Internal Hy Documentation¶
These bits are mostly useful for folks who hack on Hy itself, but can also be used for those delving deeper in macro programming.
Introduction to Hy Models¶
Hy models are a very thin layer on top of regular Python objects, representing Hy source code as data. Models only add source position information, and a handful of methods to support clean manipulation of Hy source code, for instance in macros. To achieve that goal, Hy models are mixins of a base Python class and Object Protocol.
hy.models.Object is the base class of Hy models. It only
implements one method,
replace, which replaces the source position
of the current object with the one passed as argument. This allows us to
keep track of the original position of expressions that get modified by
macros, be that in the compiler or in pure hy macros.
Object is not intended to be used directly to instantiate Hy
models, but only as a mixin for other classes.
Parenthesized and bracketed lists are parsed as compound models by the Hy parser.
Hy uses pretty-printing reprs for its compound models by default.
If this is causing issues,
it can be turned off globally by setting
or temporarily by using the
hy.models.pretty context manager.
Hy also attempts to color pretty reprs and errors using
colorama. These can
be turned off globally by setting
hy.models.Sequence is the abstract base class of “iterable” Hy
models, such as hy.models.Expression and hy.models.List.
Adding a Sequence to another iterable object reuses the class of the left-hand-side object, a useful behavior when you want to concatenate Hy objects in a macro, for instance.
Sequences are (mostly) immutable: you can’t add, modify, or remove
elements. You can still append to a variable containing a Sequence with
+= and otherwise construct new Sequences out of old ones.
hy.models.List is a Sequence Protocol for bracketed
lists, which, when used as a top-level expression, translate to Python
list literals in the compilation phase.
hy.models.Expression inherits Sequence Protocol for
() expressions. The compilation result of those
expressions depends on the first element of the list: the compiler
dispatches expressions between compiler special-forms, user-defined
macros, and regular Python function calls.
In the input stream, double-quoted strings, respecting the Python notation for strings, are parsed as a single token, which is directly parsed as a String.
An uninterrupted string of characters, excluding spaces, brackets, quotes, double-quotes and comments, is parsed as an identifier.
Identifiers are resolved to atomic models during the parsing phase in the following order:
hy.models.String represents string literals (including bracket strings),
which compile down to unicode string literals (
str) in Python.
Strings are immutable.
Hy literal strings can span multiple lines, and are considered by the parser as a single unit, respecting the Python escapes for unicode strings.
Strings have an attribute
brackets that stores the custom
delimiter used for a bracket string (e.g.,
world]==] and the empty string for
Strings that are not produced by bracket strings have their
brackets set to
hy.models.Bytes is like
String, but for sequences of bytes.
It inherits from
hy.models.Integer represents integer literals, using the
hy.models.Float represents floating-point literals.
hy.models.Complex represents complex literals.
Numeric models are parsed using the corresponding Python routine, and valid numeric python literals will be turned into their Hy counterpart.
hy.models.Symbol is the model used to represent symbols in the Hy
String, it inherits from
unicode on Python
Symbols are mangled when they are compiled to Python variable names.
hy.models.Keyword represents keywords in Hy. Keywords are
symbols starting with a
:. See keywords.
.name attribute of a
the (unmangled) string representation of the keyword without the initial
=> (setv x :foo-bar) => (print x.name) foo-bar
If needed, you can get the mangled name by calling
Hy Internal Theory¶
The Hy internals work by acting as a front-end to Python bytecode, so that Hy itself compiles down to Python Bytecode, allowing an unmodified Python runtime to run Hy code, without even noticing it.
The way we do this is by translating Hy into an internal Python AST datastructure, and building that AST down into Python bytecode using modules from the Python standard library, so that we don’t have to duplicate all the work of the Python internals for every single Python release.
Hy works in four stages. The following sections will cover each step of Hy from source to runtime.
Steps 1 and 2: Tokenizing and Parsing¶
The first stage of compiling Hy is to lex the source into tokens that we can deal with. We use a project called rply, which is a really nice (and fast) parser, written in a subset of Python called rpython.
The lexing code is all defined in
hy.lex.lexer. This code is mostly just
defining the Hy grammar, and all the actual hard parts are taken care of by
rply – we just define “callbacks” for rply in
hy.lex.parser, which takes
the tokens generated, and returns the Hy models.
You can think of the Hy models as the “AST” for Hy, it’s what Macros operate on (directly), and it’s what the compiler uses when it compiles Hy down.
Section Hy Models for more information on Hy models and what they mean.
Step 3: Hy Compilation to Python AST¶
This is where most of the magic in Hy happens. This is where we take Hy AST (the models), and compile them into Python AST. A couple of funky things happen here to work past a few problems in AST, and working in the compiler is some of the most important work we do have.
The compiler is a bit complex, so don’t feel bad if you don’t grok it on the first shot, it may take a bit of time to get right.
The main entry-point to the Compiler is
HyASTCompiler.compile. This method
is invoked, and the only real “public” method on the class (that is to say,
we don’t really promise the API beyond that method).
