blob: 9328e8f5bdfac832eb6bef446f27f8a6c6fab994 [file] [log] [blame]
# Copyright 2019 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Symbol resolver for Emboss IR.
The resolve_symbols function should be used to generate canonical resolutions
for all symbol references in an Emboss IR.
"""
import collections
from compiler.util import error
from compiler.util import ir_data
from compiler.util import ir_data_utils
from compiler.util import ir_util
from compiler.util import traverse_ir
# TODO(bolms): Symbol resolution raises an exception at the first error, but
# this is one place where it can make sense to report multiple errors.
FileLocation = collections.namedtuple("FileLocation", ["file", "location"])
def ambiguous_name_error(file_name, location, name, candidate_locations):
"""A name cannot be resolved because there are two or more candidates."""
result = [error.error(file_name, location, "Ambiguous name '{}'".format(name))
]
for location in sorted(candidate_locations):
result.append(error.note(location.file, location.location,
"Possible resolution"))
return result
def duplicate_name_error(file_name, location, name, original_location):
"""A name is defined two or more times."""
return [error.error(file_name, location, "Duplicate name '{}'".format(name)),
error.note(original_location.file, original_location.location,
"Original definition")]
def missing_name_error(file_name, location, name):
return [error.error(file_name, location, "No candidate for '{}'".format(name))
]
def array_subfield_error(file_name, location, name):
return [error.error(file_name, location,
"Cannot access member of array '{}'".format(name))]
def noncomposite_subfield_error(file_name, location, name):
return [error.error(file_name, location,
"Cannot access member of noncomposite field '{}'".format(
name))]
def _nested_name(canonical_name, name):
"""Creates a new CanonicalName with name appended to the object_path."""
return ir_data.CanonicalName(
module_file=canonical_name.module_file,
object_path=list(canonical_name.object_path) + [name])
class _Scope(dict):
"""A _Scope holds data for a symbol.
A _Scope is a dict with some additional attributes. Lexically nested names
are kept in the dict, and bookkeeping is kept in the additional attributes.
For example, each module should have a child _Scope for each type contained in
the module. `struct` and `bits` types should have nested _Scopes for each
field; `enum` types should have nested scopes for each enumerated name.
Attributes:
canonical_name: The absolute name of this symbol; e.g. ("file.emb",
"TypeName", "SubTypeName", "field_name")
source_location: The ir_data.SourceLocation where this symbol is defined.
visibility: LOCAL, PRIVATE, or SEARCHABLE; see below.
alias: If set, this name is merely a pointer to another name.
"""
__slots__ = ("canonical_name", "source_location", "visibility", "alias")
# A LOCAL name is visible outside of its enclosing scope, but should not be
# found when searching for a name. That is, this name should be matched in
# the tail of a qualified reference (the 'bar' in 'foo.bar'), but not when
# searching for names (the 'foo' in 'foo.bar' should not match outside of
# 'foo's scope). This applies to public field names.
LOCAL = object()
# A PRIVATE name is similar to LOCAL except that it is never visible outside
# its enclosing scope. This applies to abbreviations of field names: if 'a'
# is an abbreviation for field 'apple', then 'foo.a' is not a valid reference;
# instead it should be 'foo.apple'.
PRIVATE = object()
# A SEARCHABLE name is visible as long as it is in a scope in the search list.
# This applies to type names ('Foo'), which may be found from many scopes.
SEARCHABLE = object()
def __init__(self, canonical_name, source_location, visibility, alias=None):
super(_Scope, self).__init__()
self.canonical_name = canonical_name
self.source_location = source_location
self.visibility = visibility
self.alias = alias
def _add_name_to_scope(name_ir, scope, canonical_name, visibility, errors):
"""Adds the given name_ir to the given scope."""
