The Compiler

The compilation process in MicroPython involves the following steps:

  • The lexer converts the stream of text that makes up a MicroPython program into tokens.

  • The parser then converts the tokens into an abstract syntax (parse tree).

  • Then bytecode or native code is emitted based on the parse tree.

For purposes of this discussion we are going to add a simple language feature add1 that can be use in Python as:

>>> add1 3
4
>>>

The add1 statement takes an integer as argument and adds 1 to it.

Adding a grammar rule

MicroPython’s grammar is based on the CPython grammar and is defined in py/grammar.h. This grammar is what is used to parse MicroPython source files.

There are two macros you need to know to define a grammar rule: DEF_RULE and DEF_RULE_NC. DEF_RULE allows you to define a rule with an associated compile function, while DEF_RULE_NC has no compile (NC) function for it.

A simple grammar definition with a compile function for our new add1 statement looks like the following:

DEF_RULE(add1_stmt, c(add1_stmt), and(2), tok(KW_ADD1), rule(testlist))

The second argument c(add1_stmt) is the corresponding compile function that should be implemented in py/compile.c to turn this rule into executable code.

The third required argument can be or or and. This specifies the number of nodes associated with a statement. For example, in this case, our add1 statement is similar to ADD1 in assembly language. It takes one numeric argument. Therefore, the add1_stmt has two nodes associated with it. One node is for the statement itself, i.e the literal add1 corresponding to KW_ADD1, and the other for its argument, a testlist rule which is the top-level expression rule.

Note

The add1 rule here is just an example and not part of the standard MicroPython grammar.

The fourth argument in this example is the token associated with the rule, KW_ADD1. This token should be defined in the lexer by editing py/lexer.h.

Defining the same rule without a compile function is achieved by using the DEF_RULE_NC macro and omitting the compile function argument:

DEF_RULE_NC(add1_stmt, and(2), tok(KW_ADD1), rule(testlist))

The remaining arguments take on the same meaning. A rule without a compile function must be handled explicitly by all rules that may have this rule as a node. Such NC-rules are usually used to express sub-parts of a complicated grammar structure that cannot be expressed in a single rule.

Note

The macros DEF_RULE and DEF_RULE_NC take other arguments. For an in-depth understanding of supported parameters, see py/grammar.h.

Adding a lexical token

Every rule defined in the grammar should have a token associated with it that is defined in py/lexer.h. Add this token by editing the _mp_token_kind_t enum:

typedef enum _mp_token_kind_t {
    ...
    MP_TOKEN_KW_OR,
    MP_TOKEN_KW_PASS,
    MP_TOKEN_KW_RAISE,
    MP_TOKEN_KW_RETURN,
    MP_TOKEN_KW_TRY,
    MP_TOKEN_KW_WHILE,
    MP_TOKEN_KW_WITH,
    MP_TOKEN_KW_YIELD,
    MP_TOKEN_KW_ADD1,
    ...
} mp_token_kind_t;

Then also edit py/lexer.c to add the new keyword literal text:

STATIC const char *const tok_kw[] = {
    ...
    "or",
    "pass",
    "raise",
    "return",
    "try",
    "while",
    "with",
    "yield",
    "add1",
    ...
};

Notice the keyword is named depending on what you want it to be. For consistency, maintain the naming standard accordingly.

Note

The order of these keywords in py/lexer.c must match the order of tokens in the enum defined in py/lexer.h.

Parsing

In the parsing stage the parser takes the tokens produced by the lexer and converts them to an abstract syntax tree (AST) or parse tree. The implementation for the parser is defined in py/parse.c.

The parser also maintains a table of constants for use in different aspects of parsing, similar to what a symbol table does.

Several optimizations like constant folding on integers for most operations e.g. logical, binary, unary, etc, and optimizing enhancements on parenthesis around expressions are performed during this phase, along with some optimizations on strings.

It’s worth noting that docstrings are discarded and not accessible to the compiler. Even optimizations like string interning are not applied to docstrings.

Compiler passes

Like many compilers, MicroPython compiles all code to MicroPython bytecode or native code. The functionality that achieves this is implemented in py/compile.c. The most relevant method you should know about is this:

mp_obj_t mp_compile(mp_parse_tree_t *parse_tree, qstr source_file, bool is_repl) {
    // Compile the input parse_tree to a raw-code structure.
    mp_raw_code_t *rc = mp_compile_to_raw_code(parse_tree, source_file, is_repl);
    // Create and return a function object that executes the outer module.
    return mp_make_function_from_raw_code(rc, MP_OBJ_NULL, MP_OBJ_NULL);
}

The compiler compiles the code in four passes: scope, stack size, code size and emit. Each pass runs the same C code over the same AST data structure, with different things being computed each time based on the results of the previous pass.

First pass

In the first pass, the compiler learns about the known identifiers (variables) and their scope, being global, local, closed over, etc. In the same pass the emitter (bytecode or native code) also computes the number of labels needed for the emitted code.

