Why Sponsor Oils? | source | all docs for version 0.23.0 | all versions | oilshell.org

mycpp

This is a Python-to-C++ translator based on MyPy. It only handles the small subset of Python that we use in Oils.

It's inspired by both mypyc and Shed Skin. These posts give background:

• Brief Descriptions of a Python to C++ Translator
• Oil Is Being Implemented "Middle Out"

As of March 2024, the translation to C++ is done. So it's no longer experimental!

However, it's still pretty hacky. This doc exists mainly to explain the hacks. (We may want to rewrite mycpp as "yaks", although it's low priority right now.)

Source for this doc: mycpp/README.md. The code is all in mycpp/.

Instructions

Translating and Compiling oils-cpp

Running mycpp is best done on a Debian / Ubuntu-ish machine. Follow the instructions at https://github.com/oilshell/oil/wiki/Contributing to create the "dev build" first, which is DISTINCT from the C++ build. Make sure you can run:

oil$ build/py.sh all

This will give you a working shell:

oil$ bin/osh -c 'echo hi'  # running interpreted Python
hi

To run mycpp, we will build Python 3.10, clone MyPy, and install MyPy's dependencies. First install packages:

# We need libssl-dev, libffi-dev, zlib1g-dev to bootstrap Python
oil$ build/deps.sh install-ubuntu-packages

You'll also need a C++17 compiler for code generated by Souffle datalog, used by mycpp, although Oils itself only requires C++11.

Then fetch data, like the Python 3.10 tarball and MyPy repo:

oil$ build/deps.sh fetch

Then build from source:

oil$ build/deps.sh install-wedges

To build oil-native, use:

oil$ ./NINJA-config.sh
oil$ ninja              # translate and compile, may take 30 seconds

oil$ _bin/cxx-asan/osh -c 'echo hi'  # running compiled C++ !
hi

To run the tests and benchmarks:

oil$ mycpp/TEST.sh test-translator
... 200+ tasks run ...

If you have problems, post a message on #oil-dev at https://oilshell.zulipchat.com. Not many people have contributed to mycpp, so I can use your feedback!

Related:

• Oil Native Quick Start on the wiki.
• Oil Dev Cheat Sheet

Notes on the Algorithm / Architecture

There are four passes over the MyPy AST.

(1) const_pass.py: Collect string constants

Turn turn the constant in myfunc("foo") into top-level GLOBAL_STR(str1, "foo").

(2) Three passes in cppgen_pass.py.

(a) Forward Declaration Pass.

class Foo;
class Bar;

This pass also determines which methods should be declared virtual in their declarations. The virtual keyword is written in the next pass.

(b) Declaration Pass.

class Foo {
  void method();
};
class Bar {
  void method();
};

More work in this pass:

• Collect member variables and write them at the end of the definition
• Collect locals for "hoisting". Written in the next pass.

(c) Definition Pass.

void Foo:method() {
  ...
}

void Bar:method() {
  ...
}

Note: I really wish we were not using visitors, but that's inherited from MyPy.

mycpp Idioms / "Creative Hacks"

Oils is written in typed Python 2. It will run under a stock Python 2 interpreter, and it will typecheck with stock MyPy.

However, there are a few language features that don't map cleanly from typed Python to C++:

• switch statements (unfortunately we don't have the Python 3 match statement)
• C++ destructors - the RAII ptatern
• casting - MyPy has one kind of cast; C++ has static_cast and reinterpret_cast. (We don't use C-style casting.)

So this describes the idioms we use. There are some hacks in mycpp/cppgen_pass.py to handle these cases, and also Python runtime equivalents in mycpp/mylib.py.

with {,tag,str_}switch → Switch statement

We have three constructs that translate to a C++ switch statement. They use a Python context manager with Xswitch(obj) ... as a little hack.

Here are examples like the ones in mycpp/examples/test_switch.py. (ninja mycpp-logs-equal translates, compiles, and tests all the examples.)

