[mypyc] Precompute set literals for "in" ops against / iteration over set literals (#14409)

Towards mypyc/mypyc#726. (There's a Python
compatibility bug that needs to be fixed before the issue can be
closed.)

For example, the set literals here are now precomputed as frozensets at
module initialization.

```
    x in {1, 2.0, "3"}
    x not in {1, 2.0, "3"}

    for _ in {1, 2.0, "3"}:
        ...
```

Set literal items supported:

- Anything supported by `irbuild.constant_fold.constant_fold_expr()`
  - String and integer literals
  - Final references to int/str values
- Certain int and str unary/binary ops that evaluate to a constant value
- `None`, `True`, and `False`
- Float, byte, and complex literals
- Tuple literals with only items listed above

**Results** (using gcc-9 on 64-bit Ubuntu)

Master @ 98cc165

    running in_set
    ..........
    interpreted: 0.495790s (avg of 5 iterations; stdev 6.8%)
    compiled:    0.810029s (avg of 5 iterations; stdev 1.5%)

    compiled is 0.612x faster

    running set_literal_iteration

.........................................................................................
    interpreted: 0.020255s (avg of 45 iterations; stdev 2.5%)
    compiled:    0.016336s (avg of 45 iterations; stdev 1.8%)

    compiled is 1.240x faster

This PR

    running in_set
    ..........
    interpreted: 0.502020s (avg of 5 iterations; stdev 1.1%)
    compiled:    0.390281s (avg of 5 iterations; stdev 6.2%)

    compiled is 1.286x faster

    running set_literal_iteration

..............................................................................................
    interpreted: 0.019917s (avg of 47 iterations; stdev 2.2%)
    compiled:    0.007134s (avg of 47 iterations; stdev 2.6%)

    compiled is 2.792x faster

Benchmarks can be found here: mypyc/mypyc-benchmarks#32

Author: ichard26

mypyc/analysis/ircheck.py

@@ -252,6 +252,15 @@ def check_tuple_items_valid_literals(self, op: LoadLiteral, t: tuple[object, ...
             if isinstance(x, tuple):
                 self.check_tuple_items_valid_literals(op, x)
 
+    def check_frozenset_items_valid_literals(self, op: LoadLiteral, s: frozenset[object]) -> None:
+        for x in s:
+            if x is None or isinstance(x, (str, bytes, bool, int, float, complex)):
+                pass
+            elif isinstance(x, tuple):
+                self.check_tuple_items_valid_literals(op, x)
+            else:
+                self.fail(op, f"Invalid type for item of frozenset literal: {type(x)})")
+
     def visit_load_literal(self, op: LoadLiteral) -> None:
         expected_type = None
         if op.value is None:
@@ -271,6 +280,11 @@ def visit_load_literal(self, op: LoadLiteral) -> None:
         elif isinstance(op.value, tuple):
             expected_type = "builtins.tuple"
             self.check_tuple_items_valid_literals(op, op.value)
+        elif isinstance(op.value, frozenset):
+            # There's no frozenset_rprimitive type since it'd be pretty useless so we just pretend
+            # it's a set (when it's really a frozenset).
+            expected_type = "builtins.set"
+            self.check_frozenset_items_valid_literals(op, op.value)
 
         assert expected_type is not None, "Missed a case for LoadLiteral check"
 

mypyc/codegen/emitmodule.py

@@ -669,6 +669,9 @@ def generate_literal_tables(self) -> None:
         # Descriptions of tuple literals
         init_tuple = c_array_initializer(literals.encoded_tuple_values())
         self.declare_global("const int []", "CPyLit_Tuple", initializer=init_tuple)
+        # Descriptions of frozenset literals
+        init_frozenset = c_array_initializer(literals.encoded_frozenset_values())
+        self.declare_global("const int []", "CPyLit_FrozenSet", initializer=init_frozenset)
 
     def generate_export_table(self, decl_emitter: Emitter, code_emitter: Emitter) -> None:
         """Generate the declaration and definition of the group's export struct.
@@ -839,7 +842,7 @@ def generate_globals_init(self, emitter: Emitter) -> None:
         for symbol, fixup in self.simple_inits:
             emitter.emit_line(f"{symbol} = {fixup};")
 
-        values = "CPyLit_Str, CPyLit_Bytes, CPyLit_Int, CPyLit_Float, CPyLit_Complex, CPyLit_Tuple"
+        values = "CPyLit_Str, CPyLit_Bytes, CPyLit_Int, CPyLit_Float, CPyLit_Complex, CPyLit_Tuple, CPyLit_FrozenSet"
         emitter.emit_lines(
             f"if (CPyStatics_Initialize(CPyStatics, {values}) < 0) {{", "return -1;", "}"
         )

mypyc/codegen/literals.py

@@ -1,12 +1,13 @@
 from __future__ import annotations
 
-from typing import Any, Tuple, Union, cast
+from typing import Any, Dict, FrozenSet, List, Tuple, Union, cast
 from typing_extensions import Final
 
