PHP RFC: Internal operator overloading and GMP improvements

• Version: 0.1
• Date: 2013-05-12
• Author: Nikita Popov nikic@php.net
• Status: Implemented in PHP 5.6
• Patch: https://github.com/php/php-src/pull/342

PHP offers facilities for large number and decimal arithmetic (GMP and BCMath), but currently using those is a PITA. This RFC proposes to improve the situation by adding support for operator overloading in internal classes. The operator overloading is exemplarily implemented for the GMP extension, while also improving GMP in various other ways along the way.

Note: This proposal is only about internal operator overloading and not about userland overloading.

There are several reasons why overloaded operators are preferable over gmp_add($a, $b) style functions.

The first is that code using overloaded operators is simply more readable. As an example, consider the following two code snippets, one using gmp_* functions, the other using overloading:

$result = gmp_mod(
    gmp_add(
        gmp_mul($c0, gmp_mul($ms0, gmp_invert($ms0, $n0))),
        gmp_add(
            gmp_mul($c1, gmp_mul($ms1, gmp_invert($ms1, $n1))),
            gmp_mul($c2, gmp_mul($ms2, gmp_invert($ms2, $n2)))
        )
    ),
    gmp_mul($n0, gmp_mul($n1, $n2))
);
 
$result = (
    $c0 * $ms0 * gmp_invert($ms0, $n0)
  + $c1 * $ms1 * gmp_invert($ms1, $n1)
  + $c2 * $ms2 * gmp_invert($ms2, $n2)
) % ($n0 * $n1 * $n2);

Even without understanding what the above code does (it's an excerpt from a Coppersmith attack on RSA), it should be obvious that the second code is a lot clearer. It makes the structure of the code immediately clear (three multiplications are summed up and the modulus is taken), whereas the function-based code actively hides any structure in the code. For mathematical operations infix notation just comes a lot more naturally.

Another advantage of overloaded operators is that it allows polymorphism for functions doing arithmetic operations. As an example, consider manually implementing a function like gmp_powm, which calculates a power of a number modulus some other number. Here is a sample implementation (direct translation of pseudo-code on Wikipedia):

function powm($base, $exponent, $modulus) {
    $result = 1;
    while ($exponent > 0) {
        if ($exponent % 2 == 1) {
            $result = $result * $base % $modulus;
            $exponent--;
        }
        $exponent /= 2;
        $base = ($base * $base) % $modulus;
    }
    return $result;
}

With operator overloading this function will work with any type of “number” that implements the basic arithmetic operators. It can work with normal PHP integers, it can work with GMP numbers, it can work with BCMath instances (assuming the overloading API is implemented for them of course):

var_dump(powm(123, 456, 789)); // int(699)
var_dump(powm(gmp_init(123), gmp_init(456), gmp_init(789))); // GMP(699)
var_dump(powm(
    gmp_init("123456789123456789"),
    gmp_init("987654321987654321"),
    gmp_init("56789567895678956789")
)); // GMP(36912902142130032810)

Without operator overloading this would not be possible. Instead one would have to implement the same function once using the normal + style operators and once using gmp_* functions (and any other set of functions you want to support). Operator overloading brings the advantages of polymorphism (a main reason we use object oriented programming) to the basic operations of the language.

Some examples what the operator overloading capability can be used for, apart from the bignum arithmetic outlined in this RFC:

• Decimal arithmetic. This is particularly important in PHP as PHP commonly deals with monetary values which can not be represented as floating point numbers.
• Date calculations. Also very common in PHP (DateTime::add etc)
• Ratio and complex arithmetic
• Unsigned arithmetic and arithmetic on other integral types PHP does not support (e.g. cross platform 64bit integers)
• Vector and matrix calculations

Due to potential pitfalls of misusing operator overloading known from other languages (most notably C++), the use of this new feature should be limited to cases where there are clear definitions to the behavior of all overloaded operators. The application of this feature should be for mathematical use cases only (as noted above), and not 'creative' applications such as changing the white balance of a picture by incrementing or decrementing the picture object.

