ab1eb6cf58be6ace5f61c02410c1a855fdd20edc — nc 1 year, 8 months ago 561631f
begin work on codegen
2 files changed, 116 insertions(+), 14 deletions(-)

M compiler.py
M test_compiler.py
M compiler.py => compiler.py +96 -13
@@ 43,8 43,8 @@ class NilToken(str):
# EMPTY is not ''
# EMPTY == ''
EMPTY = NilToken('empty')
BLOCK_BEGIN = NilToken('block begin')
BLOCK_END = NilToken('block end')
INDENT = NilToken('indent')
DEDENT = NilToken('dedent')

### Parser combinator

@@ 351,7 351,7 @@ class Block(list):
        return self

# took inspiration from namedtuple:
# took inspiration from collections.namedtuple:
# https://github.com/python/cpython/blob/58ccd201fa74287ca9293c03136fcf1e19800ef9/Lib/collections/__init__.py#L290
def nodeclass(name, fields, hole_values=[]):
    fields = fields.replace(',', ' ').split() if type(fields) == str else fields

@@ 401,10 401,10 @@ def indentation():
            return EMPTY, new_stream
        elif len(c) > stream.indent:
            stream.indent = len(c)
            return BLOCK_BEGIN, new_stream
            return INDENT, new_stream
        elif len(c) < stream.indent:
            stream.indent = len(c)
            return BLOCK_END, new_stream
            return DEDENT, new_stream
    return indentationf

# we don't care about space, so we discard it

@@ 441,13 441,13 @@ block = convert(seq(discard(newline), one_or_more(stmt)), lambda x: x[0])
# <function> := 'def' <space> <identifier> '(' (<identifier> (',' <space> <identifier>)*) ')' ':' <newline> <block>
function = seq('def', space, identifier, char('('), intersperse(identifier, discard(seq(char(','), space))), char(')'), char(':'), block)

### A-normal form
### A-normal form normalizer

gensym_counter = 0
def gensym():
def gensym(prefix='tmp'):
    global gensym_counter
    gensym_counter += 1
    return f'tmp{gensym_counter}'
    return f'{prefix}{gensym_counter}'

def is_trivial(b):
    return type(b) in {str, int, float}

@@ 513,13 513,96 @@ def _(node: FunctionCall):
        new_args.append(maybe_hoist(a, hoisted))
    return FunctionCall(node.name, *new_args), hoisted

### Code gen
### Code gen (x86-64)

#call_registers = ['rdi', 'rsi', 'rdx', 'rcx', 'r8', 'r9']
#def emit_function_call(fcall):
#    assert fcall[0] == 'call'
#    _, function_name, args = fcall
# TODO: register allocation -- go through the tree and tag bindings with registers or memory locations

#import io
#code_gen_buffer = io.BytesIO()

class Register:
    name: str
    num: int

# FIXME: I don't like this class
#   I don't like that it holds both architectural details, as well as a buffer that is being written to...
#   TODO: remove this whole abstraction and just implement the assembly instructions inline, and append to an io buffer
class CodeGenBuffer:
    def __init__(self):
        self.buffer = []
        _reg_enc = 'rax rcx rdx rbx rsp rbp rsi rdi r8 r9 r10 r11 r12 r13 r14 r15'
        for i, r in enumerate(_reg_enc.split()):
            self.__dict__[r] = Register(r, i)
        self.call_registers = [self.rdi, self.rsi, self.rdx, self.rcx, self.r8, self.r9]

    def push(self, r):
        self.buffer.append(['push', r.name])

    def mov(self, a, b):
        self.buffer.append(['mov', a, b])

    def call(self, f):
        self.buffer.append(['call', f])

    def assign(self, a, b):
        self.buffer.append(['assign', a, b])

    def branch_if_false(self, label):
        self.buffer.append(['cmp', 'rax'])
        self.buffer.append(['bne', label])

    def branch(self, label):
        self.buffer.append(['ba', label])

    def label(self, label):

def code_gen(node, ctx, g: CodeGenBuffer):
    assert False, f'unhandled: {node} type {type(node)}'

def _(i: int, ctx, g: CodeGenBuffer):
    g.mov(i, 'rax')

def _(ifstmt: If, ctx, g: CodeGenBuffer):
    false_label = gensym('false')
    end_label = gensym('end')
    g.mov(ifstmt.cond, g.rax)

    # true branch:
    code_gen(ifstmt.then, ctx, g)

    # false branch:
    code_gen(ifstmt.otherwise, ctx, g)


def _(block: Block, ctx, g: CodeGenBuffer):
    for stmt in block: code_gen(stmt, ctx, g)

def _(assign: Assign, ctx, g: CodeGenBuffer):
    g.assign(assign.lhs, assign.rhs)

def _(f: FunctionCall, ctx, g: CodeGenBuffer):
    # TODO: remove this limit and use stack
    assert len(f.args) < len(g.call_registers), 'too many arguments'

    # spill used registers
    for r in g.call_registers[:len(f.args)]: g.push(r)

    for r, a in zip(g.call_registers, f.args):
        g.mov(a, r)


if __name__ == "__main__":
    import doctest

M test_compiler.py => test_compiler.py +20 -1
@@ 107,6 107,13 @@ else:
        self.assertEqual(d[4], 5)
        self.assertEqual(d[5], 6)

        E = nodeclass('E', 'a *b')
        e = E(1,2,3)
        self.assertEqual(e.a, 1)
        self.assertEqual(e.b, (2,3))

    # FIXME: make these tests runable again -- have to fix the problem of resetting the gensym state
    def test_normalize(self):
        tree1 = FunctionCall('a', 'b', 'c')
        self.assertEqual(normalize_expr(tree1)[0], ('a', 'b', 'c'))

@@ 140,7 147,19 @@ else:
                [('=', 'tmp8', ('b',)),
                 ('=', 'x', ('a', 'tmp8'))])


    def test_codegen(self):
        t = If(FunctionCall(FunctionCall('b', ()), 2), Return(1), Return(2))
        g = CodeGenBuffer()
        code_gen(normalize_stmt(t), None, g)
        import pprint

import doctest
def load_tests(loader, tests, ignore):
    return tests

if __name__ == '__main__':