-# Supported Python literal types. All tuple items must have supported
+# Supported Python literal types. All tuple / frozenset items must have supported
 # literal types as well, but we can't represent the type precisely.
-LiteralValue = Union[str, bytes, int, bool, float, complex, Tuple[object, ...], None]
-
+LiteralValue = Union[
+    str, bytes, int, bool, float, complex, Tuple[object, ...], FrozenSet[object], None
+]
 
 # Some literals are singletons and handled specially (None, False and True)
 NUM_SINGLETONS: Final = 3
@@ -23,6 +24,7 @@ def __init__(self) -> None:
         self.float_literals: dict[float, int] = {}
         self.complex_literals: dict[complex, int] = {}
         self.tuple_literals: dict[tuple[object, ...], int] = {}
+        self.frozenset_literals: dict[frozenset[object], int] = {}
 
     def record_literal(self, value: LiteralValue) -> None:
         """Ensure that the literal value is available in generated code."""
@@ -55,6 +57,12 @@ def record_literal(self, value: LiteralValue) -> None:
                 for item in value:
                     self.record_literal(cast(Any, item))
                 tuple_literals[value] = len(tuple_literals)
+        elif isinstance(value, frozenset):
+            frozenset_literals = self.frozenset_literals
+            if value not in frozenset_literals:
+                for item in value:
+                    self.record_literal(cast(Any, item))
+                frozenset_literals[value] = len(frozenset_literals)
         else:
             assert False, "invalid literal: %r" % value
 
@@ -86,6 +94,9 @@ def literal_index(self, value: LiteralValue) -> int:
         n += len(self.complex_literals)
         if isinstance(value, tuple):
             return n + self.tuple_literals[value]
+        n += len(self.tuple_literals)
+        if isinstance(value, frozenset):
+            return n + self.frozenset_literals[value]
         assert False, "invalid literal: %r" % value
 
     def num_literals(self) -> int:
@@ -98,6 +109,7 @@ def num_literals(self) -> int:
             + len(self.float_literals)
             + len(self.complex_literals)
             + len(self.tuple_literals)
+            + len(self.frozenset_literals)
         )
 
     # The following methods return the C encodings of literal values
@@ -119,24 +131,32 @@ def encoded_complex_values(self) -> list[str]:
         return _encode_complex_values(self.complex_literals)
 
     def encoded_tuple_values(self) -> list[str]:
-        """Encode tuple values into a C array.
+        return self._encode_collection_values(self.tuple_literals)
+
+    def encoded_frozenset_values(self) -> List[str]:
+        return self._encode_collection_values(self.frozenset_literals)
+
+    def _encode_collection_values(
+        self, values: dict[tuple[object, ...], int] | dict[frozenset[object], int]
+    ) -> list[str]:
+        """Encode tuple/frozenset values into a C array.
 
         The format of the result is like this:
 
-           <number of tuples>
-           <length of the first tuple>
+           <number of collections>
+           <length of the first collection>
            <literal index of first item>
            ...
            <literal index of last item>
-           <length of the second tuple>
+           <length of the second collection>
            ...
         """
-        values = self.tuple_literals
-        value_by_index = {index: value for value, index in values.items()}
+        # FIXME: https://github.com/mypyc/mypyc/issues/965
+        value_by_index = {index: value for value, index in cast(Dict[Any, int], values).items()}
         result = []
-        num = len(values)
-        result.append(str(num))
-        for i in range(num):
+        count = len(values)
+        result.append(str(count))
+        for i in range(count):
             value = value_by_index[i]
             result.append(str(len(value)))
             for item in value:

mypyc/ir/ops.py

@@ -39,6 +39,7 @@
 )
 
 if TYPE_CHECKING:
+    from mypyc.codegen.literals import LiteralValue
     from mypyc.ir.class_ir import ClassIR
     from mypyc.ir.func_ir import FuncDecl, FuncIR
 
@@ -588,7 +589,7 @@ class LoadLiteral(RegisterOp):
     This is used to load a static PyObject * value corresponding to
     a literal of one of the supported types.
 
-    Tuple literals must contain only valid literal values as items.
+    Tuple / frozenset literals must contain only valid literal values as items.
 
