usse/funda-scraper/venv/lib/python3.10/site-packages/mypyc/codegen/emitmodule.py

1152 lines
46 KiB
Python

"""Generate C code for a Python C extension module from Python source code."""
# FIXME: Basically nothing in this file operates on the level of a
# single module and it should be renamed.
import os
import json
from mypy.backports import OrderedDict
from typing import List, Tuple, Dict, Iterable, Set, TypeVar, Optional
from mypy.nodes import MypyFile
from mypy.build import (
BuildSource, BuildResult, State, build, sorted_components, get_cache_names,
create_metastore, compute_hash,
)
from mypy.errors import CompileError
from mypy.options import Options
from mypy.plugin import Plugin, ReportConfigContext
from mypy.fscache import FileSystemCache
from mypy.util import hash_digest
from mypyc.irbuild.main import build_ir
from mypyc.irbuild.prepare import load_type_map
from mypyc.irbuild.mapper import Mapper
from mypyc.common import (
PREFIX, TOP_LEVEL_NAME, MODULE_PREFIX, RUNTIME_C_FILES, short_id_from_name, use_fastcall,
use_vectorcall, shared_lib_name,
)
from mypyc.codegen.cstring import c_string_initializer
from mypyc.codegen.literals import Literals
from mypyc.codegen.emit import EmitterContext, Emitter, HeaderDeclaration
from mypyc.codegen.emitfunc import generate_native_function, native_function_header
from mypyc.codegen.emitclass import generate_class_type_decl, generate_class
from mypyc.codegen.emitwrapper import (
generate_wrapper_function, wrapper_function_header,
generate_legacy_wrapper_function, legacy_wrapper_function_header,
)
from mypyc.ir.ops import DeserMaps, LoadLiteral
from mypyc.ir.rtypes import RType, RTuple
from mypyc.ir.func_ir import FuncIR
from mypyc.ir.class_ir import ClassIR
from mypyc.ir.module_ir import ModuleIR, ModuleIRs, deserialize_modules
from mypyc.options import CompilerOptions
from mypyc.transform.uninit import insert_uninit_checks
from mypyc.transform.refcount import insert_ref_count_opcodes
from mypyc.transform.exceptions import insert_exception_handling
from mypyc.namegen import NameGenerator, exported_name
from mypyc.errors import Errors
# All of the modules being compiled are divided into "groups". A group
# is a set of modules that are placed into the same shared library.
# Two common configurations are that every module is placed in a group
# by itself (fully separate compilation) and that every module is
# placed in the same group (fully whole-program compilation), but we
# support finer-grained control of the group as well.
#
# In fully whole-program compilation, we will generate N+1 extension
# modules: one shim per module and one shared library containing all
# the actual code.
# In fully separate compilation, we (unfortunately) will generate 2*N
# extension modules: one shim per module and also one library containing
# each module's actual code. (This might be fixable in the future,
# but allows a clean separation between setup of the export tables
# (see generate_export_table) and running module top levels.)
#
# A group is represented as a list of BuildSources containing all of
# its modules along with the name of the group. (Which can be None
# only if we are compiling only a single group with a single file in it
# and not using shared libraries).
Group = Tuple[List[BuildSource], Optional[str]]
Groups = List[Group]
# A list of (file name, file contents) pairs.
FileContents = List[Tuple[str, str]]
class MarkedDeclaration:
"""Add a mark, useful for topological sort."""
def __init__(self, declaration: HeaderDeclaration, mark: bool) -> None:
self.declaration = declaration
self.mark = False
class MypycPlugin(Plugin):
"""Plugin for making mypyc interoperate properly with mypy incremental mode.
Basically the point of this plugin is to force mypy to recheck things
based on the demands of mypyc in a couple situations:
* Any modules in the same group must be compiled together, so we
tell mypy that modules depend on all their groupmates.
* If the IR metadata is missing or stale or any of the generated
C source files associated missing or stale, then we need to
recompile the module so we mark it as stale.
"""
def __init__(
self, options: Options, compiler_options: CompilerOptions, groups: Groups) -> None:
super().__init__(options)
self.group_map: Dict[str, Tuple[Optional[str], List[str]]] = {}
for sources, name in groups:
modules = sorted(source.module for source in sources)
for id in modules:
self.group_map[id] = (name, modules)
self.compiler_options = compiler_options
self.metastore = create_metastore(options)
def report_config_data(
self, ctx: ReportConfigContext) -> Optional[Tuple[Optional[str], List[str]]]:
