row(varchar(25))" ()
                %18 = return(%17) : ("row(varchar(25))") -> "row(varchar(25))" ()
                    {ir.terminal = "true"}
            })
            {output_names = "[""_col0""]", ir.terminal = "true"}
    })
    {ir.terminal = "true"}

[kasiafi]

I believe that the attributes should be on the top level of the dialect, as opposed to being part of an operation.
IR allows mixing dialects. In particular, an operation can have attributes from another dialect. For example, the trino.query operation has the attribute ir.terminal. The dialect is supposed to understand an attribute outside the context of an operation.

[kasiafi]

Currently, the intermediate relation row type has field names, and the fields are referenced by name with the FieldSelection operation. It will be refactored so that the row type is anonymous, and the fields will be referenced by index with the FieldReference operation.

Explanation:
The query program must work correctly with the Memo and Equivalence Classes. For that purpose, the intermediate relation type must be generic. Each operation in an Equivalence Class must derive exactly the same output type, because they all must be compatible with the downstream program.
Due to this limitation, field names in the intermediate row type aren't very useful. We must use generic sequential names, for example f_1, f_2, f_3... Using indexes would be more concise.

[kasiafi]

There is a lot of repetition of this code across different operation classes. It will be extracted and reused.