     NOTE: You can use this to load boxed (Python) int objects. Use
           Integer to load unboxed, tagged integers or fixed-width,
@@ -603,11 +604,7 @@ class LoadLiteral(RegisterOp):
     error_kind = ERR_NEVER
     is_borrowed = True
 
-    def __init__(
-        self,
-        value: None | str | bytes | bool | int | float | complex | tuple[object, ...],
-        rtype: RType,
-    ) -> None:
+    def __init__(self, value: LiteralValue, rtype: RType) -> None:
         self.value = value
         self.type = rtype
 

mypyc/ir/pprint.py

@@ -106,7 +106,18 @@ def visit_load_literal(self, op: LoadLiteral) -> str:
         # it explicit that this is a Python object.
         if isinstance(op.value, int):
             prefix = "object "
-        return self.format("%r = %s%s", op, prefix, repr(op.value))
+
+        rvalue = repr(op.value)
+        if isinstance(op.value, frozenset):
+            # We need to generate a string representation that won't vary
+            # run-to-run because sets are unordered, otherwise we may get
+            # spurious irbuild test failures.
+            #
+            # Sorting by the item's string representation is a bit of a
+            # hack, but it's stable and won't cause TypeErrors.
+            formatted_items = [repr(i) for i in sorted(op.value, key=str)]
+            rvalue = "frozenset({" + ", ".join(formatted_items) + "})"
+        return self.format("%r = %s%s", op, prefix, rvalue)
 
     def visit_get_attr(self, op: GetAttr) -> str:
         return self.format("%r = %s%r.%s", op, self.borrow_prefix(op), op.obj, op.attr)

mypyc/irbuild/builder.py

@@ -118,7 +118,7 @@
     AssignmentTargetRegister,
     AssignmentTargetTuple,
 )
-from mypyc.irbuild.util import is_constant
+from mypyc.irbuild.util import bytes_from_str, is_constant
 from mypyc.options import CompilerOptions
 from mypyc.primitives.dict_ops import dict_get_item_op, dict_set_item_op
 from mypyc.primitives.generic_ops import iter_op, next_op, py_setattr_op
@@ -296,8 +296,7 @@ def load_bytes_from_str_literal(self, value: str) -> Value:
         are stored in BytesExpr.value, whose type is 'str' not 'bytes'.
         Thus we perform a special conversion here.
         """
-        bytes_value = bytes(value, "utf8").decode("unicode-escape").encode("raw-unicode-escape")
-        return self.builder.load_bytes(bytes_value)
+        return self.builder.load_bytes(bytes_from_str(value))
 
     def load_int(self, value: int) -> Value:
         return self.builder.load_int(value)
@@ -886,7 +885,7 @@ def get_dict_base_type(self, expr: Expression) -> Instance:
         This is useful for dict subclasses like SymbolTable.
         """
         target_type = get_proper_type(self.types[expr])
-        assert isinstance(target_type, Instance)
+        assert isinstance(target_type, Instance), target_type
         dict_base = next(base for base in target_type.type.mro if base.fullname == "builtins.dict")
         return map_instance_to_supertype(target_type, dict_base)
 

mypyc/irbuild/expression.py

@@ -6,7 +6,7 @@
 
 from __future__ import annotations
 
-from typing import Callable, cast
+from typing import Callable, Sequence, cast
 
 from mypy.nodes import (
     ARG_POS,
@@ -55,6 +55,7 @@
     ComparisonOp,
     Integer,
     LoadAddress,
+    LoadLiteral,
     RaiseStandardError,
     Register,
     TupleGet,
@@ -63,12 +64,14 @@
 )
 from mypyc.ir.rtypes import (
     RTuple,
+    bool_rprimitive,
     int_rprimitive,
     is_fixed_width_rtype,
     is_int_rprimitive,
     is_list_rprimitive,
     is_none_rprimitive,
     object_rprimitive,
+    set_rprimitive,
 )
 from mypyc.irbuild.ast_helpers import is_borrow_friendly_expr, process_conditional
 from mypyc.irbuild.builder import IRBuilder, int_borrow_friendly_op
@@ -86,14 +89,15 @@
     tokenizer_printf_style,
 )
 from mypyc.irbuild.specialize import apply_function_specialization, apply_method_specialization
+from mypyc.irbuild.util import bytes_from_str
 from mypyc.primitives.bytes_ops import bytes_slice_op
 from mypyc.primitives.dict_ops import dict_get_item_op, dict_new_op, dict_set_item_op
 from mypyc.primitives.generic_ops import iter_op
 from mypyc.primitives.int_ops import int_comparison_op_mapping
 from mypyc.primitives.list_ops import list_append_op, list_extend_op, list_slice_op
 from mypyc.primitives.misc_ops import ellipsis_op, get_module_dict_op, new_slice_op, type_op
 from mypyc.primitives.registry import CFunctionDescription, builtin_names
-from mypyc.primitives.set_ops import set_add_op, set_update_op
+from mypyc.primitives.set_ops import set_add_op, set_in_op, set_update_op
 from mypyc.primitives.str_ops import str_slice_op
 from mypyc.primitives.tuple_ops import list_tuple_op, tuple_slice_op
 