# The config data we report is the group map entry for the module.
# If the data is being used to check validity, we do additional checks
# that the IR cache exists and matches the metadata cache and all
# output source files exist and are up to date.
id, path, is_check = ctx.id, ctx.path, ctx.is_check
if id not in self.group_map:
return None
# If we aren't doing validity checks, just return the cache data
if not is_check:
return self.group_map[id]
# Load the metadata and IR cache
meta_path, _, _ = get_cache_names(id, path, self.options)
ir_path = get_ir_cache_name(id, path, self.options)
try:
meta_json = self.metastore.read(meta_path)
ir_json = self.metastore.read(ir_path)
except FileNotFoundError:
# This could happen if mypyc failed after mypy succeeded
# in the previous run or if some cache files got
# deleted. No big deal, just fail to load the cache.
return None
ir_data = json.loads(ir_json)
# Check that the IR cache matches the metadata cache
if compute_hash(meta_json) != ir_data['meta_hash']:
return None
# Check that all of the source files are present and as
# expected. The main situation where this would come up is the
# user deleting the build directory without deleting
# .mypy_cache, which we should handle gracefully.
for path, hash in ir_data['src_hashes'].items():
try:
with open(os.path.join(self.compiler_options.target_dir, path), 'rb') as f:
contents = f.read()
except FileNotFoundError:
return None
real_hash = hash_digest(contents)
if hash != real_hash:
return None
return self.group_map[id]
def get_additional_deps(self, file: MypyFile) -> List[Tuple[int, str, int]]:
# Report dependency on modules in the module's group
return [(10, id, -1) for id in self.group_map.get(file.fullname, (None, []))[1]]
def parse_and_typecheck(
sources: List[BuildSource],
options: Options,
compiler_options: CompilerOptions,
groups: Groups,
fscache: Optional[FileSystemCache] = None,
alt_lib_path: Optional[str] = None
) -> BuildResult:
assert options.strict_optional, 'strict_optional must be turned on'
result = build(sources=sources,
options=options,
alt_lib_path=alt_lib_path,
fscache=fscache,
extra_plugins=[MypycPlugin(options, compiler_options, groups)])
if result.errors:
raise CompileError(result.errors)
return result
def compile_scc_to_ir(
scc: List[MypyFile],
result: BuildResult,
mapper: Mapper,
compiler_options: CompilerOptions,
errors: Errors,
) -> ModuleIRs:
"""Compile an SCC into ModuleIRs.
Any modules that this SCC depends on must have either compiled or
loaded from a cache into mapper.
Arguments:
scc: The list of MypyFiles to compile
result: The BuildResult from the mypy front-end
mapper: The Mapper object mapping mypy ASTs to class and func IRs
compiler_options: The compilation options
errors: Where to report any errors encountered
Returns the IR of the modules.
"""
if compiler_options.verbose:
print("Compiling {}".format(", ".join(x.name for x in scc)))
# Generate basic IR, with missing exception and refcount handling.
modules = build_ir(
scc, result.graph, result.types, mapper, compiler_options, errors
)
if errors.num_errors > 0:
return modules
# Insert uninit checks.
for module in modules.values():
for fn in module.functions:
insert_uninit_checks(fn)
# Insert exception handling.
for module in modules.values():
for fn in module.functions:
insert_exception_handling(fn)
# Insert refcount handling.
for module in modules.values():
for fn in module.functions:
insert_ref_count_opcodes(fn)
return modules
def compile_modules_to_ir(
result: BuildResult,
mapper: Mapper,
compiler_options: CompilerOptions,
errors: Errors,
) -> ModuleIRs:
"""Compile a collection of modules into ModuleIRs.
The modules to compile are specified as part of mapper's group_map.
Returns the IR of the modules.
"""
deser_ctx = DeserMaps({}, {})
modules = {}
# Process the graph by SCC in topological order, like we do in mypy.build
for scc in sorted_components(result.graph):
scc_states = [result.graph[id] for id in scc]
trees = [st.tree for st in scc_states if st.id in mapper.group_map and st.tree]
if not trees:
continue
fresh = all(id not in result.manager.rechecked_modules for id in scc)
if fresh:
load_scc_from_cache(trees, result, mapper, deser_ctx)
else:
scc_ir = compile_scc_to_ir(trees, result, mapper, compiler_options, errors)
modules.update(scc_ir)
return modules
def compile_ir_to_c(
groups: Groups,
modules: ModuleIRs,
result: BuildResult,
mapper: Mapper,
compiler_options: CompilerOptions,
) -> Dict[Optional[str], List[Tuple[str, str]]]:
"""Compile a collection of ModuleIRs to C source text.
Returns a dictionary mapping group names to a list of (file name,
file text) pairs.
