django/tests/utils_tests/test_jslex.py

231 lines
9.6 KiB
Python
Raw Normal View History

# -*- coding: utf-8 -*-
"""Tests for jslex."""
# originally from https://bitbucket.org/ned/jslex
from __future__ import unicode_literals
from django.test import SimpleTestCase
from django.utils.jslex import JsLexer, prepare_js_for_gettext
class JsTokensTest(SimpleTestCase):
LEX_CASES = [
# ids
("a ABC $ _ a123", ["id a", "id ABC", "id $", "id _", "id a123"]),
("\\u1234 abc\\u0020 \\u0065_\\u0067", ["id \\u1234", "id abc\\u0020", "id \\u0065_\\u0067"]),
# numbers
("123 1.234 0.123e-3 0 1E+40 1e1 .123", [
"dnum 123", "dnum 1.234", "dnum 0.123e-3", "dnum 0", "dnum 1E+40",
"dnum 1e1", "dnum .123",
]),
("0x1 0xabCD 0XABcd", ["hnum 0x1", "hnum 0xabCD", "hnum 0XABcd"]),
("010 0377 090", ["onum 010", "onum 0377", "dnum 0", "dnum 90"]),
("0xa123ghi", ["hnum 0xa123", "id ghi"]),
# keywords
("function Function FUNCTION", ["keyword function", "id Function", "id FUNCTION"]),
("const constructor in inherits", ["keyword const", "id constructor", "keyword in", "id inherits"]),
("true true_enough", ["reserved true", "id true_enough"]),
# strings
(''' 'hello' "hello" ''', ["string 'hello'", 'string "hello"']),
(r""" 'don\'t' "don\"t" '"' "'" '\'' "\"" """, [
r"""string 'don\'t'""", r'''string "don\"t"''', r"""string '"'""",
r'''string "'"''', r"""string '\''""", r'''string "\""'''
]),
(r'"ƃuıxǝ⅂ ʇdıɹɔsɐʌɐſ\""', [r'string "ƃuıxǝ⅂ ʇdıɹɔsɐʌɐſ\""']),
# comments
("a//b", ["id a", "linecomment //b"]),
("/****/a/=2//hello", ["comment /****/", "id a", "punct /=", "dnum 2", "linecomment //hello"]),
("/*\n * Header\n */\na=1;", ["comment /*\n * Header\n */", "id a", "punct =", "dnum 1", "punct ;"]),
# punctuation
("a+++b", ["id a", "punct ++", "punct +", "id b"]),
# regex
(r"a=/a*/,1", ["id a", "punct =", "regex /a*/", "punct ,", "dnum 1"]),
(r"a=/a*[^/]+/,1", ["id a", "punct =", "regex /a*[^/]+/", "punct ,", "dnum 1"]),
(r"a=/a*\[^/,1", ["id a", "punct =", r"regex /a*\[^/", "punct ,", "dnum 1"]),
(r"a=/\//,1", ["id a", "punct =", r"regex /\//", "punct ,", "dnum 1"]),
# next two are from http://www.mozilla.org/js/language/js20-2002-04/rationale/syntax.html#regular-expressions
("""for (var x = a in foo && "</x>" || mot ? z:/x:3;x<5;y</g/i) {xyz(x++);}""",
["keyword for", "punct (", "keyword var", "id x", "punct =", "id a", "keyword in",
"id foo", "punct &&", 'string "</x>"', "punct ||", "id mot", "punct ?", "id z",
"punct :", "regex /x:3;x<5;y</g", "punct /", "id i", "punct )", "punct {",
"id xyz", "punct (", "id x", "punct ++", "punct )", "punct ;", "punct }"]),
("""for (var x = a in foo && "</x>" || mot ? z/x:3;x<5;y</g/i) {xyz(x++);}""",
["keyword for", "punct (", "keyword var", "id x", "punct =", "id a", "keyword in",