The operator overloading is implemented using two new object handlers:

The do_operation handler is called for all overloadable operations which do not involve comparison. Its signature is:

typedef int (*zend_object_do_operation_t)(zend_uchar opcode, zval *result, zval *op1, zval *op2 TSRMLS_DC);

Here opcode is the opcode of the operation (e.g. ZEND_ADD), result is the target zval, op1 the first operand and op2 the second operand. For binary operations both operands are used, for unary operations the second operand is NULL. The return value can be either SUCCESS or FAILURE. If FAILURE is returned then the code falls back to the default behavior for the respective operator.

The following opcode values are supported:

+   ZEND_ADD
-   ZEND_SUB
*   ZEND_MUL
/   ZEND_DIV
%   ZEND_MOD
<<  ZEND_SL
>>  ZEND_SR
.   ZEND_CONCAT
|   ZEND_BW_OR
&   ZEND_BW_AND
^   ZEND_BW_XOR
xor ZEND_BOOL_XOR
~   ZEND_BW_NOT   (unary)
!   ZEND_BOOL_NOT (unary)

The unary + and - operators are indirectly supported by the following compiler transformations:

+$a  ==>  0 + $a
-$a  ==>  0 - $a

The compound assignment operators +=, -=, *=, /=, %=, <<=, >>=, .=, |=, &= and ^= are supported by the runtime transformation $a op= $b => $a = $a op $b.

The prefix operators ++ and -- are supported by the runtime transformations ++$a => $a = $a + 1 and --$b => $b = $b - 1. The same applies for the corresponding postfix operators, with the difference that a copy of the old value is returned (rather than the newly computed value).

The compare handler is called for comparisons. It has the following signature:

typedef int (*zend_object_compare_zvals_t)(zval *result, zval *op1, zval *op2 TSRMLS_DC); 

Here result is the target zval, op1 the first operand and op2 the second operand. The result zval has to be set to one of the longs -1 (indicating “greater than”), 0 (indicating “equal”) or -1 (indicating “less than”). The return value can either be SUCCESS or FAILURE, which indicate whether the comparison was successful. This return value will be the return value of compare_function.

The compare handler is called for the <, <=, ==, !=, >= and > operators and any other code using compare_function (e.g. sorting). The operators === and !== are explicitly not supported, as they have clearly defined semantics (same object handle) and I see now reason to break this.

The difference between the compare handler and the already existing compare_objects handler is that compare is called for all comparisons involving an object with a compare handler, whereas compare_objects is only called if both operands are objects and have the same compare_objects handler. Thus compare_objects can not be used to implement comparisons like $gmp == 0. A compare handler always takes precedence over a compare_objects handler.

Currently GMP is based on resources. This has several disadvantages:

• Cannot be serialized
• Cannot be directly cast to int/float/string
• Cannot be (meaningfully) dumped using var_dump
• Coerces to an integer by returning the resource ID. This can easily lead to bugs if you accidentally use the resource with an arithmetic operation. For example a GMP factorial test from out testsuite has been computing the factorial of the resource ID, rather than the factorial of the number.
• Cannot make use of the new operator overloading APIs
• Bad reporting on leaks. During the port I found that many functions leak resources, especially in error conditions.

This RFC proposes to make GMP use objects (of type GMP) as the underlying structure. Using this new structure, the RFC implements support for casting, dumping, serialization, cloning and overloaded operators. In the following there are examples for the new behaviors:

$n = gmp_init(42);
echo $n, "\n";         // 42
var_dump((string) $n); // string(2) "42"
var_dump((int) $n);    // int(42)
var_dump((float) $n);  // float(42)

var_dump($n = gmp_init(42));
var_dump($s = serialize($n));
var_dump(unserialize($s));
 