@@ -613,6 +617,54 @@ def transform_conditional_expr(builder: IRBuilder, expr: ConditionalExpr) -> Val
     return target
 
 
+def set_literal_values(builder: IRBuilder, items: Sequence[Expression]) -> list[object] | None:
+    values: list[object] = []
+    for item in items:
+        const_value = constant_fold_expr(builder, item)
+        if const_value is not None:
+            values.append(const_value)
+            continue
+
+        if isinstance(item, RefExpr):
+            if item.fullname == "builtins.None":
+                values.append(None)
+            elif item.fullname == "builtins.True":
+                values.append(True)
+            elif item.fullname == "builtins.False":
+                values.append(False)
+        elif isinstance(item, (BytesExpr, FloatExpr, ComplexExpr)):
+            # constant_fold_expr() doesn't handle these (yet?)
+            v = bytes_from_str(item.value) if isinstance(item, BytesExpr) else item.value
+            values.append(v)
+        elif isinstance(item, TupleExpr):
+            tuple_values = set_literal_values(builder, item.items)
+            if tuple_values is not None:
+                values.append(tuple(tuple_values))
+
+    if len(values) != len(items):
+        # Bail if not all items can be converted into values.
+        return None
+    return values
+
+
+def precompute_set_literal(builder: IRBuilder, s: SetExpr) -> Value | None:
+    """Try to pre-compute a frozenset literal during module initialization.
+
+    Return None if it's not possible.
+
+    Supported items:
+     - Anything supported by irbuild.constant_fold.constant_fold_expr()
+     - None, True, and False
+     - Float, byte, and complex literals
+     - Tuple literals with only items listed above
+    """
+    values = set_literal_values(builder, s.items)
+    if values is not None:
+        return builder.add(LoadLiteral(frozenset(values), set_rprimitive))
+
+    return None
+
+
 def transform_comparison_expr(builder: IRBuilder, e: ComparisonExpr) -> Value:
     # x in (...)/[...]
     # x not in (...)/[...]
@@ -666,6 +718,23 @@ def transform_comparison_expr(builder: IRBuilder, e: ComparisonExpr) -> Value:
             else:
                 return builder.true()
 
+    # x in {...}
+    # x not in {...}
+    if (
+        first_op in ("in", "not in")
+        and len(e.operators) == 1
+        and isinstance(e.operands[1], SetExpr)
+    ):
+        set_literal = precompute_set_literal(builder, e.operands[1])
+        if set_literal is not None:
+            lhs = e.operands[0]
+            result = builder.builder.call_c(
+                set_in_op, [builder.accept(lhs), set_literal], e.line, bool_rprimitive
+            )
+            if first_op == "not in":
+                return builder.unary_op(result, "not", e.line)
+            return result
+
     if len(e.operators) == 1:
         # Special some common simple cases
         if first_op in ("is", "is not"):

mypyc/irbuild/for_helpers.py

@@ -17,6 +17,7 @@
     Lvalue,
     MemberExpr,
     RefExpr,
+    SetExpr,
     TupleExpr,
     TypeAlias,
 )
@@ -469,12 +470,22 @@ def make_for_loop_generator(
             for_dict_gen.init(expr_reg, target_type)
             return for_dict_gen
 
+    iterable_expr_reg: Value | None = None
+    if isinstance(expr, SetExpr):
+        # Special case "for x in <set literal>".
+        from mypyc.irbuild.expression import precompute_set_literal
+
+        set_literal = precompute_set_literal(builder, expr)
+        if set_literal is not None:
+            iterable_expr_reg = set_literal
+
     # Default to a generic for loop.
-    expr_reg = builder.accept(expr)
+    if iterable_expr_reg is None:
+        iterable_expr_reg = builder.accept(expr)
     for_obj = ForIterable(builder, index, body_block, loop_exit, line, nested)
     item_type = builder._analyze_iterable_item_type(expr)
     item_rtype = builder.type_to_rtype(item_type)
-    for_obj.init(expr_reg, item_rtype)
+    for_obj.init(iterable_expr_reg, item_rtype)
     return for_obj
 
 

mypyc/irbuild/util.py

@@ -177,3 +177,13 @@ def is_constant(e: Expression) -> bool:
             )
         )
     )
+
+
+def bytes_from_str(value: str) -> bytes:
+    """Convert a string representing bytes into actual bytes.
+
+    This is needed because the literal characters of BytesExpr (the
+    characters inside b'') are stored in BytesExpr.value, whose type is
+    'str' not 'bytes'.
+    """
+    return bytes(value, "utf8").decode("unicode-escape").encode("raw-unicode-escape")