Re: Re: com php-src: Convert ASSIGN_ADD $a, $b into $a = ADD $a, $b, if possible.: ext/opcache/Optimizer/dfa_pass.c

On Wed, Mar 23, 2016 at 5:38 AM, Matt Wilmas <php_lists@realplain.com> wrote: > Hi Dmitry, > > ----- Original Message ----- > From: "Dmitry Stogov" > Sent: Tuesday, March 22, 2016 > > Commit: d8b75b0807a5d94bd7b6b175d56aba8bc5be8d7a >> Author: Dmitry Stogov <dmitry@zend.com> Tue, 22 Mar 2016 >> 23:57:26 +0300 >> Parents: 76d612129bb6b67171308f43c19d7e4791d7919e >> Branches: master >> >> Link: >> >> http://git.php.net/?p=php-src.git;a=commitdiff;h=d8b75b0807a5d94bd7b6b175d56aba8bc5be8d7a >> >> Log: >> Convert ASSIGN_ADD $a, $b into $a = ADD $a, $b, if possible. >> >> Changed paths: >> M ext/opcache/Optimizer/dfa_pass.c >> > > So this changes $a += $b into $a = $a + $b? > > What is the purpose of that...? I always use the former, because I > thought the latter created a temp value for $a + $b prior to assignment... > (Maybe I'm wrong or it's different now? :-/) > > Otherwise, is the idea to "deaden" the ASSIGN_* handlers and consolidate > stuff to a common handler, thus improving CPU instruction cache usage? > $a = $a + $b compiles to T0 = ADD $a, $b; ASSIGN $a, T0. This transformation compiles to $a = ADD $a, $b, i.e. $a is directly used as the result operand of the ADD opcode. This is only possible because we know that $a does not contain a refcounted type prior to this operation, so its value does not require destruction. $a = ADD $a, $b is faster than ASSIGN_ADD $a, $b because it doesn't have to do a whole slew of checks and has a fast-path for common cases. Furthermore (and in this instance, more importantly) the ADD opcode will be picked up by type-specialization. Nikita