rishi/scryer-prolog
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Rewrote many parts of library(json) to leave no choicepoints when generating JSON. However, the generating performance actually worsened slightly...

Browse Source
This commit is contained in:
panasenco
2021-04-22 13:31:06 -07:00
parent d3caba4073
commit 828550687c
4 changed files with 165 additions and 147 deletions
Download Patch File Download Diff File Expand all files Collapse all files

275
src/lib/json.pl
View File

@@ -47,89 +47,58 @@
:- use_module(library(reif)).
/* The DCGs are written to match the McKeeman form presented on the right side of https://www.json.org/json-en.html
almost perfectly. Note that the McKeeman form conflicts some with the pictures on the left side. */
as closely as possible. Note that the names in the McKeeman form conflict with the pictures on the site. */
json_chars(Internal) --> json_element(Internal).
/* Because it's impossible to distinguish between an empty array [] and an empty string "", we distinguish between
different types of values based on their principal functor. The principal functors match the types defined in
the JSON Schema spec here: https://json-schema.org/draft/2020-12/json-schema-validation.html#rfc.section.6.1.1
Down the line we'll incorporate more JSON Schema support, but this is it for now. */
EXCEPT we don't yet support the integer type. There are plans for more JSON Schema support in the near future. */
json_value(object(Assoc)) --> json_object(Assoc).
json_value(array(List)) --> json_array(List).
json_value(string(Chars)) --> json_string(Chars).
json_value(number(Number)) --> json_number(Number).
json_value(boolean(true)) --> "true".
json_value(boolean(false)) --> "false".
json_value(boolean(Bool)) --> json_boolean(Bool).
json_value(null) --> "null".
/* Note on variable instantiation checks (`var/1` and `nonvar/1`) used below and in Prolog in general.
Instantiation checks should never ever be used to change the logic of your program! Instead, they are one of
many tools to adjust the 'control' or 'search strategy' used by Prolog to execute the logic of your program.
Control tweaks are used for the following:
- Prevent instantiation errors.
- Prevent nontermination.
- Improve the time complexity of execution (e.g. from superexponential to linear).
For a general overview of the idea, read Bob Kowalski's "Algorithm = Logic + Control":
https://www.doc.ic.ac.uk/~rak/papers/algorithm%20=%20logic%20+%20control.pdf
For an introduction to search strategies in Prolog, read: https://www.metalevel.at/prolog/sorting#searching
This DCG definition does two things:
1. Logic: Relate an association list to a JSON object serialized in a string.
2. Control: Define the exact strategy by which we obtain an association list from a JSON string and vice versa.
This is done via instantiation checks `var/1` and `nonvar/1`.
Unfortunately, the logic and control in this DCG aren't separated cleanly like Bob Kowalski proposed.
Maybe at some point in the future we'll have a library that takes a pure logic character parsing/generating DCG
and 'injects' control strategy into it. We aren't there yet... */
json_object(EmptyAssoc) --> {empty_assoc(EmptyAssoc)}, "{", json_ws, "}".
json_object(Assoc) -->
{ ( nonvar(Assoc) ->
\+ empty_assoc(Assoc),
assoc_to_list(Assoc, [Pair|Pairs])
; true
) },
/* We pull json_boolean out into its own predicate in order to take advantage of first argument indexing and not leave
choice points. For more details, watch this video on decomposing arguments: https://youtu.be/FZLofckPu4A?t=1648 */
json_boolean(true) --> "true".
json_boolean(false) --> "false".
json_object([]) --> "{", json_ws, "}".
json_object([Pair|Pairs]) -->
"{",
json_members([Pair|Pairs]),
"}",
{ ( var(Assoc) ->
list_to_assoc([Pair|Pairs], Assoc)
; true
) }.