"id foo", "punct &&", 'string "</x>"', "punct ||", "id mot", "punct ?", "id z",
"punct /", "id x", "punct :", "dnum 3", "punct ;", "id x", "punct <", "dnum 5",
"punct ;", "id y", "punct <", "regex /g/i", "punct )", "punct {",
"id xyz", "punct (", "id x", "punct ++", "punct )", "punct ;", "punct }"]),
# Various "illegal" regexes that are valid according to the std.
(r"""/????/, /++++/, /[----]/ """, ["regex /????/", "punct ,", "regex /++++/", "punct ,", "regex /[----]/"]),
# Stress cases from http://stackoverflow.com/questions/5533925/what-javascript-constructs-does-jslex-incorrectly-lex/5573409#5573409 # NOQA
(r"""/\[/""", [r"""regex /\[/"""]),
(r"""/[i]/""", [r"""regex /[i]/"""]),
(r"""/[\]]/""", [r"""regex /[\]]/"""]),
(r"""/a[\]]/""", [r"""regex /a[\]]/"""]),
(r"""/a[\]]b/""", [r"""regex /a[\]]b/"""]),
(r"""/[\]/]/gi""", [r"""regex /[\]/]/gi"""]),
(r"""/\[[^\]]+\]/gi""", [r"""regex /\[[^\]]+\]/gi"""]),
(r"""
rexl.re = {
NAME: /^(?![0-9])(?:\w)+|^"(?:[^"]|"")+"/,
UNQUOTED_LITERAL: /^@(?:(?![0-9])(?:\w|\:)+|^"(?:[^"]|"")+")\[[^\]]+\]/,
QUOTED_LITERAL: /^'(?:[^']|'')*'/,
NUMERIC_LITERAL: /^[0-9]+(?:\.[0-9]*(?:[eE][-+][0-9]+)?)?/,
SYMBOL: /^(?:==|=|<>|<=|<|>=|>|!~~|!~|~~|~|!==|!=|!~=|!~|!|&|\||\.|\:|,|\(|\)|\[|\]|\{|\}|\?|\:|;|@|\^|\/\+|\/|\*|\+|-)/
};
""", # NOQA
["id rexl", "punct .", "id re", "punct =", "punct {",
"id NAME", "punct :", r"""regex /^(?![0-9])(?:\w)+|^"(?:[^"]|"")+"/""", "punct ,",
"id UNQUOTED_LITERAL", "punct :", r"""regex /^@(?:(?![0-9])(?:\w|\:)+|^"(?:[^"]|"")+")\[[^\]]+\]/""",
"punct ,",
"id QUOTED_LITERAL", "punct :", r"""regex /^'(?:[^']|'')*'/""", "punct ,",
"id NUMERIC_LITERAL", "punct :", r"""regex /^[0-9]+(?:\.[0-9]*(?:[eE][-+][0-9]+)?)?/""", "punct ,",
"id SYMBOL", "punct :", r"""regex /^(?:==|=|<>|<=|<|>=|>|!~~|!~|~~|~|!==|!=|!~=|!~|!|&|\||\.|\:|,|\(|\)|\[|\]|\{|\}|\?|\:|;|@|\^|\/\+|\/|\*|\+|-)/""", # NOQA
"punct }", "punct ;"
]),
(r"""
rexl.re = {
NAME: /^(?![0-9])(?:\w)+|^"(?:[^"]|"")+"/,
UNQUOTED_LITERAL: /^@(?:(?![0-9])(?:\w|\:)+|^"(?:[^"]|"")+")\[[^\]]+\]/,
QUOTED_LITERAL: /^'(?:[^']|'')*'/,
NUMERIC_LITERAL: /^[0-9]+(?:\.[0-9]*(?:[eE][-+][0-9]+)?)?/,
SYMBOL: /^(?:==|=|<>|<=|<|>=|>|!~~|!~|~~|~|!==|!=|!~=|!~|!|&|\||\.|\:|,|\(|\)|\[|\]|\{|\}|\?|\:|;|@|\^|\/\+|\/|\*|\+|-)/
};
str = '"';
""", # NOQA
["id rexl", "punct .", "id re", "punct =", "punct {",
"id NAME", "punct :", r"""regex /^(?![0-9])(?:\w)+|^"(?:[^"]|"")+"/""", "punct ,",
"id UNQUOTED_LITERAL", "punct :", r"""regex /^@(?:(?![0-9])(?:\w|\:)+|^"(?:[^"]|"")+")\[[^\]]+\]/""",
"punct ,",
"id QUOTED_LITERAL", "punct :", r"""regex /^'(?:[^']|'')*'/""", "punct ,",
"id NUMERIC_LITERAL", "punct :", r"""regex /^[0-9]+(?:\.[0-9]*(?:[eE][-+][0-9]+)?)?/""", "punct ,",
