[ruby-core:81569] [Ruby trunk Feature#12589] VM performance improvement proposal

Issue #12589 has been updated by vmakarov (Vladimir Makarov).


normalperson (Eric Wong) wrote:
>  
>  Btw, have you explored the GNU lightning JIT at all?
>  	http://www.gnu.org/software/lightning/
>  I'm on the mailing list and it doesn't seem very active, though...
>  
>
  
  Yes, I know about GNULighting, Eric.  It is an old project.  It is
just a portable assembler.

  Using it for JIT, it is like building a car having only a wheel.  To
get a good performance result for JIT, you still need to write a lot
of optimizations.  To get at least 50% performance result of GCC or
LLVM, somebody should spend many years to implement 10-20 most useful
optimizations.  Using tracing JITs could simplify the work as a
compiled code has a very simple control flow graph (extended basic
blocks).  But still it is a lot of work (at least 10 man years)
especially if you need achieve a good reliability and portability.

  It is possible to make a port of GCC to GNUlighting to use GCC
optimizations but it has no sense as GCC can directly generate a code
to targets supported by GNUlighting and even to much more targets.

I've been studying JITs for many years and did some research on it and
I don't see a better approach than using GCC (GCC became 30 year old
this year) or LLVM.  A huge amount of efforts of hundreds of
developers were spent on these compilers to get a reliable, portable,
and highly optimizing compilers.

  There is a myth that JVM JIT creates a better code than GCC/LLVM.  I
saw reports saying that JVM JIT server compiler achieves only 40% of
performance of GCC/LLVM on some code (e.g. a rendering code) of
statically typed languages.  That is why Azul (LLVM based java) exists
despite legal issues.

  I think Graal performance based on some articles I read is somewhere
in the middle between JVM client and server JIT compilers.  Probably
OMR is approximately the same (although I know it less than Graal).

  So saying using GCC/LLVM is the best option in my opinion, still
there is an open question how to use them.  GCC has libjit and LLVM
has MCJIT.  Their API might change in future.  It is better to use C
as the input because C definition is *stable*.

It looks like libjit and MCJIT is a shortcut in comparison with C
compilation.  But it is not that big as the biggest CPU consumers in
GCC and LLVM are optimizations not lexical analysis or parsing.  I
minimize this difference even more with a few techniques, e.g. using
precompiled headers for the environment (declarations and definitions
needed for C code compiled from a Ruby method by JIT).  By the way JVM
uses analogous approach (class data sharing) for faster startup.

Using C as JIT input makes an easy switch from GCC to LLVM and vise
verse.  It makes JIT debugging easier.  It simplifies the
environment creation, e.g. libjit would need a huge number of tedious
calls to the API to do the same.  Libjit has no ability to do inlining
too which would prevent inlining on Ruby->C->Ruby path.

So I see more upsides than downsides of my approach.  The current
performance are also encouraging -- **I have better performance on many
tests than JRuby or Graal Ruby using much less computer resources**
although I did not start yet to work on Ruby->Ruby inlining and
Ruby->C->Ruby inlining.

>  I encountered a new compatibility problem with gcc 4.9 on
>  Debian stable with -Werror=incompatible-pointer-types not
>  being supported.
>  
>  Also, my previous comment about C99 "restrict" not working  on
>  my setup still applies.
>  

My project is just at the initial stages.  There are a lot of things
to do.  When I implement inlining I will focus on JIT reliability and
stability.  I don't think MJIT can be used right now for more serious
programs.

I should remove -Werror=incompatible-pointer-types from the script and
restrict added by me.  They are not important.

The code is currently tuned for my major environment (FC25 Linux).  I
very rarely check OSX.  Some work should be done for configuring MRI
to use right options depending on the environment.

Eric, thank you for trying my code and giving a feedback.  I really
appreciate it.


----------------------------------------
Feature #12589: VM performance improvement proposal
https://bugs.ruby-lang.org/issues/12589#change-65271

* Author: vmakarov (Vladimir Makarov)
* Status: Open
* Priority: Normal
* Assignee: 
* Target version: 
----------------------------------------
  Hello.  I'd like to start a big MRI project but I don't want to
disrupt somebody else plans.  Therefore I'd like to have MRI
developer's opinion on the proposed project or information if somebody
is already working on an analogous project.

  Basically I want to improve overall MRI VM performance:

  * First of all, I'd like to change VM insns and move from
    **stack-based** insns to **register transfer** ones.  The idea behind
    it is to decrease VM dispatch overhead as approximately 2 times
    less RTL insns are necessary than stack based insns for the same
    program (for Ruby it is probably even less as a typical Ruby program
    contains a lot of method calls and the arguments are passed through
    the stack).

    But *decreasing memory traffic* is even more important advantage
    of RTL insns as an RTL insn can address temporaries (stack) and
    local variables in any combination.  So there is no necessity to
    put an insn result on the stack and then move it to a local
    variable or put variable value on the stack and then use it as an
    insn operand.  Insns doing more also provide a bigger scope for C
    compiler optimizations.

    The biggest changes will be in files compile.c and insns.def (they
    will be basically rewritten).  **So the project is not a new VM
    machine.  MRI VM is much more than these 2 files.**

    The disadvantage of RTL insns is a bigger insn memory footprint
    (which can be upto 30% more) although as I wrote there are fewer
    number of RTL insns.