json_members(Pairs, Pair),
"}".
/* Why have both `json_members//1` and `json_members_//2`? Wouldn't it be less confusing to have just
`json_members//1`?
In fact in the first version of the code there was just this simple definition of `json_members//1`:
/* `json_members//2` below is implemented with a lagged argument to take advantage of first argument indexing.
This is a pure performance-driven decision that doesn't affect the logic. The predicate could equivalently be
implementes as `json_members//1` below:
```
json_members([Key-Value]) --> json_member(Key, Value).
json_members([Key-Value | Pairs]) --> json_member(Key, Value), ",", json_members(Pairs).
json_members([Key-Value]) --> json_member(Key, Value).
json_members([Key-Value, Pair2 | Pairs]) --> json_member(Key, Value), ",", json_members([Pair2 | Pairs]).
```
The problem with this definition was that there's no way for Prolog to distinguish between the two DCG heads,
because [Key-Value] unifies with [Key-Value|[]], which unifies with [Key-Value|Pairs].
Therefore, such a representation is defaulty, and is actually misleading because when you look at it you think
that a list with only one pair would apply to only the first definition, but it actually applies to both!
For more info on clean vs defaulty representations, read: https://www.metalevel.at/prolog/data#clean
The below definition, while longer, cleanly distinguishes member lists with just one value from member lists
with two or more values.
That's a logically equivalent and equally clean representation to the lagged argument. However, it leaves
choice points, while using the lagged argument doesn't. For more info, watch: https://youtu.be/FZLofckPu4A?t=1737
*/
json_members([Pair|Pairs]) --> json_members_(Pairs, Pair).
json_members_([], Key-Value) --> json_member(Key, Value).
json_members_([NextPair|Pairs], Key-Value) -->
json_members([], Key-Value) --> json_member(Key, Value).
json_members([NextPair|Pairs], Key-Value) -->
json_member(Key, Value),
",",
json_members_(Pairs, NextPair).
json_members(Pairs, NextPair).
json_member(Key, Value) --> json_ws, json_string(Key), json_ws, ":", json_element(Value).
json_array([]) --> "[", json_ws, "]".
json_array([Value|Values]) --> "[", json_elements([Value|Values]), "]".
json_array([Value|Values]) --> "[", json_elements(Values, Value), "]".
json_elements([Value|Values]) --> json_elements_(Values, Value).
json_elements_([], Value) --> json_element(Value).
json_elements_([NextValue|Values], Value) -->
/* Also using a lagged argument with `json_elements//2` to take advantage of first-argument indexing */
json_elements([], Value) --> json_element(Value).
json_elements([NextValue|Values], Value) -->
json_element(Value),
",",
json_elements_(Values, NextValue).
json_elements(Values, NextValue).
json_element(Value) --> json_ws, json_value(Value), json_ws.
@@ -138,49 +107,55 @@ json_string(Chars) --> "\"", json_characters(Chars), "\"".
json_characters("") --> "".
json_characters([Char|Chars]) --> json_character(Char), json_characters(Chars).
/* A directly printable character is defined by the JSON spec as a character between 0020 and 10FFFF except the
escaped characters.
Note that `char_code/2` throws an instantiation error if both its arguments are undefined, so we delay
calling it until we've seen both the generating and the parsing sides of the DCG.
If we moved the block containing `char_code/2` up before `[PrintChar]`, we would still be able to generate JSON,
but attempting to parse JSON would cause an instantiation error. */
escape_map([
'"' - '"',
('\\') - ('\\'),
('/') - ('/'), /* Forward slash parsed with or without a preceding backslash, but always generated with. */
'\b' - 'b',
'\f' - 'f',
'\n' - 'n',
'\r' - 'r',
'\t' - 't' ]).
letter_escape('"', '"').
letter_escape('\\', '\\').
letter_escape('/', '/').
letter_escape('\b', 'b').