// outputs
 
object(GMP)#%d (1) {
  ["num"]=>
  string(2) "42"
}
string(33) "O:3:"GMP":1:{s:3:"num";s:2:"42";}"
object(GMP)#%d (1) {
  ["num"]=>
  string(2) "42"
}

$a = gmp_init(3);
$b = clone $a;
gmp_clrbit($a, 0);
var_dump($a, $b);
 
// Output: (Note that $b is still 3)
 
object(GMP)#1 (1) {
  ["num"]=>
  string(1) "2"
}
object(GMP)#2 (1) {
  ["num"]=>
  string(1) "3"
}

$a = gmp_init(42);
$b = gmp_init(17);
 
var_dump($a + $b);
var_dump($a + 17);
var_dump(42 + $b);
 
// Outputs the following 3 times:
 
object(GMP)#%d (1) {
  ["num"]=>
  string(2) "59"
}

The following operators are supported: +, -, *, /, %, |, &, ^, ~, << and >>. All operators work with two GMP values or one GMP value and one GMP-coercible value (e.g. strings and integers).

$a = gmp_init(42);
var_dump($a == 42,   $a == 17,    $a < 40,     $a < 100);
//       bool(true), bool(false), bool(false), bool(true)

Comparison is supported via the ==, !=, <, >, <= and >= operators. Sorting and other comparison-based operations work as well:

$arr = [gmp_init(0), -3, gmp_init(2), 1];
sort($arr);
var_dump($arr);
 
// Outputs
 
array(4) {
  [0]=>
  int(-3)
  [1]=>
  object(GMP)#1 (1) {
    ["num"]=>
    string(1) "0"
  }
  [2]=>
  int(1)
  [3]=>
  object(GMP)#2 (1) {
    ["num"]=>
    string(1) "2"
  }
}

During the refactoring of the implementation a few additional, small changes were done:

• If you pass a GMP instance to gmp_fact you will get the factorial of the GMP number (and not the factorial of the resource ID).
• Previously some functions like gmp_mod returned a long result if the second argument was long. This inconsistent and partially buggy behavior is no longer present and a GMP instance is always returned. As the GMP instance is castable to a long this does not break compatibility with scripts relying on the old behavior.
• Due to the previous change gmp_div_r no longer returns an incorrect result in some rounding modes.
• If you pass an invalid rounding mode to a function, you will now get a warning.

The addition of operator overloading does not break backwards compatibility.

The switch from GMP resources to objects can break scripts that checked whether something is a GMP integer using code like is_resource($a) && get_resource_type($a) == 'GMP integer'.

The addition of operator overloading does not affect performance (or at least I couldn't find a measurable difference). This is not surprising as the overloading code is usually placed in a rarely reached error/catch-all case.

The changes to GMP improve performance in all scenarios I measured (4M runs each):

                     NEW    OLD
a) gmp_add($a, $b)   1.07   1.25
b) gmp_add($a, 17)   1.02   1.21
c) gmp_add(42, $b)   1.20   1.84
d) $a + $b           0.76   ---

The difference between tests b) and c) is that the former makes use of an operator specialized on integers rather than creating a temporary GMP instance.

The pull request for this RFC can be found here: https://github.com/php/php-src/pull/342

The vote started on 10.06.2013 and ended on 17.06.2013. Both proposals are accepted.

Should these changes be applied for PHP 5.6?
Real name	Internal operator overloading	GMP changes	None
auroraeosrose (auroraeosrose)
dm (dm)
dragoonis (dragoonis)
gooh (gooh)
ircmaxell (ircmaxell)
kalle (kalle)
laruence (laruence)
levim (levim)
lstrojny (lstrojny)
mike (mike)
nikic (nikic)
pajoye (pajoye)
patrickallaert (patrickallaert)
peehaa (peehaa)
pierrick (pierrick)
salathe (salathe)
stas (stas)
Final result:	14	15	2
This poll has been closed.

http://markmail.org/message/y7rq5vcd5ucsbcyb: This is a rather old discussion on userland operator overloading (so not really the same as this). The patch discussed there is no longer accessible.