Safe Temporary Cleanup on Exception (Variable Liveness)

https://gist.github.com/dstogov/43eaa3f3b68583877322It looks like the
last version of the patch https://github.com/php/php-src/pull/1634 solves
all the problems (except for reported by Nikita "unknown liveness in
finally" - Zend/tests/temporary_cleaning_010.phpt).

Please, analyse the patch and try to craft a script not covered by the
patch.
If nobody object and no problems found, I'll commit this on Friday.

Thanks. Dmitry.

I've committed unrelated and obvious parts (without semantic changes).
The updated and extended patch now is much more clean
https://gist.github.com/dstogov/43eaa3f3b68583877322
Anyway, it's still incomplete.

Thanks. Dmitry.

On Tue, Nov 10, 2015 at 1:42 PM, Nikita Popov nikita.ppv@gmail.com
wrote:

Hi,

Please take a look into
https://gist.github.com/dstogov/43eaa3f3b68583877322

This is variation on Bob's idea about variable liveness.
The patch is incomplete. Liveness construction is still naive. However,
all tests are passed.

Ranges are represented by start/end/variable cartages. (I suppose this
representation is more compact, than in the original proposal).
In general, it's possible to use few ranges for the same variable, if
its whole live range is not linear. (this is not implemented yet).

Related ideas and tests are welcome.

Thanks. Dmitry.

One leak this does not cover yet is a discarded delayed return value:

function test() {
try {
$a = [1, 2, 3];
return $a + [];
} finally {
throw new Exception;
}
}

try {
test();
} catch (Exception $e) {}

This case is a bit tricky because it violates the usual invariant that
all temporaries have well-defined liveness ranges (either alive or not, no
maybe). We'd have to change that first.

I see the problem. We can't statically define live-ranges for finally
code, because we don't know where it's going to be called from.
May be we may reconstruct this at run-time.

Thanks. Dmitry,

Nikita

Hey:

On Tue, Nov 10, 2015 at 1:42 PM, Nikita Popov nikita.ppv@gmail.com
wrote:

Hi,

Please take a look into
https://gist.github.com/dstogov/43eaa3f3b68583877322

This is variation on Bob's idea about variable liveness.
The patch is incomplete. Liveness construction is still naive. However,
all tests are passed.

Ranges are represented by start/end/variable cartages. (I suppose this
representation is more compact, than in the original proposal).
In general, it's possible to use few ranges for the same variable, if
its whole live range is not linear. (this is not implemented yet).

Related ideas and tests are welcome.

Thanks. Dmitry.

One leak this does not cover yet is a discarded delayed return value:

function test() {
try {
$a = [1, 2, 3];
return $a + [];
} finally {
throw new Exception;
}
}

try {
test();
} catch (Exception $e) {}

This case is a bit tricky because it violates the usual invariant that
all temporaries have well-defined liveness ranges (either alive or not, no
maybe). We'd have to change that first.

I see the problem. We can't statically define live-ranges for finally
code, because we don't know where it's going to be called from.
May be we may reconstruct this at run-time.

Or maybe we could separate ZEND_RETURN(_BY_REF) into two parts

the first part, set the return value to EX(return_value), the second part
do the really leave(leave_helper).

thus, we can keep the liveness of the variable won't be intrruptted?

like: previously we emit: ZEND_ADD, ZEND_FAST_CALL, ZEND_RETURN.

    now we emit: ZEND_ADD,  ZEND_RETURN, ZEND_FAST_CALL, ZEND_RETURN

s #ZEND_RETURN$#ZEND_LEAVE#

thanks

    then in HANDLE_EXCEPTION we can dtor EX(return_value) if it is

already set?

this could also fixed double return memory leak
like Zend/tests/try/bug70228.phpt

thanks

Thanks. Dmitry,

Nikita

--
Xinchen Hui
@Laruence
http://www.laruence.com/

--
Xinchen Hui
@Laruence
http://www.laruence.com/