"""
source_paths = {source.module: result.graph[source.module].xpath
for sources, _ in groups for source in sources}
names = NameGenerator([[source.module for source in sources] for sources, _ in groups])
# Generate C code for each compilation group. Each group will be
# compiled into a separate extension module.
ctext: Dict[Optional[str], List[Tuple[str, str]]] = {}
for group_sources, group_name in groups:
group_modules = [(source.module, modules[source.module]) for source in group_sources
if source.module in modules]
if not group_modules:
ctext[group_name] = []
continue
generator = GroupGenerator(
group_modules, source_paths,
group_name, mapper.group_map, names,
compiler_options
)
ctext[group_name] = generator.generate_c_for_modules()
return ctext
def get_ir_cache_name(id: str, path: str, options: Options) -> str:
meta_path, _, _ = get_cache_names(id, path, options)
return meta_path.replace('.meta.json', '.ir.json')
def get_state_ir_cache_name(state: State) -> str:
return get_ir_cache_name(state.id, state.xpath, state.options)
def write_cache(
modules: ModuleIRs,
result: BuildResult,
group_map: Dict[str, Optional[str]],
ctext: Dict[Optional[str], List[Tuple[str, str]]],
) -> None:
"""Write out the cache information for modules.
Each module has the following cache information written (which is
in addition to the cache information written by mypy itself):
* A serialized version of its mypyc IR, minus the bodies of
functions. This allows code that depends on it to use
these serialized data structures when compiling against it
instead of needing to recompile it. (Compiling against a
module requires access to both its mypy and mypyc data
structures.)
* The hash of the mypy metadata cache file for the module.
This is used to ensure that the mypyc cache and the mypy
cache are in sync and refer to the same version of the code.
This is particularly important if mypyc crashes/errors/is
stopped after mypy has written its cache but before mypyc has.
* The hashes of all of the source file outputs for the group
the module is in. This is so that the module will be
recompiled if the source outputs are missing.
"""
hashes = {}
for name, files in ctext.items():
hashes[name] = {file: compute_hash(data) for file, data in files}
# Write out cache data
for id, module in modules.items():
st = result.graph[id]
meta_path, _, _ = get_cache_names(id, st.xpath, result.manager.options)
# If the metadata isn't there, skip writing the cache.
try:
meta_data = result.manager.metastore.read(meta_path)
except OSError:
continue
newpath = get_state_ir_cache_name(st)
ir_data = {
'ir': module.serialize(),
'meta_hash': compute_hash(meta_data),
'src_hashes': hashes[group_map[id]],
}
result.manager.metastore.write(newpath, json.dumps(ir_data))
result.manager.metastore.commit()
def load_scc_from_cache(
scc: List[MypyFile],
result: BuildResult,
mapper: Mapper,
ctx: DeserMaps,
) -> ModuleIRs:
"""Load IR for an SCC of modules from the cache.
Arguments and return are as compile_scc_to_ir.
"""
cache_data = {
k.fullname: json.loads(
result.manager.metastore.read(get_state_ir_cache_name(result.graph[k.fullname]))
)['ir'] for k in scc
}
modules = deserialize_modules(cache_data, ctx)
load_type_map(mapper, scc, ctx)
return modules
def compile_modules_to_c(
result: BuildResult,
compiler_options: CompilerOptions,
errors: Errors,
groups: Groups,
) -> Tuple[ModuleIRs, List[FileContents]]:
"""Compile Python module(s) to the source of Python C extension modules.
This generates the source code for the "shared library" module
for each group. The shim modules are generated in mypyc.build.
Each shared library module provides, for each module in its group,
a PyCapsule containing an initialization function.
Additionally, it provides a capsule containing an export table of
pointers to all of the group's functions and static variables.
Arguments:
result: The BuildResult from the mypy front-end
compiler_options: The compilation options
errors: Where to report any errors encountered
groups: The groups that we are compiling. See documentation of Groups type above.
ops: Optionally, where to dump stringified ops for debugging.
Returns the IR of the modules and a list containing the generated files for each group.
"""
# Construct a map from modules to what group they belong to
group_map = {source.module: lib_name for group, lib_name in groups for source in group}
mapper = Mapper(group_map)
# Sometimes when we call back into mypy, there might be errors.
# We don't want to crash when that happens.
result.manager.errors.set_file('<mypyc>', module=None, scope=None)
modules = compile_modules_to_ir(result, mapper, compiler_options, errors)
ctext = compile_ir_to_c(groups, modules, result, mapper, compiler_options)
if errors.num_errors == 0:
write_cache(modules, result, group_map, ctext)
return modules, [ctext[name] for _, name in groups]
def generate_function_declaration(fn: FuncIR, emitter: Emitter) -> None:
emitter.context.declarations[emitter.native_function_name(fn.decl)] = HeaderDeclaration(
f'{native_function_header(fn.decl, emitter)};',
needs_export=True)
if fn.name != TOP_LEVEL_NAME:
if is_fastcall_supported(fn, emitter.capi_version):
emitter.context.declarations[PREFIX + fn.cname(emitter.names)] = HeaderDeclaration(
f'{wrapper_function_header(fn, emitter.names)};')
else:
emitter.context.declarations[PREFIX + fn.cname(emitter.names)] = HeaderDeclaration(
f'{legacy_wrapper_function_header(fn, emitter.names)};')
def pointerize(decl: str, name: str) -> str:
"""Given a C decl and its name, modify it to be a declaration to a pointer."""