name = name_ir.text
new_scope = _Scope(canonical_name, name_ir.source_location, visibility)
if name in scope:
errors.append(duplicate_name_error(
scope.canonical_name.module_file, name_ir.source_location, name,
FileLocation(scope[name].canonical_name.module_file,
scope[name].source_location)))
else:
scope[name] = new_scope
return new_scope
def _add_name_to_scope_and_normalize(name_ir, scope, visibility, errors):
"""Adds the given name_ir to scope and sets its canonical_name."""
name = name_ir.name.text
canonical_name = _nested_name(scope.canonical_name, name)
ir_data_utils.builder(name_ir).canonical_name.CopyFrom(canonical_name)
return _add_name_to_scope(name_ir.name, scope, canonical_name, visibility,
errors)
def _add_struct_field_to_scope(field, scope, errors):
"""Adds the name of the given field to the scope."""
new_scope = _add_name_to_scope_and_normalize(field.name, scope, _Scope.LOCAL,
errors)
if field.HasField("abbreviation"):
_add_name_to_scope(field.abbreviation, scope, new_scope.canonical_name,
_Scope.PRIVATE, errors)
value_builtin_name = ir_data.Word(
text="this",
source_location=ir_data.Location(is_synthetic=True),
)
# In "inside field" scope, the name `this` maps back to the field itself.
# This is important for attributes like `[requires]`.
_add_name_to_scope(value_builtin_name, new_scope,
field.name.canonical_name, _Scope.PRIVATE, errors)
def _add_parameter_name_to_scope(parameter, scope, errors):
"""Adds the name of the given parameter to the scope."""
_add_name_to_scope_and_normalize(parameter.name, scope, _Scope.LOCAL, errors)
def _add_enum_value_to_scope(value, scope, errors):
"""Adds the name of the enum value to scope."""
_add_name_to_scope_and_normalize(value.name, scope, _Scope.LOCAL, errors)
def _add_type_name_to_scope(type_definition, scope, errors):
"""Adds the name of type_definition to the given scope."""
new_scope = _add_name_to_scope_and_normalize(type_definition.name, scope,
_Scope.SEARCHABLE, errors)
return {"scope": new_scope}
def _set_scope_for_type_definition(type_definition, scope):
"""Sets the current scope for an ir_data.AddressableUnit."""
return {"scope": scope[type_definition.name.name.text]}
def _add_module_to_scope(module, scope):
"""Adds the name of the module to the given scope."""
module_symbol_table = _Scope(
ir_data.CanonicalName(module_file=module.source_file_name,
object_path=[]),
None,
_Scope.SEARCHABLE)
scope[module.source_file_name] = module_symbol_table
return {"scope": scope[module.source_file_name]}
def _set_scope_for_module(module, scope):
"""Adds the name of the module to the given scope."""
return {"scope": scope[module.source_file_name]}
def _add_import_to_scope(foreign_import, table, module, errors):
if not foreign_import.local_name.text:
# This is the prelude import; ignore it.
return
_add_alias_to_scope(foreign_import.local_name, table, module.canonical_name,
[foreign_import.file_name.text], _Scope.SEARCHABLE,
errors)
def _construct_symbol_tables(ir):
"""Constructs per-module symbol tables for each module in ir."""
symbol_tables = {}
errors = []
traverse_ir.fast_traverse_ir_top_down(
ir, [ir_data.Module], _add_module_to_scope,
parameters={"errors": errors, "scope": symbol_tables})
traverse_ir.fast_traverse_ir_top_down(
ir, [ir_data.TypeDefinition], _add_type_name_to_scope,
incidental_actions={ir_data.Module: _set_scope_for_module},
parameters={"errors": errors, "scope": symbol_tables})
if errors:
# Ideally, we would find duplicate field names elsewhere in the module, even
# if there are duplicate type names, but field/enum names in the colliding
# types also end up colliding, leading to spurious errors. E.g., if you
# have two `struct Foo`s, then the field check will also discover a
# collision for `$size_in_bytes`, since there are two `Foo.$size_in_bytes`.
return symbol_tables, errors
traverse_ir.fast_traverse_ir_top_down(
ir, [ir_data.EnumValue], _add_enum_value_to_scope,
incidental_actions={
ir_data.Module: _set_scope_for_module,
ir_data.TypeDefinition: _set_scope_for_type_definition,
},
parameters={"errors": errors, "scope": symbol_tables})
traverse_ir.fast_traverse_ir_top_down(
ir, [ir_data.Field], _add_struct_field_to_scope,
incidental_actions={
ir_data.Module: _set_scope_for_module,
ir_data.TypeDefinition: _set_scope_for_type_definition,
},
parameters={"errors": errors, "scope": symbol_tables})
traverse_ir.fast_traverse_ir_top_down(
ir, [ir_data.RuntimeParameter], _add_parameter_name_to_scope,
incidental_actions={
ir_data.Module: _set_scope_for_module,
ir_data.TypeDefinition: _set_scope_for_type_definition,
},
parameters={"errors": errors, "scope": symbol_tables})
return symbol_tables, errors
def _add_alias_to_scope(name_ir, table, scope, alias, visibility, errors):
"""Adds the given name to the scope as an alias."""
name = name_ir.text
new_scope = _Scope(_nested_name(scope, name), name_ir.source_location,
visibility, alias)
scoped_table = table[scope.module_file]
for path_element in scope.object_path:
scoped_table = scoped_table[path_element]
if name in scoped_table:
errors.append(duplicate_name_error(
scoped_table.canonical_name.module_file, name_ir.source_location, name,
FileLocation(scoped_table[name].canonical_name.module_file,
scoped_table[name].source_location)))
else:
scoped_table[name] = new_scope
return new_scope
def _resolve_head_of_field_reference(field_reference, table, current_scope,
visible_scopes, source_file_name, errors):
return _resolve_reference(
field_reference.path[0], table, current_scope,
visible_scopes, source_file_name, errors)
def _resolve_reference(reference, table, current_scope, visible_scopes,
source_file_name, errors):
"""Sets the canonical name of the given reference."""
if reference.HasField("canonical_name"):
# This reference has already been resolved by the _resolve_field_reference
# pass.
return
target = _find_target_of_reference(reference, table, current_scope,
visible_scopes, source_file_name, errors)
if target is not None:
assert not target.alias
ir_data_utils.builder(reference).canonical_name.CopyFrom(target.canonical_name)
def _find_target_of_reference(reference, table, current_scope, visible_scopes,
source_file_name, errors):
"""Returns the resolved name of the given reference."""
found_in_table = None
name = reference.source_name[0].text
for scope in visible_scopes:
scoped_table = table[scope.module_file]
for path_element in scope.object_path or []:
scoped_table = scoped_table[path_element]
if (name in scoped_table and
(scope == current_scope or
scoped_table[name].visibility == _Scope.SEARCHABLE)):
# Prelude is "", so explicitly check for None.
if found_in_table is not None:
# TODO(bolms): Currently, this catches the case where a module tries to
# use a name that is defined (at the same scope) in two different
# modules. It may make sense to raise duplicate_name_error whenever two
# modules define the same name (whether it is used or not), and reserve
# ambiguous_name_error for cases where a name is found in multiple
# scopes.
errors.append(ambiguous_name_error(
source_file_name, reference.source_location, name, [FileLocation(
found_in_table[name].canonical_name.module_file,
found_in_table[name].source_location), FileLocation(
scoped_table[name].canonical_name.module_file, scoped_table[
name].source_location)]))
continue
found_in_table = scoped_table
if reference.is_local_name:
# This is a little hacky. When "is_local_name" is True, the name refers
# to a type that was defined inline. In many cases, the type should be
# found at the same scope as the field; e.g.:
#
# struct Foo:
# 0 [+1] enum bar:
# BAZ = 1
#
# In this case, `Foo.bar` has type `Foo.Bar`. Unfortunately, things
# break down a little bit when there is an inline type in an anonymous
# `bits`:
#
# struct Foo:
# 0 [+1] bits:
# 0 [+7] enum bar:
# BAZ = 1
#
# Types inside of anonymous `bits` are hoisted into their parent type,
# so instead of `Foo.EmbossReservedAnonymous1.Bar`, `bar`'s type is just
# `Foo.Bar`. Unfortunately, the field is still
# `Foo.EmbossReservedAnonymous1.bar`, so `bar`'s type won't be found in
# `bar`'s `current_scope`.
#
# (The name `bar` is exposed from `Foo` as an alias virtual field, so
# perhaps the correct answer is to allow type aliases, so that `Bar` can
# be found in both `Foo` and `Foo.EmbossReservedAnonymous1`. That would
# involve an entirely new feature, though.)
#
# The workaround here is to search scopes from the innermost outward,
# and just stop as soon as a match is found. This isn't ideal, because
# it relies on other bits of the front end having correctly added the
# inline type to the correct scope before symbol resolution, but it does
# work. Names with False `is_local_name` will still be checked for
# ambiguity.
break
if found_in_table is None:
errors.append(missing_name_error(
source_file_name, reference.source_name[0].source_location, name))
if not errors:
for subname in reference.source_name:
if subname.text not in found_in_table:
errors.append(missing_name_error(source_file_name,
subname.source_location, subname.text))
return None
found_in_table = found_in_table[subname.text]
while found_in_table.alias:
referenced_table = table
for name in found_in_table.alias:
referenced_table = referenced_table[name]
# TODO(bolms): This section should really be a recursive lookup
# function, which would be able to handle arbitrary aliases through
# other aliases.
#
# This should be fine for now, since the only aliases here should be
# imports, which can't refer to other imports.
assert not referenced_table.alias, "Alias found to contain alias."
found_in_table = referenced_table
return found_in_table
return None
def _resolve_field_reference(field_reference, source_file_name, errors, ir):
"""Resolves the References inside of a FieldReference."""
if field_reference.path[-1].HasField("canonical_name"):
# Already done.
return
previous_field = ir_util.find_object_or_none(field_reference.path[0], ir)
previous_reference = field_reference.path[0]
for ref in field_reference.path[1:]:
while ir_util.field_is_virtual(previous_field):
if (previous_field.read_transform.WhichOneof("expression") ==
"field_reference"):
# Pass a separate error list into the recursive _resolve_field_reference
# call so that only one copy of the error for a particular reference
# will actually surface: in particular, the one that results from a
# direct call from traverse_ir_top_down into _resolve_field_reference.
new_errors = []
_resolve_field_reference(
previous_field.read_transform.field_reference,
previous_field.name.canonical_name.module_file, new_errors, ir)
# If the recursive _resolve_field_reference was unable to resolve the
# field, then bail. Otherwise we get a cascade of errors, where an
# error in `x` leads to errors in anything trying to reach a member of
# `x`.
if not previous_field.read_transform.field_reference.path[-1].HasField(
"canonical_name"):
return
previous_field = ir_util.find_object(
previous_field.read_transform.field_reference.path[-1], ir)
else:
errors.append(
noncomposite_subfield_error(source_file_name,
previous_reference.source_location,
previous_reference.source_name[0].text))
return
if previous_field.type.WhichOneof("type") == "array_type":
errors.append(
array_subfield_error(source_file_name,
previous_reference.source_location,
previous_reference.source_name[0].text))
return
assert previous_field.type.WhichOneof("type") == "atomic_type"
member_name = ir_data_utils.copy(
previous_field.type.atomic_type.reference.canonical_name)
ir_data_utils.builder(member_name).object_path.extend([ref.source_name[0].text])
previous_field = ir_util.find_object_or_none(member_name, ir)
if previous_field is None:
errors.append(
missing_name_error(source_file_name,
ref.source_name[0].source_location,
ref.source_name[0].text))
return
ir_data_utils.builder(ref).canonical_name.CopyFrom(member_name)
previous_reference = ref
def _set_visible_scopes_for_type_definition(type_definition, visible_scopes):
"""Sets current_scope and visible_scopes for the given type_definition."""
