PHP: rfc:jit-ir

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
====== PHP RFC: A new JIT implementation based on IR Framework ======
  * Version: 1.0
  * Date: 2023-09-20
  * Author: Dmitry Stogov, dmitry@php.net
  * Status: Accepted
  * First Published at: http://wiki.php.net/rfc/php-ir

===== Introduction =====
The first version of JIT for PHP was released in PHP-8.0 (https://wiki.php.net/rfc/jit). That time we took a quite simple approach. We are generating native code directly from Zend VM byte code. This approach allowed to deliver JIT in a quite short time, but it also created significant limitations in possible optimizations. Smart optimizations require a more formal and detailed **Intermediate Representation** (IR). The introduction of the AArch64 JIT back-end in PHP-8.1 made the maintenance of JIT a bit more complex, because now we often have to update assembler code in two back-ends.
 
A smarter JIT compiler with some Intermediate Representation was planned for a long time, and the real work was started in January 2022.
This proposal is the result of these 1.75 years of work.
 
===== Proposal =====
I propose a new JIT implementation that is based on a separately developed [[https://github.com/dstogov/ir|IR Framework]].

Instead of separate back-ends that generated x86 and AArch64 code, now we have a single back-end that constructs IR.

This IR is passed to the framework that performs: machine independent optimizations, register allocation, scheduling, and native code generation. The new back-end doesn't have to care about many old low-level details, e.g.: register assignment, different calling conventions, etc.

On the other hand, sometimes the construction of IR may be more complex because all the dependencies must be consistently reflected. This helps to find many errors during formal IR verification.

The new JIT implementation will affect only a few core PHP developers who are involved in the JIT development or maintenance.

JIT is a self-contained subsystem, and the changes in the implementation details (even big changes) don't affect others.

The proposed [[https://github.com/php/php-src/pull/12079|patch]] makes few changes outside **ext/opcache/jit**. It also very carefully adds new IR related code into **zend_jit.c** and **zend_jit_trace.c**, keeping the ability to work with old and new back-ends.

At the current state the PR doesn't remove the sources of the old JIT implementation. It is possible to switch to the old implementation by configuring PHP with **--disable-opcache-jit-ir**.

JIT is a quite complex subsystem. Its first implementation will likely have bugs, but I am confident that these can be addressed. We have more time to fix potential issues the sooner this proposal passes, and the patch is merged.

The main goal of a separate IR Framework development is the collaboration with other compiler experts (sharing expertise, contributions, etc.). I'm doing my best to make the IR project interesting outside the PHP world and involve others. In case of success, this will reduce the PHP "bus-factor".

The necessary part of the IR Framework is embedded into the PHP source tree, and won't introduce any new external dependencies.

The details of the IR framework are complex. This [[https://www.researchgate.net/publication/374470404_IR_JIT_Framework_a_base_for_the_next_generation_JIT_for_PHP|presentation]] explains design ideas and provides an overview of the most important implementation details.

===== Backward Incompatible Changes =====

There are no BC breaks in regard to user land PHP code.

===== Proposed PHP Version(s) =====

next (master branch)

===== RFC Impact =====

==== Performance ====

The current new PHP JIT implements exactly the same set of PHP related features and optimizations as the old one.

Because the IR Framework provides more optimizations and has a smarter Register Allocator, the JIT produces a bit faster (5-10%) and smaller code.

This is visible with **bench.php** and **micro_bench.php**. The speed of real-life applications is not affected.

The compilation speed of the Tracing JIT is almost the same. The compilation speed of the Function JIT is up to 4 times slower (tested on Wordpress).

This is quite a good result for an Optimizing JIT compiler. In my tests, PHP with the new JIT may produce ~15MB of native code per second.   

==== Zend VM ====

On JIT supported platforms (x86, x86_64, AArch64), the HYBRID VM interpreter starts to save and restore all the persistent CPU registers at **execute_ex()** prologue and epilogue.

This allows elimination of the similar save/restore code in each JIT-ed function or trace. CLANG and MSVC builds that use CALL VM are not impacted. 

==== To Opcache ====

JIT is implemented as a part of Opcache, so the PR changes the build process for Opcache.
No other changes are present.

==== php.ini Defaults ====

The PR does not introduce new php.ini directives or change any default values.

It adds few new IR related JIT debugging capabilities that may be enabled through **opcache.jit_debug** directive bits:

  * 0x01000000 - print the initial IR graph right after construction 
  * 0x02000000 - print the final IR graph right before code emission
  * 0x04000000 - print Control Flow Graph
  * 0x08000000 - print information about Live Ranges and Allocated Registers
  * 0x10000000 - print IR graph after Sparse Conditional Constants Propagation pass
  * 0x20000000 - print IR graph after Global Code Motion and linearization pass
  * 0x40000000 - print IR graph after Register Allocation pass
  * 0x80000000 - annotate IR graph with information about CFG, Registers and Code Generation rules

==== To SAPIs ====

SAPIs are not affected.

==== To Existing Extensions ====

PHP extensions are not affected.
 
==== New Constants ====

There are no new constants.

===== Open Issues =====

Should we keep the old implementation of the JIT for a while, or change the patch to replace it straight away?

===== Unaffected PHP Functionality =====

JIT is developed as a self-contained feature, and doesn't affect anything else directly.

===== Future Scope =====

The usage of IR opens the door for more powerful optimizations. Some of these are going to be PHP independent (e.g. the planned introduction of new loop optimization passes, improved LOAD/STORE motion and redundancy, and improvements to the code generator).

The other part is PHP specific (for example, we can extend CPU register usage for more instructions).

It is also possible to support new JIT targets (e.g. RISCV), almost independently of PHP.

Finally, we may try to completely avoid the manual IR construction in JIT. We may introduce a single formal specification for VM instructions in a C-like language, convert it to IR and then use partial evaluation to generate VM and JIT handlers (similar to Truffle). 

===== Proposed Voting Choices =====

Voting opened 2023-10-06 and closes 2023-10-19.

<doodle title="Merge the new JIT implementation into PHP master?" auth="user" voteType="single" closed="true">
   * Yes
   * No
</doodle>


Secondary vote: Should we keep the old JIT implementation for a while or remove it right after merge?

Note, that old implementation is not going to be tested (by CI) and we can't guarantee its functionality anyway.

<doodle title="Remove the old JIT implementation right after merging the new one?" auth="user" voteType="single" closed="true">
   * Yes
   * No
</doodle>


===== Patches and Tests =====

  - https://github.com/php/php-src/pull/12079
  - All CI tests (including nightly workflow) passed: https://github.com/dstogov/php-src/actions/runs/6123676012/job/16638279521 

===== Implementation =====

   - merged into master via https://github.com/php/php-src/commit/caf102dfaef1ef4191c76a4502c6b4fef20cba46
   - old JIT implementation is removed via https://github.com/php/php-src/commit/be275433d936e41d95b2fd656464bcc4d7138b76

===== References =====

https://github.com/dstogov/ir
Page Tools