Simple switch:

myint = 99
with switch(myint) as case:
    if case(42, 43):
        print('forties')
    else:
        print('other')

Switch on object type, which goes well with ASDL sum types:

val = value.Str('foo)  # type: value_t
with tagswitch(val) as case:
    if case(value_e.Str, value_e.Int):
        print('string or int')
    else:
        print('other')

We usually need to apply the UP_val pattern here, described in the next section.

Switch on string, which generates a fast two-level dispatch -- first on length, and then with str_equals_c():

s = 'foo'
with str_switch(s) as case:
    if case("foo")
        print('FOO')
    else:
        print('other')

val → UP_val → val Downcasting pattern

Summary: variable names like UP_* are special in our Python code.

Consider the downcasts marked BAD:

val = value.Str('foo)  # type: value_t

with tagswitch(obj) as case:
    if case(value_e.Str):
        val = cast(value.Str, val)  # BAD: conflicts with first declaration
        print('s = %s' % val.s)

    elif case(value_e.Int):
        val = cast(value.Int, val)  # BAD: conflicts with both
        print('i = %d' % val.i)

    else:
        print('other')

MyPy allows this, but it translates to invalid C++ code. C++ can't have a variable named val, with 2 related types value_t and value::Str.

So we use this idiom instead, which takes advantage of local vars in case blocks in C++:

val = value.Str('foo')  # type: value_t

UP_val = val  # temporary variable that will be casted

with tagswitch(val) as case:
    if case(value_e.Str):
        val = cast(value.Str, UP_val)  # this works
        print('s = %s' % val.s)

    elif case(value_e.Int):
        val = cast(value.Int, UP_val)  # also works
        print('i = %d' % val.i)

    else:
        print('other')

This translates to something like:

value_t* val = Alloc<value::Str>(str42);
value_t* UP_val = val;

switch (val->tag()) {
    case value_e::Str: {
        // DIFFERENT local var
        value::Str* val = static_cast<value::Str>(UP_val);
        print(StrFormat(str43, val->s))
    }
        break;
    case value_e::Int: {
        // ANOTHER DIFFERENT local var
        value::Int* val = static_cast<value::Int>(UP_val);
        print(StrFormat(str44, val->i))
    }
        break;
    default:
        print(str45);
}

This works because there's no problem having different variables with the same name within each case { } block.

Again, the names UP_* are special. If the name doesn't start with UP_, the inner blocks will look like:

    case value_e::Str: {
        val = static_cast<value::Str>(val);  // BAD: val reused
        print(StrFormat(str43, val->s))
    }

And they will fail to compile. It's not valid C++ because the superclass value_t doesn't have a field val->s. Only the subclass value::Str has it.

(Note that Python has a single flat scope per function, while C++ has nested scopes.)

Python context manager → C++ constructor and destructor (RAII)

This Python code:

with ctx_Foo(42):
  f()

translates to this C++ code:

{
  ctx_Foo tmp(42);
  f()

  // destructor ~ctx_Foo implicitly called
}

MyPy "Shimming" Technique

We have an interesting way of "writing Python and C++ at the same time":

1. First, all Python code must pass the MyPy type checker, and run with a stock Python 2 interpreter.
   • This is the source of truth — the source of our semantics.
2. We translate most .py files to C++, except some files, in particular mycpp/mylib.py and files starting with py like core/{pyos.pyutil}.py.
3. In C++, we can substitute custom implementations with the properties we want, like Dict<K, V> being ordered, BigInt being distinct from C int, BufWriter being efficient, etc.

The MyPy type system is very powerful! It lets us do all this.

NewDict() for ordered dicts

Dicts in Python 2 aren't ordered, but we make them ordered at runtime by using mylib.NewDict(), which returns collections_.OrderedDict.

The static type is still Dict[K, V], but change the "spec" to be an ordered dict.

In C++, Dict<K, V> is implemented as an ordered dict. (Note: we don't implement preserving order on deletion, which seems OK.)

• TODO: iteritems() could go away

StackArray[T]

TODO: describe this when it works.

BigInt

• In Python, it's simply defined a a class with an integer, in mylib/mops.py.
• In C++, it's currently typedef int64_t BigInt, but we want to make it a big integer.