letter_escape('\f', 'f').
letter_escape('\n', 'n').
letter_escape('\r', 'r').
letter_escape('\t', 't').
json_character(PrintChar) -->
{ ( nonvar(PrintChar) ->
dif(PrintChar, '/') /* Don't generate forward slash without preceding backslash */
/* Note on variable instantiation checks (`var/1` and `nonvar/1`) used below and in Prolog in general.
Instantiation checks should ideally never be used to change the logic of your program. Instead, they are one of
many tools to adjust the 'control' or 'search strategy' used by Prolog to execute the logic of your program.
For a general overview of the idea, read Bob Kowalski's "Algorithm = Logic + Control":
https://www.doc.ic.ac.uk/~rak/papers/algorithm%20=%20logic%20+%20control.pdf
For an introduction to search strategies in Prolog, read: https://www.metalevel.at/prolog/sorting#searching
However, when dealing with a real-world data format standard, real differences arise in how a string should be
parsed vs generated. Usually, parsing should allow multiple ways of doing things, while generating should only
happen in one best way.
JSON characters are parsed/generated in one of three ways:
1. Directly. All characters in the range 20.10FFFF, except '"' and '\\' must be generated and parsed directly,
escape for the forward slash '/', which must not be generated directly, but can be parsed directly.
2. Backslash followed by a single special character defined in the escape map - both parsing and generating.
3. Backslash followed by 'u' and 4 hex values defining the character code of the internal character.
When generating, only allow range 0.20 excepting characters in the escape map.
When parsing, allow any value.
In order to take advantage of first argument indexing, we must reify this distinction in a single predicate. */
json_character(InternalChar) -->
{ ( nonvar(InternalChar) ->
( letter_escape(InternalChar, _) ->
Type = letter_escape
; char_code(InternalChar, InternalCharCode),
( InternalCharCode >= 32 ->
Type = direct
; Type = hex_escape
)
)
; true
) },
[PrintChar],
{ dif(PrintChar, '"'),
dif(PrintChar, '\\'),
char_code(PrintChar, PrintCharCode),
PrintCharCode >= 32 /* 20.10FFFF */ }.
json_character(EscapeChar) --> "\\", json_escape(EscapeChar).
json_character(Type, InternalChar).
json_escape(EscapeChar) -->
[PrintChar],
{ escape_map(EscapeMap),
member(EscapeChar-PrintChar, EscapeMap) }.
json_escape(EscapeChar) -->
"u",
{ /* Control: Get the code of the escape character if we can. Otherwise we'll end up backtracking over 65,536
possible hex values.
Logic: Only the first 32 Unicode characters not escaped in the escape map are eligible for \u-escaping
when generating. However, we want to be able to parse any of the 65,536 \u-escaped values when parsing. */
( nonvar(EscapeChar) ->
json_character(direct, PrintChar) --> [PrintChar].
json_character(letter_escape, EscapeChar) -->
{ letter_escape(EscapeChar, PrintChar) },
"\\",
[PrintChar].
json_character(hex_escape, EscapeChar) -->
"\\u",
{ ( nonvar(EscapeChar) ->
char_code(EscapeChar, EscapeCharCode),
EscapeCharCode < 32,
escape_map(EscapeMap),
\+ member(EscapeChar-_, EscapeMap),
H1 = 0,
H2 = 0,
H3 is EscapeCharCode // 16,
@@ -191,62 +166,69 @@ json_escape(EscapeChar) -->
json_hex(H2),
json_hex(H3),
json_hex(H4),
/* Control + Logic: Get the escape character atom from the character code computed from the hexes. */
{ ( var(EscapeChar) ->
EscapeCharCode is H1 * 16^3 + H2 * 16^2 + H3 * 16 + H4,
char_code(EscapeChar, EscapeCharCode)
; true
) }.
json_hex(Digit) --> json_digit(Digit).
json_hex(10) --> "a".
json_hex(11) --> "b".
json_hex(12) --> "c".
json_hex(13) --> "d".
json_hex(14) --> "e".
json_hex(15) --> "f".
json_hex(10) --> "A".
json_hex(11) --> "B".
json_hex(12) --> "C".
json_hex(13) --> "D".
json_hex(14) --> "E".
json_hex(15) --> "F".
json_hex(Value) -->