return {
"current_scope": type_definition.name.canonical_name,
# In order to ensure that the iteration through scopes in
# _find_target_of_reference will go from innermost to outermost, it is
# important that the current scope (type_definition.name.canonical_name)
# precedes the previous visible_scopes here.
"visible_scopes": (type_definition.name.canonical_name,) + visible_scopes,
}
def _set_visible_scopes_for_module(module):
"""Sets visible_scopes for the given module."""
self_scope = ir_data.CanonicalName(module_file=module.source_file_name)
extra_visible_scopes = []
for foreign_import in module.foreign_import:
# Anonymous imports are searched for top-level names; named imports are not.
# As of right now, only the prelude should be imported anonymously; other
# modules must be imported with names.
if not foreign_import.local_name.text:
extra_visible_scopes.append(
ir_data.CanonicalName(module_file=foreign_import.file_name.text))
return {"visible_scopes": (self_scope,) + tuple(extra_visible_scopes)}
def _set_visible_scopes_for_attribute(attribute, field, visible_scopes):
"""Sets current_scope and visible_scopes for the attribute."""
del attribute # Unused
if field is None:
return
return {
"current_scope": field.name.canonical_name,
"visible_scopes": (field.name.canonical_name,) + visible_scopes,
}
def _module_source_from_table_action(m, table):
return {"module": table[m.source_file_name]}
def _resolve_symbols_from_table(ir, table):
"""Resolves all references in the given IR, given the constructed table."""
errors = []
# Symbol resolution is broken into five passes. First, this code resolves any
# imports, and adds import aliases to modules.
traverse_ir.fast_traverse_ir_top_down(
ir, [ir_data.Import], _add_import_to_scope,
incidental_actions={
ir_data.Module: _module_source_from_table_action,
},
parameters={"errors": errors, "table": table})
if errors:
return errors
# Next, this resolves all absolute references (e.g., it resolves "UInt" in
# "0:1 UInt field" to [prelude]::UInt).
traverse_ir.fast_traverse_ir_top_down(
ir, [ir_data.Reference], _resolve_reference,
skip_descendants_of=(ir_data.FieldReference,),
incidental_actions={
ir_data.TypeDefinition: _set_visible_scopes_for_type_definition,
ir_data.Module: _set_visible_scopes_for_module,
ir_data.Attribute: _set_visible_scopes_for_attribute,
},
parameters={"table": table, "errors": errors, "field": None})
# Lastly, head References to fields (e.g., the `a` of `a.b.c`) are resolved.
traverse_ir.fast_traverse_ir_top_down(
ir, [ir_data.FieldReference], _resolve_head_of_field_reference,
incidental_actions={
ir_data.TypeDefinition: _set_visible_scopes_for_type_definition,
ir_data.Module: _set_visible_scopes_for_module,
ir_data.Attribute: _set_visible_scopes_for_attribute,
},
parameters={"table": table, "errors": errors, "field": None})
return errors
def resolve_field_references(ir):
"""Resolves structure member accesses ("field.subfield") in ir."""
errors = []
traverse_ir.fast_traverse_ir_top_down(
ir, [ir_data.FieldReference], _resolve_field_reference,
incidental_actions={
ir_data.TypeDefinition: _set_visible_scopes_for_type_definition,
ir_data.Module: _set_visible_scopes_for_module,
ir_data.Attribute: _set_visible_scopes_for_attribute,
},
parameters={"errors": errors, "field": None})
return errors
def resolve_symbols(ir):
"""Resolves the symbols in all modules in ir."""
symbol_tables, errors = _construct_symbol_tables(ir)
if errors:
return errors
return _resolve_symbols_from_table(ir, symbol_tables)