ByteAt(), ByteEquals(), ...

Hand optimization to reduce 1-byte strings. For IFS algorithm, LooksLikeGlob(), GlobUnescape().

File / LineReader / BufWriter

TODO: describe how this works.

Can it be more type safe? I think we can cast File to both LineReader and BufWriter.

Or can we invert the relationship, so File derives from both LineReader and BufWriter?

Fast JSON - avoid intermediate allocations

• pyj8.WriteString() is shimmed so we don't create encoded J8 string objects, only to throw them away and write to mylib.BufWriter. Instead, we append an encoded strings directly to the BufWriter.
• Likewise, we have BufWriter::write_spaces to avoid temporary allocations when writing indents.
  • This could be generalized to BufWriter::write_repeated(' ', 42).
• We may also want BufWriter::write_slice()

Limitations Requiring Source Rewrites

mycpp itself may cause limitations on expressiveness, or the C++ language may be able express what we want.

• C++ doesn't have try / except / else, or finally
  • Use the with ctx_Foo pattern instead.
• if mylist tests if the pointer is non-NULL; use if len(mylist) for non-empty test
• Functions can have at most one keyword / optional argument.
  • We generate two methods: f(x) which calls f(x, y) with the default value of y
  • If there are two or more optional arguments:
    • For classes, you can use the "builder pattern", i.e. add an Init_MyMember() method
    • If the arguments are booleans, translate it to a single bitfield argument
• C++ has nested scope and Python has flat function scope. This can cause name collisions.
  • Could enforce this if it becomes a problem

Also see mycpp/examples/invalid_* for Python code that fails to translate.

WARNING: Assumptions Not Checked

Global Constants Can't Be Mutated

We translate top level constants to statically initialized C data structures (zero startup cost):

gStr = 'foo'   
gList = [1, 2]  # type: List[int]
gDict = {'bar': 42}  # type: Dict[str, int]

Even though List and Dict are mutable in general, you should NOT mutate these global instances! The C++ code will break at runtime.

Gotcha about Returning Variants (Subclasses) of a Type

MyPy will accept this code:

if cond:
  sig = proc_sig.Open  # type: proc_sig_t
                       # bad because mycpp HOISTS this
else:
  sig = proc_sig.Closed.CreateNull()
  sig.words = words    # assignment fails
return sig

It will translate to C++, but fail to compile. Instead, rewrite it like this:

sig = None  # type: proc_sig_t
if cond:
  sig = proc_sig.Open  # type: proc_sig_t
                       # bad because mycpp HOISTS this
else:
  closed = proc_sig.Closed.CreateNull()
  closed.words = words    # assignment fails
  sig = closed
return sig

Exceptions Can't Leave Destructors / Python __exit__

Context managers like with ctx_Foo(): translate to C++ constructors and destructors.

In C++, a destructor can't "leave" an exception. It results in a runtime error.

You can throw and CATCH an exception WITHIN a destructor, but you can't let it propagate outside.

This means you must be careful when coding the __exit__ method. For example, in vm::ctx_Redirect, we had this bug due to IOError being thrown and not caught when restoring/popping redirects.

To fix the bug, we rewrote the code to use an out param List[IOError_OSError].

Related:

• https://akrzemi1.wordpress.com/2011/09/21/destructors-that-throw/

More Translation Notes

Hacky Heuristics

• callable(arg) to either:
  • function call f(arg)
  • instantiation Alloc<T>(arg)
• name.attr to either:
  • obj->member
  • module::Func
• cast(MyType, obj) to either
  • static_cast<MyType*>(obj)
  • reinterpret_cast<MyType*>(obj)

Hacky Hard-Coded Names

These are signs of coupling between mycpp and Oils, which ideally shouldn't exist.

• mycpp_main.py
  • ModulesToCompile() -- some files have to be ordered first, like the ASDL runtime.
    • TODO: Pea can respect parameter order? So we do that outside the project?
    • Another ordering constraint comes from inheritance. The forward declaration is NOT sufficient in that case.
• cppgen_pass.py
  • _GetCastKind() has some hard-coded names
  • AsdlType::Create() is special cased to ::, not ->
  • Default arguments e.g. scope_e::Local need a repeated using.

