Added index of error to LexError

This allows for easier writing of tools empowered by the `sly` library. Now you can extend the library to get a better context of the lexed data.
This commit is contained in:
Daniel Roy Greenfeld 2017-02-17 09:35:21 -08:00 committed by GitHub
parent b6357d2491
commit 732d0819d9

View File

@ -41,11 +41,13 @@ class LexError(Exception):
'''
Exception raised if an invalid character is encountered and no default
error handler function is defined. The .text attribute of the exception
contains all remaining untokenized text.
contains all remaining untokenized text. The .error_index is the index
location of the error.
'''
def __init__(self, message, text):
def __init__(self, message, text, error_index):
self.args = (message,)
self.text = text
self.error_index = error_index
class PatternError(Exception):
'''
@ -79,7 +81,7 @@ class LexerMetaDict(OrderedDict):
value.pattern = self[key]
else:
raise AttributeError(f'Name {key} redefined')
super().__setitem__(key, value)
class LexerMeta(type):
@ -160,7 +162,7 @@ class Lexer(metaclass=LexerMeta):
pattern = value.pattern
cls._token_funcs[tokname] = value
# Form the regular expression component
# Form the regular expression component
part = f'(?P<{tokname}>{pattern})'
# Make sure the individual regex compiles properly
@ -185,7 +187,7 @@ class Lexer(metaclass=LexerMeta):
# Verify that that ignore and literals specifiers match the input type
if not isinstance(cls.ignore, str):
raise LexerBuildError('ignore specifier must be a string')
if not all(isinstance(lit, str) for lit in cls.literals):
raise LexerBuildError('literals must be specified as strings')
@ -252,4 +254,4 @@ class Lexer(metaclass=LexerMeta):
# Default implementations of the error handler. May be changed in subclasses
def error(self, value):
raise LexError(f'Illegal character {value[0]!r} at index {self.index}', value)
raise LexError(f'Illegal character {value[0]!r} at index {self.index}', value, self.index)