# This doesn't work in general but does work for all our types...
if '(' in decl:
# Function pointer. Stick an * in front of the name and wrap it in parens.
return decl.replace(name, f'(*{name})')
else:
# Non-function pointer. Just stick an * in front of the name.
return decl.replace(name, f'*{name}')
def group_dir(group_name: str) -> str:
"""Given a group name, return the relative directory path for it. """
return os.sep.join(group_name.split('.')[:-1])
class GroupGenerator:
def __init__(self,
modules: List[Tuple[str, ModuleIR]],
source_paths: Dict[str, str],
group_name: Optional[str],
group_map: Dict[str, Optional[str]],
names: NameGenerator,
compiler_options: CompilerOptions) -> None:
"""Generator for C source for a compilation group.
The code for a compilation group contains an internal and an
external .h file, and then one .c if not in multi_file mode or
one .c file per module if in multi_file mode.)
Arguments:
modules: (name, ir) pairs for each module in the group
source_paths: Map from module names to source file paths
group_name: The name of the group (or None if this is single-module compilation)
group_map: A map of modules to their group names
names: The name generator for the compilation
multi_file: Whether to put each module in its own source file regardless
of group structure.
"""
self.modules = modules
self.source_paths = source_paths
self.context = EmitterContext(names, group_name, group_map)
self.names = names
# Initializations of globals to simple values that we can't
# do statically because the windows loader is bad.
self.simple_inits: List[Tuple[str, str]] = []
self.group_name = group_name
self.use_shared_lib = group_name is not None
self.compiler_options = compiler_options
self.multi_file = compiler_options.multi_file
@property
def group_suffix(self) -> str:
return '_' + exported_name(self.group_name) if self.group_name else ''
@property
def short_group_suffix(self) -> str:
return '_' + exported_name(self.group_name.split('.')[-1]) if self.group_name else ''
def generate_c_for_modules(self) -> List[Tuple[str, str]]:
file_contents = []
multi_file = self.use_shared_lib and self.multi_file
# Collect all literal refs in IR.
for _, module in self.modules:
for fn in module.functions:
collect_literals(fn, self.context.literals)
base_emitter = Emitter(self.context)
# Optionally just include the runtime library c files to
# reduce the number of compiler invocations needed
if self.compiler_options.include_runtime_files:
for name in RUNTIME_C_FILES:
base_emitter.emit_line(f'#include "{name}"')
base_emitter.emit_line(f'#include "__native{self.short_group_suffix}.h"')
base_emitter.emit_line(f'#include "__native_internal{self.short_group_suffix}.h"')
emitter = base_emitter
self.generate_literal_tables()
for module_name, module in self.modules:
if multi_file:
emitter = Emitter(self.context)
emitter.emit_line(f'#include "__native{self.short_group_suffix}.h"')
emitter.emit_line(
f'#include "__native_internal{self.short_group_suffix}.h"')
self.declare_module(module_name, emitter)
self.declare_internal_globals(module_name, emitter)
self.declare_imports(module.imports, emitter)
for cl in module.classes:
if cl.is_ext_class:
generate_class(cl, module_name, emitter)
# Generate Python extension module definitions and module initialization functions.
self.generate_module_def(emitter, module_name, module)
for fn in module.functions:
emitter.emit_line()
generate_native_function(fn, emitter, self.source_paths[module_name], module_name)
if fn.name != TOP_LEVEL_NAME:
emitter.emit_line()
if is_fastcall_supported(fn, emitter.capi_version):
generate_wrapper_function(
fn, emitter, self.source_paths[module_name], module_name)
else:
generate_legacy_wrapper_function(
fn, emitter, self.source_paths[module_name], module_name)
if multi_file:
name = (f'__native_{emitter.names.private_name(module_name)}.c')
file_contents.append((name, ''.join(emitter.fragments)))