// Compile pass 1.
comp->emit = emit_bc;
comp->emit_method_table = &emit_bc_method_table;

uint max_num_labels = 0;
for (scope_t *s = comp->scope_head; s != NULL && comp->compile_error == MP_OBJ_NULL; s = s->next) {
    if (s->emit_options == MP_EMIT_OPT_ASM) {
        compile_scope_inline_asm(comp, s, MP_PASS_SCOPE);
    } else {
        compile_scope(comp, s, MP_PASS_SCOPE);

        // Check if any implicitly declared variables should be closed over.
        for (size_t i = 0; i < s->id_info_len; ++i) {
            id_info_t *id = &s->id_info[i];
            if (id->kind == ID_INFO_KIND_GLOBAL_IMPLICIT) {
                scope_check_to_close_over(s, id);
            }
        }
    }
    ...
}

Second and third passes

The second and third passes involve computing the Python stack size and code size for the bytecode or native code. After the third pass the code size cannot change, otherwise jump labels will be incorrect.

for (scope_t *s = comp->scope_head; s != NULL && comp->compile_error == MP_OBJ_NULL; s = s->next) {
    ...

    // Pass 2: Compute the Python stack size.
    compile_scope(comp, s, MP_PASS_STACK_SIZE);

    // Pass 3: Compute the code size.
    if (comp->compile_error == MP_OBJ_NULL) {
        compile_scope(comp, s, MP_PASS_CODE_SIZE);
    }

    ...
}

Just before pass two there is a selection for the type of code to be emitted, which can either be native or bytecode.

// Choose the emitter type.
switch (s->emit_options) {
    case MP_EMIT_OPT_NATIVE_PYTHON:
    case MP_EMIT_OPT_VIPER:
        if (emit_native == NULL) {
            emit_native = NATIVE_EMITTER(new)(&comp->compile_error, &comp->next_label, max_num_labels);
        }
        comp->emit_method_table = NATIVE_EMITTER_TABLE;
        comp->emit = emit_native;
        break;

    default:
        comp->emit = emit_bc;
        comp->emit_method_table = &emit_bc_method_table;
        break;
}

The bytecode option is the default but something unique to note for the native code option is that there is another option via VIPER. See the Emitting native code section for more details on viper annotations.

There is also support for inline assembly code, where assembly instructions are written as Python function calls but are emitted directly as the corresponding machine code. This assembler has only three passes (scope, code size, emit) and uses a different implementation, not the compile_scope function. See the inline assembler tutorial for more details.

Fourth pass

The fourth pass emits the final code that can be executed, either bytecode in the virtual machine, or native code directly by the CPU.

for (scope_t *s = comp->scope_head; s != NULL && comp->compile_error == MP_OBJ_NULL; s = s->next) {
    ...

    // Pass 4: Emit the compiled bytecode or native code.
    if (comp->compile_error == MP_OBJ_NULL) {
        compile_scope(comp, s, MP_PASS_EMIT);
    }
}

Emitting bytecode

Statements in Python code usually correspond to emitted bytecode, for example a + b generates “push a” then “push b” then “binary op add”. Some statements do not emit anything but instead affect other things like the scope of variables, for example global a.

The implementation of a function that emits bytecode looks similar to this:

void mp_emit_bc_unary_op(emit_t *emit, mp_unary_op_t op) {
    emit_write_bytecode_byte(emit, 0, MP_BC_UNARY_OP_MULTI + op);
}

We use the unary operator expressions for an example here but the implementation details are similar for other statements/expressions. The method emit_write_bytecode_byte() is a wrapper around the main function emit_get_cur_to_write_bytecode() that all functions must call to emit bytecode.

Emitting native code

Similar to how bytecode is generated, there should be a corresponding function in py/emitnative.c for each code statement:

STATIC void emit_native_unary_op(emit_t *emit, mp_unary_op_t op) {
     vtype_kind_t vtype;
     emit_pre_pop_reg(emit, &vtype, REG_ARG_2);
     if (vtype == VTYPE_PYOBJ) {
         emit_call_with_imm_arg(emit, MP_F_UNARY_OP, op, REG_ARG_1);
         emit_post_push_reg(emit, VTYPE_PYOBJ, REG_RET);
     } else {
         adjust_stack(emit, 1);
         EMIT_NATIVE_VIPER_TYPE_ERROR(emit,
             MP_ERROR_TEXT("unary op %q not implemented"), mp_unary_op_method_name[op]);
     }
}

The difference here is that we have to handle viper typing. Viper annotations allow us to handle more than one type of variable. By default all variables are Python objects, but with viper a variable can also be declared as a machine-typed variable like a native integer or pointer. Viper can be thought of as a superset of Python, where normal Python objects are handled as usual, while native machine variables are handled in an optimised way by using direct machine instructions for the operations. Viper typing may break Python equivalence because, for example, integers become native integers and can overflow (unlike Python integers which extend automatically to arbitrary precision).