1
Fork 0

add notes to lex.lhs

This commit is contained in:
Jake Wheat 2016-02-13 11:40:49 +02:00
parent 91e7012c47
commit f17053c7d9

View file

@ -4,9 +4,37 @@ The parser uses a separate lexer for two reasons:
1. sql syntax is very awkward to parse, the separate lexer makes it
easier to handle this in most places (in some places it makes it
harder or impossible, the fix is to switch to something better than
parsec
parsec)
2. using a separate lexer gives a huge speed boost because it reduces
backtracking. (We could get this by making the parsing code a lot more
complex also.)
= Lexing and dialects
The main dialect differences:
symbols follow different rules in different dialects
e.g. postgresql has a flexible extensible-ready syntax for operators
which are parsed here as symbols
sql server using [] for quoting identifiers, and so they don't parse
as symbols here (in other dialects including ansi, these are used for
array operations)
quoting of identifiers is different in different dialects
there are various other identifier differences:
ansi has :host_param
there are variants on these like in @sql_server adn in #oracle
string quoting follows different rules in different dialects,
e.g. postgresql has $$ quoting
todo: public documentation on dialect definition - and dialect flags
2. using a separate lexer gives a huge speed boost
> -- | This is the module contains a Lexer for SQL.
> {-# LANGUAGE TupleSections #-}
@ -355,3 +383,184 @@ Some helper combinators
> peekSatisfy :: (Char -> Bool) -> Parser ()
> peekSatisfy p = do
> void $ lookAhead (satisfy p)
postgresql notes:
u&
SELECT 'foo'
'bar';
is equivalent to:
SELECT 'foobar';
SELECT 'foo' 'bar';
is invalid
(this should be in ansi also)
definitely do major review and docs:
when can escapes and prefixes be using with syntactic string literals
when can they be combined
when can e.g. dollar quoting be used
what escaping should there be, including unicode escapes
E'string'
with a range of escapes which should appear in the dialect data type
dollar quoted strings
never with prefixes/escapes
B''
X''
numbers
:: cast
type 'string' - literals only, not array types
ansi allows for some specific types
'string'::type
cast('string' as type)
can use dollar quoting here
typename('string') (not all types)
check these in the parser for keyword issues
extended operator rules
$1 positional parameter
()
[]
,
;
: array slices and variable names/hostparam
*
.
some operator precedence notes
SELECT 3 OPERATOR(pg_catalog.+) 4;
diff from ansi:
all the same symbols + more + different rules about parsing multi char
symbols (ansi is trivial here, postgresql is not trivial and
extensible)
identifiers: same, doublecheck the u&
hostparam: same, but with implementation issues because : is also a
symbol in postgresql. this might be a little tricky to deal with
string literals:
u&?, does pg support n?
numbers: same
whitespace, comments: same
make sure there is a list of lexical syntax which is valid in postgres
and not in ansi, and vice versa, and have explicit tests for
these. There might also be situations here where a string is valid in
both, but lexes differently. There is definitely cases like this in
the main syntax.
action plan:
no abstract syntax changes are needed
write down a spec for ansi and for postgresql lexical syntax
create a list of tests for postgresql
include eveything from ansi which is the same: maybe refactor the
tests to make this maintainable
design for escaping issues
(affects ansi also)
design for string literal-like syntax and for continuation strings
(affects ansi also)
the test approach in general is first to parse basic examples of each
kind of token, then to manually come up with some edge cases to test,
and then to generate a good representative set of tokens (probably the
same set as the previous two categories), and create the cross product
of pairs of these tokens, eliminate ones when the tokens are next to
each other and it doesn't parse as the two separate tokens, using
manually written rules (want to be super accurate here - no false
positives or negatives), then test these all parse good as
well. Separating out the lexing in this way and doing this approach I
think gives a very good chance of minimising bugs in the basic
parsing, especially in the hairy bits.
= lexical syntax
One possible gotcha: there isn't a one-one correpsondence between e.g
identifiers and string literals in the lexical syntax, and identifiers
and string literals in the main syntax.
== ansi
=== symbol
=== identifier
+ escaping
=== quoted identifier
+ escaping, prefixes
=== host param
=== string literal-like
+ escaping, prefixes
=== number literals
=== whitespace
=== comments
== postgresql
=== symbol
=== identifier
== postgresql
=== symbol
extended set of symbols + extensibility + special cases
: is a symbol and also part of host param
=== identifier
same as ansi? is the character set the same?
=== quoted identifier
same as ansi?
=== host param
same as ansi (check char set)
=== string literal-like
dollar quoting
E quoting
missing n'?
=== number literals
same as ansi, i think
=== whitespace
same as ansi
=== comments
same as ansi
=== additions
$1 positional parameter
---- find what else is in hssqlppp to support mysql, oracle, sql
server