# The external header file contains type declarations while
# the internal contains declarations of functions and objects
# (which are shared between shared libraries via dynamic
# exports tables and not accessed directly.)
ext_declarations = Emitter(self.context)
ext_declarations.emit_line(f'#ifndef MYPYC_NATIVE{self.group_suffix}_H')
ext_declarations.emit_line(f'#define MYPYC_NATIVE{self.group_suffix}_H')
ext_declarations.emit_line('#include <Python.h>')
ext_declarations.emit_line('#include <CPy.h>')
declarations = Emitter(self.context)
declarations.emit_line(f'#ifndef MYPYC_NATIVE_INTERNAL{self.group_suffix}_H')
declarations.emit_line(f'#define MYPYC_NATIVE_INTERNAL{self.group_suffix}_H')
declarations.emit_line('#include <Python.h>')
declarations.emit_line('#include <CPy.h>')
declarations.emit_line(f'#include "__native{self.short_group_suffix}.h"')
declarations.emit_line()
declarations.emit_line('int CPyGlobalsInit(void);')
declarations.emit_line()
for module_name, module in self.modules:
self.declare_finals(module_name, module.final_names, declarations)
for cl in module.classes:
generate_class_type_decl(cl, emitter, ext_declarations, declarations)
for fn in module.functions:
generate_function_declaration(fn, declarations)
for lib in sorted(self.context.group_deps):
elib = exported_name(lib)
short_lib = exported_name(lib.split('.')[-1])
declarations.emit_lines(
'#include <{}>'.format(
os.path.join(group_dir(lib), f"__native_{short_lib}.h")
),
f'struct export_table_{elib} exports_{elib};'
)
sorted_decls = self.toposort_declarations()
emitter = base_emitter
self.generate_globals_init(emitter)
emitter.emit_line()
for declaration in sorted_decls:
decls = ext_declarations if declaration.is_type else declarations
if not declaration.is_type:
decls.emit_lines(
f'extern {declaration.decl[0]}', *declaration.decl[1:])
# If there is a definition, emit it. Otherwise repeat the declaration
# (without an extern).
if declaration.defn:
emitter.emit_lines(*declaration.defn)
else:
emitter.emit_lines(*declaration.decl)
else:
decls.emit_lines(*declaration.decl)
if self.group_name:
self.generate_export_table(ext_declarations, emitter)
self.generate_shared_lib_init(emitter)
ext_declarations.emit_line('#endif')
declarations.emit_line('#endif')
output_dir = group_dir(self.group_name) if self.group_name else ''
return file_contents + [
(os.path.join(output_dir, f'__native{self.short_group_suffix}.c'),
''.join(emitter.fragments)),
(os.path.join(output_dir, f'__native_internal{self.short_group_suffix}.h'),
''.join(declarations.fragments)),
(os.path.join(output_dir, f'__native{self.short_group_suffix}.h'),
''.join(ext_declarations.fragments)),
]
def generate_literal_tables(self) -> None:
"""Generate tables containing descriptions of Python literals to construct.
We will store the constructed literals in a single array that contains
literals of all types. This way we can refer to an arbitrary literal by
its index.
"""
literals = self.context.literals
# During module initialization we store all the constructed objects here
self.declare_global('PyObject *[%d]' % literals.num_literals(), 'CPyStatics')
# Descriptions of str literals
init_str = c_string_array_initializer(literals.encoded_str_values())
self.declare_global('const char * const []', 'CPyLit_Str', initializer=init_str)
# Descriptions of bytes literals
init_bytes = c_string_array_initializer(literals.encoded_bytes_values())
self.declare_global('const char * const []', 'CPyLit_Bytes', initializer=init_bytes)
# Descriptions of int literals
init_int = c_string_array_initializer(literals.encoded_int_values())
self.declare_global('const char * const []', 'CPyLit_Int', initializer=init_int)
# Descriptions of float literals
init_floats = c_array_initializer(literals.encoded_float_values())
self.declare_global('const double []', 'CPyLit_Float', initializer=init_floats)
# Descriptions of complex literals
init_complex = c_array_initializer(literals.encoded_complex_values())
self.declare_global('const double []', 'CPyLit_Complex', initializer=init_complex)
# Descriptions of tuple literals
init_tuple = c_array_initializer(literals.encoded_tuple_values())
self.declare_global('const int []', 'CPyLit_Tuple', initializer=init_tuple)
def generate_export_table(self, decl_emitter: Emitter, code_emitter: Emitter) -> None:
"""Generate the declaration and definition of the group's export struct.
To avoid needing to deal with deeply platform specific issues
involving dynamic library linking (and some possibly
insurmountable issues involving cyclic dependencies), compiled
code accesses functions and data in other compilation groups
via an explicit "export struct".
Each group declares a struct type that contains a pointer to
every function and static variable it exports. It then
populates this struct and stores a pointer to it in a capsule
stored as an attribute named 'exports' on the group's shared
library's python module.
On load, a group's init function will import all of its
dependencies' exports tables using the capsule mechanism and
copy the contents into a local copy of the table (to eliminate
the need for a pointer indirection when accessing it).
Then, all calls to functions in another group and accesses to statics
from another group are done indirectly via the export table.
For example, a group containing a module b, where b contains a class B
and a function bar, would declare an export table like:
struct export_table_b {
PyTypeObject **CPyType_B;
PyObject *(*CPyDef_B)(CPyTagged cpy_r_x);
CPyTagged (*CPyDef_B___foo)(PyObject *cpy_r_self, CPyTagged cpy_r_y);
tuple_T2OI (*CPyDef_bar)(PyObject *cpy_r_x);
char (*CPyDef___top_level__)(void);
};
that would be initialized with:
static struct export_table_b exports = {
&CPyType_B,
&CPyDef_B,
&CPyDef_B___foo,
&CPyDef_bar,
&CPyDef___top_level__,
};
To call `b.foo`, then, a function in another group would do
`exports_b.CPyDef_bar(...)`.
