HHVM: The Fiddly Bits

Ian Littman

@iansltx

http://ian.im/hh15

whoami

• Back-end dev for Access Empowerment
• I build APIs...and occasionally stuff in front of them
• I use HHVM in production for a non-AE project
• iansltx on Twitter, GitHub, most other places

What we'll talk about

• HHVM for PHP
  • The theory behind it all
  • Basic setup
  • Performance (benchmarks!)
  • Configuration gotchas
  • Compatibility gotchas
• HHVM for Hack
  • Basic setup
  • Language features
  • What does it look like in practice?
  • Is it faster than PHP on HHVM?
• Case studies

The Theory

• Code doesn't change that much during runtime, even for a dynamic language like PHP
• If we run the code a bit, we'll figure out the parts that are predictable from a type definition perspective
• If we can make (correct) assumptions about types, etc. within a block of code, we can omit type-checking code within that block
• In some cases, we can go all the way to machine code with the above optimizations...for a nice performance boost

The Implementation

1. Profile the code a bit in interpreter-only mode
   1. 11 times by default
   2. hhvm.eval.jit_warmup_requests = int
   3. Figure out which types are seen most
2. JIT compile hot code at runtime
   1. HHBC -> HHIR
   2. HHIR -> VASM (arch independent)
   3. VASM -> ASM/machine code (x86-64)
3. Bail out to interpreter mode if needed

How did we get to HHBC?

• Depends on whether we're in Repo Authoritative mode
• Interpreter mode (PHP5/7-like)
  • On-request...
  • Parse, lex, build Abstract Syntax Tree
  • Optimize obvious cases
  • Build HipHop Assembly codes
  • Build HipHop Bytecode (HHBC)
• RepoAuth mode (ahead-of-time analysis step)
  • Static analysis to eliminate dead code, optimize
  • All of the above, up to HHBC
  • Store HHBC in a SQLite database, read on runtime

Config Gotchas

• Thread count is defaulted to 2 * core count
  • For smaller machines, kills high-concurrency perf
  • hhvm.server.thread_count = int
• Logging (hhvm.error_handling.)
  • call_user_handler_on_fatals = 0
    • Error code 2^24 if turned on...catch all the fatals!
  • throw_exception_on_bad_method_call = 1
    • SPL exception...different in some cases than PHP5
• func_get_args() will get you current arg values if changed
  • PHP 7 does this too
• No MySQL PDO over SSL
• No easter_date()

Benchmarks!

• Repo Auth
• Standard 1
• Standard 2
• Hack
• PHP 5.6.6 + OPCache w\no change detection
• Vultr 768M Ubuntu 14.10 x64 VPS in Dallas
• HHVM 3.6.0
• iansltx/raphple as of today
• 1-500 users over 1 minute hitting the home page
• Twilio needed a couple mods for RepoAuth

Setup

• Packages exist for Ubuntu
• Uses FastCGI, just like php-fpm
• Compiling takes an hour or two; see the GitHub wiki for instructions
  • CentOS 7
  • Amazon Linux
  • Short-circuit Intel TBB's test suite to go faster
• Drop-in replacement for PHP5

Hacklang

• The reason PHP 7 has a bunch of new language features
• A superset of PHP, minus a few ugly bits
• Currently syntactic sugar, plus async support, plus static analysis and type-checking
• Minimal performance impact for sequentially executed code...for now

How do I write it?

• Plugins
  • Sublime
  • GitHub Atom
  • Vim
  • Emacs
• Support in PHPStorm coming soon

Language Features

• Types...types, everywhere!
  • Scalar typehints (rather strict)
  • Return typehints (including scalars)
  • Nullable types, e.g. function bar(?string $bar)
  • Generics
  • Collection types
    • Vector
    • Map
    • Set
    • Pair

Language Features

• Shapes (lightweight typed structures)
• Type aliases
• Annotations (not docblocks)
  • <<__annotation>> syntax
  • e.g. <<__Memoize>>
• XHP - HTML as a first-class citizen
  • Escaping for "free"
  • Extensible syntax, e.g. for Bootstrap
  • Available as a PHP extension

Language Features

• async/await
  • Multithreading without callback hell
  • Non-blocking I/O
    • curl
    • files
    • memcached
    • mysql (as of 3.6!)
  • Here's the big perf. improvement vs. PHP

Language Features

• Type checker (hh_client)
  • Watches file system
  • On-save, not run at compile/runtime
  • Detailed information on type errors etc.
  • Doesn't catch everything...
  • ...but that may be because I mixed Hack and PHP

Typechecker output

/usr/share/nginx/html/public/index.php:75:26,37: Too few arguments (Typing[4104])
  /tmp/hh_server_root/hhi_1e44a56a/collections/Vector.hhi:50:19,29: Definition is here
/usr/share/nginx/html/public/index.php:80:33,36: Could not find method push in an object of type Vector (Typing[4053])
  /usr/share/nginx/html/public/index.php:75:26,37: This is why I think it is an object of type Vector
  /tmp/hh_server_root/hhi_1e44a56a/collections/Vector.hhi:45:13,18: Declaration of Vector is here
/usr/share/nginx/html/public/index.php:80:33,36: Could not find method push in an object of type Vector (Typing[4053])
  /usr/share/nginx/html/public/index.php:75:26,37: This is why I think it is an object of type Vector
  /tmp/hh_server_root/hhi_1e44a56a/collections/Vector.hhi:45:13,18: Declaration of Vector is here
/usr/share/nginx/html/public/index.php:92:34,38: Invalid argument (Typing[4110])
  /usr/share/nginx/html/public/index.php:102:34,36: This is an int
  /usr/share/nginx/html/public/index.php:91:33,38: It is incompatible with a string
/usr/share/nginx/html/public/index.php:92:34,38: Invalid argument (Typing[4110])
  /usr/share/nginx/html/public/index.php:102:34,36: This is an int

Yep, you can mix 'em

• Can call Hack from PHP, and vice versa
• Hackificator (the reason most of Fb's code is Hack)
• Can translate back and forth automatically*
  • * Can't do with async/await
• ...so you can start with a single file and work outward
• Here's what I did in an afternoon...raphple runs on it

That's it!

Questions?

@iansltx / http://ian.im/hh15