Fix parsing of semi-reserved tokens at offset > 4GB

[TysonAndre]

Possibly related to GH-5668
Also affects php 8.0 but this changes the
datastructures in public(?) header files.
(binary ABI change)
PHP 7.4 wouldn't even parse a class with the below example.

1. Fix parsing semi-reserved tokens such as namespace
   with an offset that's larger than 2**32
2. Fix parsing some tokens when their length is larger than 2**32

Offset is relative to yy_start(the start of the file?)

Low priority because 4GB php files and the configured memory limits to support them are extremely rare - this was only noticed when looking into edge cases in the parser - this test case is artificial

Instead of parsing the token 'namespace', the character range from 4GB before is used due to truncation to 32 bits

php > ini_set('memory_limit', '12G');
php > eval(str_repeat(' ' , 2**32) . 'class MyClass{ public static function namespace() {}}');
php > MyClass::namespace();

Warning: Uncaught Error: Call to undefined method MyClass::namespace() in php shell code:1
Stack trace:
  thrown in php shell code on line 1
php > var_export((new ReflectionClass('MyClass'))->getMethods());
array (
  0 =>
  ReflectionMethod::__set_state(array(
     'name' => '         ',
     'class' => 'MyClass',
  )),
)

[nikic]

This used 32-bit integers do avoid doubling the size of the parser stack elements...

[TysonAndre]

This can probably use 64-bit integers just for the offset and continue using 32-bit integers for the length?

Is it possible to reuse the memory for zend_string * elem->str (and possibly num) instead from stack elements?
For example, make it a union of zend_string * and size_t offset, and make the token type determine whether it's used as an ident or a zend_string? (EDIT: missed that this was already a union: typedef union _zend_parser_stack_elem {)

emit_token_with_ident:
	if (PARSER_MODE()) {
		elem->ident.offset = SCNG(yy_text) - SCNG(yy_start);
		elem->ident.len = SCNG(yy_leng);
	}
	if (SCNG(on_event)) {
		SCNG(on_event)(ON_TOKEN, token, start_line, yytext, yyleng, SCNG(on_event_context));
	}
	return token;

[nikic]

This can probably use 64-bit integers just for the offset and continue using 32-bit integers for the length?

That makes no difference, due to padding.

A way to preserve the current size would be to store a char* to the start of the keyword, and then determine the length in zend_lex_tstring. After all, keywords are [a-zA-Z_], so it's not hard to determine the length from that.

TBH I have no idea how important it is to preserve the size, it's just why I did this at the time.

[nikic]

I committed a fix using the approach I mentioned in my previous comment in cc3e03c. Please tell me if you see any further issues with this.