{ ( nonvar(Value) ->
( between(0, 9, Value) ->
Code is Value + 48
; ( between(10, 15, Value) ->
Code is Value + 87
; false
)
),
char_code(Char, Code)
; true
)
},
[Char],
{ ( var(Value) ->
char_code(Char, Code),
( between(48, 57, Code) ->
Value is Code - 48
; ( between(65, 70, Code) ->
Value is Code - 55
; ( between(97, 102, Code) ->
Value is Code - 87
; false
)
)
)
; true
) }.
/* Here we are going to write completely different DCGs for parsing and generating, and rely on built-in
predicates. However, the underlying logic remains the same. */
/* Here we are going to simply rely on `number_chars/2` when generating. */
json_number(Number) -->
{ ( nonvar(Number) ->
number_chars(Number, NumberChars)
; false
) },
NumberChars.
json_number(Number) -->
{ var(Number) },
json_sign_noplus(Sign),
json_integer(Integer),
json_fraction(Fraction),
json_exponent(Exponent),
{ ( Exponent >= 0 ->
Base = 10
; Base = 10.0
),
Number is Sign * (Integer + Fraction) * Base ^ Exponent }.
( { nonvar(Number) } ->
{ number_chars(Number, NumberChars) },
NumberChars
; json_sign_noplus(Sign),
json_integer(Integer),
json_fraction(Fraction),
json_exponent(Exponent),
{ ( Exponent >= 0 ->
Base = 10
; Base = 10.0
),
Number is Sign * (Integer + Fraction) * Base ^ Exponent }
).
json_integer(Digit) --> json_digit(Digit).
json_integer(TotalValue) -->
json_onenine(FirstDigit),
json_digits(RemainingValue, Power),
{ TotalValue is FirstDigit * 10 ^ (Power + 1) + RemainingValue }.
json_onenine(FirstDigit),
json_digits(RemainingValue, Power),
{ TotalValue is FirstDigit * 10 ^ (Power + 1) + RemainingValue }.
json_digits(Digit, 0) --> json_digit(Digit).
json_digits(Value, Power) -->
json_digit(FirstDigit),
json_digits(RemainingValue, NextPower),
{ Power is NextPower + 1,
Value is FirstDigit * 10^Power + RemainingValue }.
json_digit(0) --> "0".
json_digit(Digit) --> json_onenine(Digit).
json_digit(FirstDigit),
json_digits(RemainingValue, NextPower),
{ Power is NextPower + 1,
Value is FirstDigit * 10^Power + RemainingValue }.
json_onenine(1) --> "1".
json_onenine(2) --> "2".
@@ -258,6 +240,17 @@ json_onenine(7) --> "7".
json_onenine(8) --> "8".
json_onenine(9) --> "9".
json_digit(0) --> "0".
json_digit(1) --> "1".
json_digit(2) --> "2".
json_digit(3) --> "3".
json_digit(4) --> "4".
json_digit(5) --> "5".
json_digit(6) --> "6".
json_digit(7) --> "7".
json_digit(8) --> "8".
json_digit(9) --> "9".
json_fraction(0) --> "".
json_fraction(Fraction) -->
".",
@@ -280,8 +273,6 @@ json_sign_noplus(-1) --> "-".
json_sign(Sign) --> json_sign_noplus(Sign).
json_sign(1) --> "+".
/* Make sure json_ws doesn't attempt to generate whitespace and succeeds without choicepoints when generating */
json_ws --> [C], {nonvar(C), member(C, " \n\r\t")}, json_ws.
json_ws --> "".
json_ws --> " ", json_ws.
json_ws --> "\n", json_ws.
json_ws --> "\r", json_ws.
json_ws --> "\t", json_ws.

33
src/tests/json/README.md
View File

@@ -1,13 +1,38 @@
## Benchmarks
### With CLP(Z):
### Read
With CLP(Z):
```
?- test_json_read.
?- test_json:test_json_read.
% CPU time: 41.522 seconds
```
### After removing CLP(Z):
After removing CLP(Z):
```
?- test_json_read.
?- test_json:test_json_read.
% CPU time: 0.444 seconds
```
With first argument indexing optimizations:
```
?- test_json:test_json_read.
% CPU time: 0.310 seconds
```
### Write
Without first argument indexing optimizations:
```
?- test_json:test_json_minify.
% CPU time: 0.014 seconds
```
With first argument indexing optimizations:
```
?- test_json:test_json_minify.
% CPU time: 0.015 seconds
```

2
src/tests/json/test_json.pl
View File

@@ -34,7 +34,7 @@ test_json_minify :-
read_line_to_chars(RefMin, RefChars, []),
close(RefMin),
name_parse("pass_everything.json", Json),
time(once(phrase(json_chars(Json), MinChars))),
time(phrase(json_chars(Json), MinChars)),
RefChars = MinChars.
test_json_int_float :-