"""
decls = decl_emitter.context.declarations
decl_emitter.emit_lines(
'',
f'struct export_table{self.group_suffix} {{',
)
for name, decl in decls.items():
if decl.needs_export:
decl_emitter.emit_line(pointerize('\n'.join(decl.decl), name))
decl_emitter.emit_line('};')
code_emitter.emit_lines(
'',
f'static struct export_table{self.group_suffix} exports = {{',
)
for name, decl in decls.items():
if decl.needs_export:
code_emitter.emit_line(f'&{name},')
code_emitter.emit_line('};')
def generate_shared_lib_init(self, emitter: Emitter) -> None:
"""Generate the init function for a shared library.
A shared library contains all of the actual code for a
compilation group.
The init function is responsible for creating Capsules that
wrap pointers to the initialization function of all the real
init functions for modules in this shared library as well as
the export table containing all of the exported functions and
values from all the modules.
These capsules are stored in attributes of the shared library.
"""
assert self.group_name is not None
emitter.emit_line()
emitter.emit_lines(
'PyMODINIT_FUNC PyInit_{}(void)'.format(
shared_lib_name(self.group_name).split('.')[-1]),
'{',
('static PyModuleDef def = {{ PyModuleDef_HEAD_INIT, "{}", NULL, -1, NULL, NULL }};'
.format(shared_lib_name(self.group_name))),
'int res;',
'PyObject *capsule;',
'PyObject *tmp;',
'static PyObject *module;',
'if (module) {',
'Py_INCREF(module);',
'return module;',
'}',
'module = PyModule_Create(&def);',
'if (!module) {',
'goto fail;',
'}',
'',
)
emitter.emit_lines(
'capsule = PyCapsule_New(&exports, "{}.exports", NULL);'.format(
shared_lib_name(self.group_name)),
'if (!capsule) {',
'goto fail;',
'}',
'res = PyObject_SetAttrString(module, "exports", capsule);',
'Py_DECREF(capsule);',
'if (res < 0) {',
'goto fail;',
'}',
'',
)
for mod, _ in self.modules:
name = exported_name(mod)
emitter.emit_lines(
f'extern PyObject *CPyInit_{name}(void);',
'capsule = PyCapsule_New((void *)CPyInit_{}, "{}.init_{}", NULL);'.format(
name, shared_lib_name(self.group_name), name),
'if (!capsule) {',
'goto fail;',
'}',
f'res = PyObject_SetAttrString(module, "init_{name}", capsule);',
'Py_DECREF(capsule);',
'if (res < 0) {',
'goto fail;',
'}',
'',
)
for group in sorted(self.context.group_deps):
egroup = exported_name(group)
emitter.emit_lines(
'tmp = PyImport_ImportModule("{}"); if (!tmp) goto fail; Py_DECREF(tmp);'.format(
shared_lib_name(group)),
'struct export_table_{} *pexports_{} = PyCapsule_Import("{}.exports", 0);'.format(
egroup, egroup, shared_lib_name(group)),
f'if (!pexports_{egroup}) {{',
'goto fail;',
'}',
'memcpy(&exports_{group}, pexports_{group}, sizeof(exports_{group}));'.format(
group=egroup),
'',
)
emitter.emit_lines(
'return module;',
'fail:',
'Py_XDECREF(module);',
'return NULL;',
'}',
)
def generate_globals_init(self, emitter: Emitter) -> None:
emitter.emit_lines(
'',
'int CPyGlobalsInit(void)',
'{',
'static int is_initialized = 0;',
'if (is_initialized) return 0;',
''
)
emitter.emit_line('CPy_Init();')
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'
emitter.emit_lines(f'if (CPyStatics_Initialize(CPyStatics, {values}) < 0) {{',
'return -1;',
'}')
emitter.emit_lines(
'is_initialized = 1;',
'return 0;',
'}',
)
def generate_module_def(self, emitter: Emitter, module_name: str, module: ModuleIR) -> None:
"""Emit the PyModuleDef struct for a module and the module init function."""
# Emit module methods
module_prefix = emitter.names.private_name(module_name)
emitter.emit_line(f'static PyMethodDef {module_prefix}module_methods[] = {{')
for fn in module.functions:
if fn.class_name is not None or fn.name == TOP_LEVEL_NAME:
continue
name = short_id_from_name(fn.name, fn.decl.shortname, fn.line)
if is_fastcall_supported(fn, emitter.capi_version):
flag = 'METH_FASTCALL'
else:
flag = 'METH_VARARGS'
emitter.emit_line(
('{{"{name}", (PyCFunction){prefix}{cname}, {flag} | METH_KEYWORDS, '
'NULL /* docstring */}},').format(
name=name,
cname=fn.cname(emitter.names),
prefix=PREFIX,
flag=flag))
emitter.emit_line('{NULL, NULL, 0, NULL}')
emitter.emit_line('};')
emitter.emit_line()
# Emit module definition struct
emitter.emit_lines(f'static struct PyModuleDef {module_prefix}module = {{',
'PyModuleDef_HEAD_INIT,',
f'"{module_name}",',
'NULL, /* docstring */',
'-1, /* size of per-interpreter state of the module,',
' or -1 if the module keeps state in global variables. */',
f'{module_prefix}module_methods',
'};')
emitter.emit_line()
