PHP RFC: 64 bit platform improvements for string length and integer in zval

• Version: 2.0
• Date: 2013-12-05
• Authors: Anatol Belski ab@php.net, Matt Ficken mattficken@php.net, Stephen A. Zarkos szarkos@php.net
• Status: Rejected

Current PHP zval datatype implementation uses long datatype to handle signed integer and int datatype to handle string length. The most 64 bit capable platforms PHP is used on are LP64 (Linux/most Unix-like), LLP64 (Windows), ILP64 (SPARC64). The data model used for the relevant datatypes on those platforms looks as follows

Regarding PHP that means today, even on 64 bit platforms the following features lack on consistency

• handling of strings >= 2^31
• handling of 64 bit integers
• large file support
• handling of numeric 64 bit hash keys

Thus, the current situation contains a roadblock on the way to the overall consistent 64 bit platforms support and further improvement of PHP.

To bring everything inline, a dynamic types model is being suggested. Besides the platform inconsistency elimination, this will enable several further moves (see the Future scope). Performance improvements for 64 bit platforms, just to name one, could be then the subject of a new RFC and would have the base for the further development.

The following datatypes are suggested for use in zval

Those datatypes are guaranteed to be 64 bit which makes PHP work consistent on any 64 bit platforms. The usage of this datatypes is integrated into the appropriate places across PHP. The size of zval, with default field alignment, is expected to grow by 4 bytes on LLP64 platforms only.

For the consistent LFS support, a set of portable datatypes and macros was invented and integrated.

For consistent 64 bit numeric hash keys support, the appropriate 64 bit unsigned was integrated.

The usage of long datatype continues on 32 bit platforms, the only change there is the usage of unsigned for the string length.

No further configure options are needed, the platform and compiler will be automatically recognized and an appropriate set of datatypes and macros will be activated.

• Zend/zend_int.h
• Zend/zend_types.h
• Zend/zend_stream.h
• main/php_streams.h

Generally speaking, every occurence mentioning “long” in macros or function names should be replaced with a corresponding neutral keyword, suggested “int”, in further like “lval” with “ival”, etc.

NOTE: the macro renamings change is up to the vote (Vote 2). Depending on the vote results, the macro renamings can be reverted and will become a topic of the later cleanup RFC.

NOTE: please take in account the voting option for retaining compatibility with old specs (Vote 3). Should the availability of the older specs be voted positive, the patch will be extended with them as aliases. The actual removal will be then the subject of the later cleanup RFC.

New spprintf modifier 'p' was implemented to platform independently output php_int_t datatype. That modifier can be used with 'd', 'u', 'x' and 'o' printf format specs with spprintf, snprintf and the wrapping printf implementations.

This modifier is of course available in all the spprintf/snprintf derivatives. Any of the introduced new datatypes can be used with the appropriate format spec.

• long and int in zval have to be replaced with the new integer datatypes
• code working with hashes and arrays has to use php_uint_t
• code working with filesystem objects has to use appropriate portable macros and datatypes
• code working with string lengths has to predominatly use php_size_t

NOTE the items below are to be considered in the vote (Vote 2 and 3), depending on it the patch might change. The essential part about the usage of new datatypes (Vote 1) will remain unaffected, the votes 2 and 3 are merely about the semantic changes.

• 'l', 'L', 's', 'p' parameter formats aren't available anymore
• Z_STRLEN*, *LONG*, etc. older macros aren't available anymore

PHP 5.6

• possible collisions with dependency libraries using 32 bit integer datatypes (range checks needed)
• existing extensions have to adapt zend_parsing_parameters() format
• existing extensions have to use new APIs and macros for string length and integer
• in the user space some function names need to be adapted on the new semantics, for instance long2ip() should be int2ip()

Some dead SAPIs are present in the core. They was not ported. A decision based on whether the authors are willing to support them has to be met. Then porting or removal of those SAPIs can be scheduled. The separate RFC https://wiki.php.net/rfc/removal_of_dead_sapis was created to handle this issue.

It has to do with squeezing anything possible from the 64 bit platforms, for maximal PHP benefit. No real features are going to be changed, removed or added to the PHP language.

The numbers here are the test scores one already might have seen in the other performance tests.

*, ** Some issues with the x64 versions of 5.5.8 and str_size_and_int64 when testing with opcache enabled. However issues of this kind are well known on windows (for instance #64926) and are due to some unluckily choosen memory address. So the cause persists in the mainstream and is not because of this patch.

Tutorial, tools and compatibility header to ease the migration of the PECL extensions are available. The goal is to make the same source in the new semantic compatible with older PHP versions.

	php_int_t i0, i1;
	char *s0, p0;
	php_size_t s0_len, p0_len;
 
	if(zend_parse_parameters(ZEND_NUM_ARGS() TSRMLS_CC, "iISP", &i0, &i1, &s0, &s0_len, &p0, &p0_len) == FAILURE) {
		return;
	}

	php_int_t i0;
 
	if(zend_parse_parameters(ZEND_NUM_ARGS() TSRMLS_CC, "i"), &i0) == FAILURE) {
		return;
	}
 
	if (INT_MAX < i0 || INT_MIN > i0) {
		php_error_docref(NULL TSRMLS_CC, E_WARNING, "Value '" ZEND_INT_FMT "' is out of range", i0);
		return;
	}

	php_error_docref(NULL TSRMLS_CC, E_WARNING, "Value '%pd' is out of range", i0);

	char *s0;
	php_size_t s0_len;
	php_int_t max_len;
 
	if(zend_parse_parameters(ZEND_NUM_ARGS() TSRMLS_CC, "Si", &s0, &s0_len, &max_len) == FAILURE) {
		return;
	}
 
	if (max_len >= 0 && s0_len > max_len) {
		return;
	}

char *dup_substr(zval *s, zval *i)
{
	php_size_t len;
	php_int_t max;
	char ret;
 
	convert_to_string(s);
	convert_to_int(i);
 
	len = Z_STRSIZE_P(s);
	max = Z_IVAL(i);
 
	if (max < 0 || max >= 0 && max > len)
	{
	return NULL;
	}
 
	ret = emalloc((max + 1) * sizeof(char));
 
	if (!ret) {
		return NULL;
	}
 
	memmove(ret, Z_STRVAL_P(s), max);
	ret[max] = '\0';
 
	return ret;
}

• in far perspective - easier to implement 128 bit support
• in near perspective - excellent base for 64 bit performance optimization
• easier integration on rarely used platforms
• easier integration on new platforms

The Vote 1 is the main vote for this patch. The both Votes 2 and 3 are merely to decide about the semantical replacements choosen for the patch. Should the Votes 2 and 3 result in reverting of that semantical changes, the essential patch part about the 64 bit support will not be hurt. Fixing the macro names or zpp specs is only the naming issue.

The vote begins Monday, 27 January 2014, 21:30 CET and ends Monday, 03 February 2014, 21:30 CET.

Feature branch
Windows builds
Test reports

Patch progress page
PECL porting docs'n'tools

Initial discussion brought up by Anthony Ferrara
Discussion after implementation start

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

-