Issue on mycpp improvements: https://github.com/oilshell/oil/issues/568

Major Features

• Python int and bool → C++ int and bool
  • None → nullptr
• Statically Typed Python Collections
  • str → Str*
  • List[T] → List<T>*
  • Dict[K, V] → Dict<K, V>*
  • tuples → Tuple2<A, B>, Tuple3<A, B, C>, etc.
• Collection literals turn into initializer lists
  • And there is a C++ type inference issue which requires an explicit std::initializer_list<int>{1, 2, 3}, not just {1, 2, 3}
• Python's polymorphic iteration → StrIter, ListIter<T>, DictIter<K, V
  • d.iteritems() is rewritten mylib.iteritems() → DictIter
    • TODO: can we be smarter about this?
  • reversed(mylist) → ReverseListIter
• Python's in operator:
  • s in mystr → str_contains(mystr, s)
  • x in mylist → list_contains(mylist, x)
• Classes and inheritance
  • __init__ method becomes a constructor. Note: initializer lists aren't used.
  • Detect virtual methods
  • TODO: could we detect abstract methods? (NotImplementedError)
• Python generators Iterator[T] → eager List<T> accumulators
• Python Exceptions → C++ exceptions
• Python Modules → C++ namespace (we assume a 2-level hierarchy)
  • TODO: mycpp need real modules, because our oils_for_unix.mycpp.cc translation unit is getting big.
  • And cpp/preamble.h is a hack to work around the lack of modules.

Minor Translations

• s1 == s2 → str_equals(s1, s2)
• 'x' * 3 → str_repeat(globalStr, 3)
• [None] * 3 → list_repeat(nullptr, 3)
• Omitted:
  • If the LHS of an assignment is _, then the statement is omitted
    • This is for _ = log, which shuts up Python lint warnings for 'unused import'
  • Code under if __name__ == '__main__'

Optimizations

• Returning Tuples by value. To reduce GC pressure, we we return Tuple2<A, B> instead of Tuple2<A, B>*, and likewise for Tuple3 and Tuple4.

Rooting Policy

The translated code roots local variables in every function

StackRoots _r({&var1, &var2});

We have two kinds of hand-written code:

1. Methods like Str::strip() in mycpp/
2. OS bindings like stat() in cpp/

Neither of them needs any rooting! This is because we use manual collection points in the interpreter, and these functions don't call any functions that can collect. They are "leaves" in the call tree.

The mycpp Runtime

The mycpp translator targets a runtime that's written from scratch. It implements garbage-collected data structures like:

• Typed records
  • Python classes
  • ASDL product and sum types
• Str (immutable, as in Python)
• List<T>
• Dict<K, V>
• Tuple2<A, B>, Tuple3<A, B, C>, ...

It also has functions based on CPython's:

• mycpp/gc_builtins.{h,cc} corresponds roughly to Python's __builtin__ module, e.g. int() and str()
• mycpp/gc_mylib.{h,cc} corresponds mylib.py
  • mylib.BufWriter is a bit like cStringIO.StringIO

Differences from CPython

• Integers either C int or mylib.BigInt, not Python's arbitrary size integers
• NUL bytes are allowed in arguments to syscalls like open(), unlike in CPython
• s.strip() is defined in terms of ASCII whitespace, which does not include say \v.
  • This is done to be consistent with JSON and J8 Notation.

C++ Notes

Gotchas

• C++ classes can have 2 member variables of the same name! From the base class and derived class.
• Failing to declare methods virtual can involve the wrong one being called at runtime

Minor Features Used

In addition to classes, templates, exceptions, etc. mentioned above, we use:

• static_cast and reinterpret_cast
• enum class for ASDL
• Function overloading
  • For equality and hashing?
• offsetof for introspection of field positions for garbage collection
• std::initializer_list for StackRoots()
  • Should we get rid of this?

Not Used

• I/O Streams, RTTI, etc.
• const
• Smart pointers