Avoid many string duplications, use interning (browscap-local, of
course), reduce pattern size, use more compact key-value
representation, build result array only on demand.
Testing corpus provided by Anatol against current browscap.ini lite.
About 30% of all agents are not recognized by this browscap.ini,
but this should give us decent coverage.
The parent process was releasing the child semaphore
after the child process continued execution. Now
the child semaphore is released before the child
process continues execution.
This bug was caused by the fact that dom_get_elements_by_tag_name_ns_raw
uses an empty string to filter on the default namespace (as NULL means
'no filter'), whereas in the node itself the default namespace is
signalled by nodep->ns being null.
* 'PHP-7.0' of git.php.net:/php-src: (146 commits)
Flush stderr on win32 in cli_log_message
Fixed bug #73154
FIx bug #70213
Fix dom class can't be inherited by the internal class
Another try at making concat_003 more reliable
Fix flaky openssl_pkey_new test
Make Opcache tests using the cli server more reliable
Revert "Fix #73530: Unsetting result set may reset other result set"
define php_ap_map_http_request_error function for older httpd only
add old versions of httpd support
Disable AppVeyor fast_finish
Makes the sapi web server and curl tests more reliable
Fixes the curl tests to be more reliable in Travis CI
Interpretation of curl_setopt values for boolean parameters
Fixes#65689. PDO_Firebrid / exec() does not free allocated statement.
Fix alpn_ctx leaking in openssl
Fixed bug #73373 (deflate_add does not verify that output was not truncated)
Fix IS_UNDEF comparisons in opcache
Fixed bug #73704 (phpdbg shows the wrong line in files with shebang)
Increase timing quota for small string concat test
...
In Travis CI the nowait.phpt test sometimes fails because of
the usleep() not being long enough ensure proper
execution order. Instead of depending on sleeping the
test now uses two semaphores to guarantee execution order.
Same fix already applied to ext/curl/tests/server.inc
and sapi/cli/tests/php_cli_server.inc
1. Increases the amount of time for the PHP built-in server to accept a connection
2. Outputs an error if the PHP built-in server fails
This reverts commit eb570294a2.
That commit caused a regression, so it's probably best to revert it, and
to tackle the issue for the next minor release.
1. Increased the time for the sapi/cli/tests/php_cli_server.inc to accept connections.
2. sapi/cli/tests/php_cli_server.inc has errors messages and a `stop` function.
3. bug67429.phpt uses the `stop` function to shutdown the webserver before starting a new one.
4. Removed ext/curl/tests/bug48203_multi-win32.phpt test now that ext/curl/tests/bug48203_multi.phpt
runs on Windows also.
1. Increases the amount of time for the PHP built-in server to accept a connection
2. Outputs an error if the PHP built-in server fails
3. In bug48203_multi.phpt the test no longer starts and stops multiple PHP built-in servers
These conditions were formerly `!q->pData` and `!p->pData`, and should
now be detected as undefined variables, using the special type IS_UNDEF.
Incidentally, this syntax raised a logical-not-parentheses compiler
warning, now gone.
Particularly on slower VMs, the sporadic fails can still happen.
The timing is kept in an uncritical range, but allows the tests
pass there. Mayby, it'd make sense to introduce a new group for
this kind of tests, so tests requiring exact time measurement
can be avoided on unsuitable environments.
Insert a lot more DEREFs. Some may not be necessary, but better
safe than sorry. I'm thinking the soap_hash_str_find_deref function
should become part of the zend_hash API -- we need this in many
places.
The problem was manifestated only with BIT columns and only when more than
one row was fetched. The problem was coming from the fact that in pre-7.0
times mysqlnd was using a no-copy optimization. This optimization kept the
strings (and also the BIT mask equivalents as strings) in the packet and the
zval referred to them. 7.0+ zvals cannot use no-copy and always copy. Because
of this the allocated memory for the packet was reduced by 1 by the person who
ported the driver, but the starting address of the bit area wasn't reduced.
Because of this the bit_area started at wrong address and the length decoded
wrong.
Nikita Popov