# Emit module init function. If we are compiling just one module, this
# will be the C API init function. If we are compiling 2+ modules, we
# generate a shared library for the modules and shims that call into
# the shared library, and in this case we use an internal module
# initialized function that will be called by the shim.
if not self.use_shared_lib:
declaration = f'PyMODINIT_FUNC PyInit_{module_name}(void)'
else:
declaration = f'PyObject *CPyInit_{exported_name(module_name)}(void)'
emitter.emit_lines(declaration,
'{')
emitter.emit_line('PyObject* modname = NULL;')
# Store the module reference in a static and return it when necessary.
# This is separate from the *global* reference to the module that will
# be populated when it is imported by a compiled module. We want that
# reference to only be populated when the module has been successfully
# imported, whereas this we want to have to stop a circular import.
module_static = self.module_internal_static_name(module_name, emitter)
emitter.emit_lines(f'if ({module_static}) {{',
f'Py_INCREF({module_static});',
f'return {module_static};',
'}')
emitter.emit_lines(f'{module_static} = PyModule_Create(&{module_prefix}module);',
f'if (unlikely({module_static} == NULL))',
' goto fail;')
emitter.emit_line(
'modname = PyObject_GetAttrString((PyObject *){}, "__name__");'.format(
module_static))
module_globals = emitter.static_name('globals', module_name)
emitter.emit_lines(f'{module_globals} = PyModule_GetDict({module_static});',
f'if (unlikely({module_globals} == NULL))',
' goto fail;')
# HACK: Manually instantiate generated classes here
type_structs: List[str] = []
for cl in module.classes:
type_struct = emitter.type_struct_name(cl)
type_structs.append(type_struct)
if cl.is_generated:
emitter.emit_lines(
'{t} = (PyTypeObject *)CPyType_FromTemplate('
'(PyObject *){t}_template, NULL, modname);'
.format(t=type_struct))
emitter.emit_lines(f'if (unlikely(!{type_struct}))',
' goto fail;')
emitter.emit_lines('if (CPyGlobalsInit() < 0)',
' goto fail;')
self.generate_top_level_call(module, emitter)
emitter.emit_lines('Py_DECREF(modname);')
emitter.emit_line(f'return {module_static};')
emitter.emit_lines('fail:',
f'Py_CLEAR({module_static});',
'Py_CLEAR(modname);')
for name, typ in module.final_names:
static_name = emitter.static_name(name, module_name)
emitter.emit_dec_ref(static_name, typ, is_xdec=True)
undef = emitter.c_undefined_value(typ)
emitter.emit_line(f'{static_name} = {undef};')
# the type objects returned from CPyType_FromTemplate are all new references
# so we have to decref them
for t in type_structs:
emitter.emit_line(f'Py_CLEAR({t});')
emitter.emit_line('return NULL;')
emitter.emit_line('}')
def generate_top_level_call(self, module: ModuleIR, emitter: Emitter) -> None:
"""Generate call to function representing module top level."""
# Optimization: we tend to put the top level last, so reverse iterate
for fn in reversed(module.functions):
if fn.name == TOP_LEVEL_NAME:
emitter.emit_lines(
f'char result = {emitter.native_function_name(fn.decl)}();',
'if (result == 2)',
' goto fail;',
)
break
def toposort_declarations(self) -> List[HeaderDeclaration]:
"""Topologically sort the declaration dict by dependencies.
Declarations can require other declarations to come prior in C (such as declaring structs).
In order to guarantee that the C output will compile the declarations will thus need to
be properly ordered. This simple DFS guarantees that we have a proper ordering.
This runs in O(V + E).
"""
result = []
marked_declarations: Dict[str, MarkedDeclaration] = OrderedDict()
for k, v in self.context.declarations.items():
marked_declarations[k] = MarkedDeclaration(v, False)
def _toposort_visit(name: str) -> None:
decl = marked_declarations[name]
if decl.mark:
return
for child in decl.declaration.dependencies:
_toposort_visit(child)
result.append(decl.declaration)
decl.mark = True
for name, marked_declaration in marked_declarations.items():
_toposort_visit(name)
return result
def declare_global(self, type_spaced: str, name: str,
*,
initializer: Optional[str] = None) -> None:
if '[' not in type_spaced:
base = f'{type_spaced}{name}'
else:
a, b = type_spaced.split('[', 1)
base = f'{a}{name}[{b}'
if not initializer:
defn = None
else:
defn = [f'{base} = {initializer};']
if name not in self.context.declarations:
self.context.declarations[name] = HeaderDeclaration(
f'{base};',
defn=defn,
)
def declare_internal_globals(self, module_name: str, emitter: Emitter) -> None:
static_name = emitter.static_name('globals', module_name)
self.declare_global('PyObject *', static_name)
def module_internal_static_name(self, module_name: str, emitter: Emitter) -> str:
return emitter.static_name(module_name + '_internal', None, prefix=MODULE_PREFIX)
def declare_module(self, module_name: str, emitter: Emitter) -> None:
# We declare two globals for each module:
# one used internally in the implementation of module init to cache results
# and prevent infinite recursion in import cycles, and one used
# by other modules to refer to it.
internal_static_name = self.module_internal_static_name(module_name, emitter)
self.declare_global('CPyModule *', internal_static_name, initializer='NULL')
static_name = emitter.static_name(module_name, None, prefix=MODULE_PREFIX)
self.declare_global('CPyModule *', static_name)
self.simple_inits.append((static_name, 'Py_None'))
def declare_imports(self, imps: Iterable[str], emitter: Emitter) -> None:
for imp in imps:
self.declare_module(imp, emitter)
def declare_finals(
self, module: str, final_names: Iterable[Tuple[str, RType]], emitter: Emitter) -> None:
for name, typ in final_names:
static_name = emitter.static_name(name, module)
emitter.context.declarations[static_name] = HeaderDeclaration(
f'{emitter.ctype_spaced(typ)}{static_name};',
[self.final_definition(module, name, typ, emitter)],
needs_export=True)
def final_definition(
self, module: str, name: str, typ: RType, emitter: Emitter) -> str:
static_name = emitter.static_name(name, module)
# Here we rely on the fact that undefined value and error value are always the same
if isinstance(typ, RTuple):
# We need to inline because initializer must be static
undefined = '{{ {} }}'.format(''.join(emitter.tuple_undefined_value_helper(typ)))
else:
undefined = emitter.c_undefined_value(typ)
return f'{emitter.ctype_spaced(typ)}{static_name} = {undefined};'
def declare_static_pyobject(self, identifier: str, emitter: Emitter) -> None:
symbol = emitter.static_name(identifier, None)
self.declare_global('PyObject *', symbol)
def sort_classes(classes: List[Tuple[str, ClassIR]]) -> List[Tuple[str, ClassIR]]:
mod_name = {ir: name for name, ir in classes}
irs = [ir for _, ir in classes]
deps: Dict[ClassIR, Set[ClassIR]] = OrderedDict()
for ir in irs:
if ir not in deps:
deps[ir] = set()
if ir.base:
deps[ir].add(ir.base)
deps[ir].update(ir.traits)
sorted_irs = toposort(deps)
return [(mod_name[ir], ir) for ir in sorted_irs]
T = TypeVar('T')
def toposort(deps: Dict[T, Set[T]]) -> List[T]:
"""Topologically sort a dict from item to dependencies.
This runs in O(V + E).
"""
result = []
visited: Set[T] = set()
def visit(item: T) -> None:
if item in visited:
return
for child in deps[item]:
visit(child)
result.append(item)
visited.add(item)
for item in deps:
visit(item)
return result
def is_fastcall_supported(fn: FuncIR, capi_version: Tuple[int, int]) -> bool:
if fn.class_name is not None:
if fn.name == '__call__':
# We can use vectorcalls (PEP 590) when supported
return use_vectorcall(capi_version)
# TODO: Support fastcall for __init__.
return use_fastcall(capi_version) and fn.name != '__init__'
return use_fastcall(capi_version)
def collect_literals(fn: FuncIR, literals: Literals) -> None:
"""Store all Python literal object refs in fn.
Collecting literals must happen only after we have the final IR.
This way we won't include literals that have been optimized away.
"""
for block in fn.blocks:
for op in block.ops:
if isinstance(op, LoadLiteral):
literals.record_literal(op.value)
def c_array_initializer(components: List[str]) -> str:
"""Construct an initializer for a C array variable.
Components are C expressions valid in an initializer.
For example, if components are ["1", "2"], the result
would be "{1, 2}", which can be used like this:
int a[] = {1, 2};
If the result is long, split it into multiple lines.
"""
res = []
current: List[str] = []
cur_len = 0
for c in components:
if not current or cur_len + 2 + len(c) < 70:
current.append(c)
cur_len += len(c) + 2
else:
res.append(', '.join(current))
current = [c]
cur_len = len(c)
if not res:
# Result fits on a single line
return '{%s}' % ', '.join(current)
# Multi-line result
res.append(', '.join(current))
return '{\n ' + ',\n '.join(res) + '\n}'
def c_string_array_initializer(components: List[bytes]) -> str:
result = []
result.append('{\n')
for s in components:
result.append(' ' + c_string_initializer(s) + ',\n')
result.append('}')
return ''.join(result)