"id SYMBOL", "punct :", r"""regex /^(?:==|=|<>|<=|<|>=|>|!~~|!~|~~|~|!==|!=|!~=|!~|!|&|\||\.|\:|,|\(|\)|\[|\]|\{|\}|\?|\:|;|@|\^|\/\+|\/|\*|\+|-)/""", # NOQA
"punct }", "punct ;",
"id str", "punct =", """string '"'""", "punct ;",
]),
(r""" this._js = "e.str(\"" + this.value.replace(/\\/g, "\\\\").replace(/"/g, "\\\"") + "\")"; """,
["keyword this", "punct .", "id _js", "punct =", r'''string "e.str(\""''', "punct +", "keyword this",
"punct .", "id value", "punct .", "id replace", "punct (", r"regex /\\/g", "punct ,", r'string "\\\\"',
"punct )",
"punct .", "id replace", "punct (", r'regex /"/g', "punct ,", r'string "\\\""', "punct )", "punct +",
r'string "\")"', "punct ;"]),
]
def make_function(input, toks):
def test_func(self):
lexer = JsLexer()
result = ["%s %s" % (name, tok) for name, tok in lexer.lex(input) if name != 'ws']
self.assertListEqual(result, toks)
return test_func
2016-11-13 01:11:23 +08:00
for i, (input, toks) in enumerate(JsTokensTest.LEX_CASES):
setattr(JsTokensTest, "test_case_%d" % i, make_function(input, toks))
GETTEXT_CASES = (
(
r"""
a = 1; /* /[0-9]+/ */
b = 0x2a0b / 1; // /[0-9]+/
c = 3;
""",
r"""
a = 1; /* /[0-9]+/ */
b = 0x2a0b / 1; // /[0-9]+/
c = 3;
"""
), (
r"""
a = 1.234e-5;
/*
* /[0-9+/
*/
b = .0123;
""",
r"""
a = 1.234e-5;
/*
* /[0-9+/
*/
b = .0123;
"""
), (
r"""
x = y / z;
alert(gettext("hello"));
x /= 3;
""",
r"""
x = y / z;
alert(gettext("hello"));
x /= 3;
"""
), (
r"""
s = "Hello \"th/foo/ere\"";
s = 'He\x23llo \'th/foo/ere\'';
s = 'slash quote \", just quote "';
""",
r"""
s = "Hello \"th/foo/ere\"";
s = "He\x23llo \'th/foo/ere\'";
s = "slash quote \", just quote \"";
"""
), (
r"""
s = "Line continuation\
continued /hello/ still the string";/hello/;
""",
r"""
s = "Line continuation\
continued /hello/ still the string";"REGEX";
"""
), (
r"""
var regex = /pattern/;
var regex2 = /matter/gm;
var regex3 = /[*/]+/gm.foo("hey");
""",
r"""
var regex = "REGEX";
var regex2 = "REGEX";
var regex3 = "REGEX".foo("hey");
"""
), (
r"""
for (var x = a in foo && "</x>" || mot ? z:/x:3;x<5;y</g/i) {xyz(x++);}
for (var x = a in foo && "</x>" || mot ? z/x:3;x<5;y</g/i) {xyz(x++);}
""",
r"""
for (var x = a in foo && "</x>" || mot ? z:"REGEX"/i) {xyz(x++);}
for (var x = a in foo && "</x>" || mot ? z/x:3;x<5;y<"REGEX") {xyz(x++);}
"""
), (
"""
\\u1234xyz = gettext('Hello there');
""", r"""
Uu1234xyz = gettext("Hello there");
"""
)
)
class JsToCForGettextTest(SimpleTestCase):
pass
def make_function(js, c):
def test_func(self):
self.assertMultiLineEqual(prepare_js_for_gettext(js), c)
return test_func
2016-11-13 01:11:23 +08:00
for i, pair in enumerate(GETTEXT_CASES):
setattr(JsToCForGettextTest, "test_case_%d" % i, make_function(*pair))