PHP RFC: More precise float value handling

• Version: 1.1
• Date: 2015-07-30
• Author: Yasuo Ohgaki yohgaki@ohgaki.net, Jakub Zelenka bukka@php.net
• Status: Under Discussion
• First Published at: http://wiki.php.net/rfc/precise_float_value

This RFC is based on the discussion about displaying float values in json_encode and proposes more precise float value handling overall.

JSON is used to exchange data between systems. Although JSON RFC "6 Numbers" does not require specific implementation for float/int type, float value should be handled as precise as possible by default.

Currently, json_encode() uses EG(precision) which is set to 14. That means that 14 digits max are used for displaying (printing) the number. IEEE 754 double supports higher precision and serialize()/var_export() uses PG(serialize_precision)=17 to be more precise. Since json_encode() uses EG(precision), json_encode() removes lower digits of fraction parts and destroys original value even if PHP's float could hold more precise float value.

<?php
$j = '{ "v": 0.1234567890123456789 }';
var_dump(json_decode($j));
var_dump(json_encode(json_decode($j)));
ini_set('precision', 20);
var_dump(json_decode($j));
var_dump(json_encode(json_decode($j)));
var_dump(0.1234567890123456789);
?>

object(stdClass)#1 (1) {
  ["v"]=>
  float(0.12345678901235)
}
string(22) "{"v":0.12345678901235}"
object(stdClass)#1 (1) {
  ["v"]=>
  float(0.12345678901234567737)
}
string(28) "{"v":0.12345678901234567737}"
float(0.12345678901234567737)

PHP's float type stores “raw” IEEE 754 double and could display accurate fraction value up to 17 digits.

Current PHP outputs meaningless values for oversized EG(precision)/PG(serialize_precision).

<?php
$v = 0.12345678901234567890;
var_dump($v);
ini_set('precision', 100);
var_dump($v);
?>

float(0.12345678901235)
float(0.12345678901234567736988623209981597028672695159912109375)

That is caused by used mode for double to string conversion.

This RFC proposes to introduce EG(precision)=-1 and PG(serialize_precision)=-1 that uses zend_dtoa()'s mode 0 which uses better algorigthm for rounding float numbers (-1 is used to indicate 0 mode).

The RFC also proposes changing ini for JSON precision to PG(serialize_precision).

Followings are sample codes and outputs of the proposed patch.

NEW behavior

<?php
$v = 10.0000000000001;
 
ini_set('precision', -1);
ini_set('serialize_precision', -1);
 
var_dump($v);
echo var_export($v, true), PHP_EOL;
echo json_encode($v), PHP_EOL;
echo $v, PHP_EOL;
?>

float(10.0000000000001)
10.0000000000001
10.0000000000001
10.0000000000001

OLD behavior

<?php
$v = 10.00000000000001;
 
ini_set('precision', 14);
ini_set('serialize_precision', 17);
 
var_dump($v);
echo var_export($v, true), PHP_EOL;
ini_set('serialize_precision', 14);
echo json_encode($v), PHP_EOL;
ini_set('serialize_precision', 17);
echo $v, PHP_EOL;
?>

float(10)
10.000000000000011
10
10

Please note that IEEE float cannot store exactly precise values. e.g. Result of “10/3” - see phpt of the patch. Even with this proposal, there will be rounding errors, but the behavior becomes similar to other languages and values are more precise in many cases.

Setting mode 0 as default can mean that the rounding will be more precise which also means that the rounding might be different in var_export()/serialize().

The BC break could happen only if someone would rely on exact output but that shouldn't be the case. All our existing tests passes when 0 mode is used.

None when old INI value is used.

• PHP 7.1

None.

PHP overall

• 0 mode (EG(precision)= -1) float outputs values rounded to nearest.

Standard module and JSON

• serialize(), var_export(), json_encode - Use PG(serialize_precision) and 0 mode by default.

Not affected.

None.

precision

• hardcoded default values : 14 Unmodified
• php.ini-development values : 14 Unmodified
• php.ini-production values : 14 Unmodified

serialize_precision

• hardcoded default values : -1
• php.ini-development values : -1
• php.ini-production values : -1

None.

PHP uses “raw” IEEE 754 value internally regardless of precision settings. Therefore, this RFC does not affect internal computation.

WDDX

• wddx_serialize_vars/value() - Use PG(serialize_precision) and 0 mode. It uses EG(precision) currently.

XML_RPC

• xmlrpc_encode() - Use PG(serialize_precision) and 0 mode. It uses EG(precision) currently.

Requires a 50%+1 majority

There will be two votings

• whether to introduce mode 0 and use it as default for serialize_precision
• should PG(serialize_precision) be used instead of EG(precision) in json_encode.

The initial PR can be found here:

• https://github.com/php/php-src/pull/1455

Note that the PR is currently outdate but it will be updated if the RFC is accepted.

After the project is implemented, this section should contain

1. the version(s) it was merged to
2. a link to the git commit(s)
3. a link to the PHP manual entry for the feature

• https://marc.info/?l=php-internals&m=143786181229714&w=2
• https://bugs.php.net/bug.php?id=71473
• http://blog.reverberate.org/2016/02/06/floating-point-demystified-part2.html

Keep this updated with features that were discussed on the mail lists.