In fact, even internally, we don’t recurse directly hardly ever, we almost
always force the Hy tree through
compile, and will often do this with
sub-elements of an expression that we have. It’s up to the Type-based dispatcher
to properly dispatch sub-elements.
All methods that preform a compilation are marked with the
decorator. You can either pass the class of the Hy model that it compiles,
or you can use a string for expressions. I’ll clear this up in a second.
First Stage Type-Dispatch¶
Let’s start in the
compile method. The first thing we do is check the
Type of the thing we’re building. We look up to see if we have a method that
can build the
type() that we have, and dispatch to the method that can
handle it. If we don’t have any methods that can build that type, we raise
For instance, if we have a
String, we have an almost 1-to-1 mapping of
Hy AST to Python AST. The
compile_string method takes the
ast.Str() that’s populated with the correct line-numbers and
If we get a
Expression, we’ll attempt to see if this is a known
Macro, and push to have it expanded by invoking
push the result back into
Second Stage Expression-Dispatch¶
The only special case is the
Expression, since we need to create different
AST depending on the special form in question. For instance, when we hit an
(if True True False), we need to generate a
ast.If, and properly
compile the sub-nodes. This is where the
@builds() with a String as an
argument comes in.
compile_expression (which is defined with an
@builds(Expression)) will dispatch based on the string of the first
argument. If, for some reason, the first argument is not a string, it will
properly handle that case as well (most likely by raising an
If the String isn’t known to Hy, it will default to create an
which will try to do a runtime call (in Python, something like
Issues Hit with Python AST¶
Python AST is great; it’s what’s enabled us to write such a powerful project on top of Python without having to fight Python too hard. Like anything, we’ve had our fair share of issues, and here’s a short list of the common ones you might run into.
Python differentiates between Statements and Expressions.
This might not sound like a big deal – in fact, to most Python programmers, this will shortly become a “Well, yeah” moment.
In Python, doing something like:
print for x in range(10): pass, because
for isn’t an expression, it’s a control flow statement. Things like
1 + 1 are Expressions, as is
lambda x: 1 + x, but other language
features, such as
while are statements.
Since they have no “value” to Python, this makes working in Hy hard, since
doing something like
(print (if True True False)) is not just common, it’s
As a result, we reconfigure things using a
Result object, where we offer
ast.stmt that need to get run, and a single
ast.expr that can
be used to get the value of whatever was just run. Hy does this by forcing
assignment to things while running.
As example, the Hy:
(print (if True True False))
Will turn into:
if True: _temp_name_here = True else: _temp_name_here = False print(_temp_name_here)
OK, that was a bit of a lie, since we actually turn that statement into:
print(True if True else False)
By forcing things into an
ast.expr if we can, but the general idea holds.
Step 4: Python Bytecode Output and Runtime¶
After we have a Python AST tree that’s complete, we can try and compile it to
Python bytecode by pushing it through
eval. From here on out, we’re no
longer in control, and Python is taking care of everything. This is why things
like Python tracebacks, pdb and django apps work.
Using gensym for Safer Macros¶
When writing macros, one must be careful to avoid capturing external variables or using variable names that might conflict with user code.
We will use an example macro
nif (see http://letoverlambda.com/index.cl/guest/chap3.html#sec_5
for a more complete description.)
nif is an example, something like a numeric
where based on the expression, one of the 3 forms is called depending on if the
expression is positive, zero or negative.
A first pass might be something like:
(defmacro nif [expr pos-form zero-form neg-form] `(do (setv obscure-name ~expr) (cond [(pos? obscure-name) ~pos-form] [(zero? obscure-name) ~zero-form] [(neg? obscure-name) ~neg-form])))
obscure-name is an attempt to pick some variable name as not to
conflict with other code. But of course, while well-intentioned,
this is no guarantee.
gensym is designed to generate a new, unique symbol for just
such an occasion. A much better version of
nif would be:
(defmacro nif [expr pos-form zero-form neg-form] (setv g (gensym)) `(do (setv ~g ~expr) (cond [(pos? ~g) ~pos-form] [(zero? ~g) ~zero-form] [(neg? ~g) ~neg-form])))
This is an easy case, since there is only one symbol. But if there is
a need for several gensym’s there is a second macro
basically expands to a
(with-gensyms [a b c] ...)
(do (setv a (gensym) b (gensym) c (gensym)) ...)
so our re-written
nif would look like:
(defmacro nif [expr pos-form zero-form neg-form] (with-gensyms [g] `(do (setv ~g ~expr) (cond [(pos? ~g) ~pos-form] [(zero? ~g) ~zero-form] [(neg? ~g) ~neg-form]))))
Finally, though we can make a new macro that does all this for us.
will take all symbols that begin with
g! and automatically call
gensym with the
remainder of the symbol. So
g!a would become
Our final version of
nif, built with
(defmacro/g! nif [expr pos-form zero-form neg-form] `(do (setv ~g!res ~expr) (cond [(pos? ~g!res) ~pos-form] [(zero? ~g!res) ~zero-form] [(neg? ~g!res) ~neg-form])))