    Another disadvantage of RTL insns *specifically* for Ruby is that
    insns for call sequences will be basically the same stack based
    ones but only bigger as they address the stack explicitly.

  * Secondly, I'd like to **combine some frequent insn sequences** into
    bigger insns.  Again it decreases insn dispatch overhead and
    memory traffic even more.  Also it permits to remove some type
    checking.

    The first thing on my mind is a sequence of a compare insn and a
    branch and using immediate operands besides temporary (stack) and
    local variables.  Also it is not a trivial task for Ruby as the
    compare can be implemented as a method.

  I already did some experiments.  RTL insns & combining insns permits
to speed the following micro-benchmark in more 2 times:

```
i = 0
while i<30_000_000 # benchmark loop 1
  i += 1
end
```

The generated RTL insns for the benchmark are

```
== disasm: #<ISeq:<main>@while.rb>======================================
== catch table
| catch type: break  st: 0007 ed: 0020 sp: 0000 cont: 0020
| catch type: next   st: 0007 ed: 0020 sp: 0000 cont: 0005
| catch type: redo   st: 0007 ed: 0020 sp: 0000 cont: 0007
|------------------------------------------------------------------------
local table (size: 2, temp: 1, argc: 0 [opts: 0, rest: -1, post: 0, block: -1, kw: -1@-1, kwrest: -1])
[ 2] i
0000 set_local_val    2, 0                                            (   1)
0003 jump             13                                              (   2)
0005 jump             13
0007 plusi            <callcache>, 2, 2, 1, -1                        (   3)
0013 btlti            7, <callcache>, -1, 2, 30000000, -1             (   2)
0020 local_ret        2, 0                                            (   3)
```

In this experiment I ignored trace insns (that is another story) and a
complication that a integer compare insn can be re-implemented as a
Ruby method.  Insn bflti is combination of LT immediate compare and
branch true.

A modification of fib benchmark is sped up in 1.35 times:

```
def fib_m n
  if n < 1
    1
  else
    fib_m(n-1) * fib_m(n-2)
  end
end

fib_m(40)
```

The RTL code of fib_m looks like

```
== disasm: #<ISeq:[e-mail]>==========================================
local table (size: 2, temp: 3, argc: 1 [opts: 0, rest: -1, post: 0, block: -1, kw: -1@-1, kwrest: -1])
[ 2] n<Arg>
0000 bflti            10, <callcache>, -1, 2, 1, -1                   (   2)
0007 val_ret          1, 16
0010 minusi           <callcache>, -2, 2, 1, -2                       (   5)
0016 simple_call_self <callinfo!mid:fib_m, argc:1, FCALL|ARGS_SIMPLE>, <callcache>, -1
0020 minusi           <callcache>, -3, 2, 2, -3
0026 simple_call_self <callinfo!mid:fib_m, argc:1, FCALL|ARGS_SIMPLE>, <callcache>, -2
0030 mult             <callcache>, -1, -1, -2, -1
0036 temp_ret         -1, 16
```

In reality, the improvement of most programs probably will be about
10%.  That is because of very dynamic nature of Ruby (a lot of calls,
checks for redefinition of basic type operations, checking overflows
to switch to GMP numbers).  For example, integer addition can not be
less than about x86-64 17 insns out of the current 50 insns on the
fast path.  So even if you make the rest (33) insns 2 times faster,
the improvement will be only 30%.

A very important part of MRI performance improvement is to make calls
fast because there are a lot of them in Ruby but as I read in some
Koichi Sasada's presentations he pays a lot of attention to it.  So I
don't want to touch it.

  * Thirdly.  I want to implement the insns as small inline functions
    for future AOT compiler, of course, if the projects described
    above are successful.  It will permit easy AOT generation of C code
    which will be basically calls of the functions.

    I'd like to implement AOT compiler which will generate a Ruby
    method code, call a C compiler to generate a binary shared code
    and load it into MRI for subsequent calls.  The key is to minimize
    the compilation time.  There are many approaches to do it but I
    don't want to discuss it right now.

    C generation is easy and most portable implementation of AOT but
    in future it is possible to use GCC JIT plugin or LLVM IR to
    decrease overhead of C scanner/parser.

    C compiler will see a bigger scope (all method insns) to do
    optimizations.  I think using AOT can give another 10%
    improvement.  It is not that big again because of dynamic nature
    of Ruby and any C compiler is not smart enough to figure out
    aliasing for typical generated C program.

    The life with the performance point of view would be easy if Ruby
    did not permit to redefine basic operations for basic types,
    e.g. plus for integer.  In this case we could evaluate types of
    operands and results using some data flow analysis and generate
    faster specialized insns.  Still a gradual typing if it is
    introduced in future versions of Ruby would help to generate such
    faster insns.

  Again I wrote this proposal for discussion as I don't want to be in
a position to compete with somebody else ongoing big project.  It
might be counterproductive for MRI development.  Especially I don't
want it because the project is big and long and probably will have a
lot of tehcnical obstacles and have a possibilty to be a failure.




-- 
https://bugs.ruby-lang.org/

Unsubscribe: <mailto:[e-mail]?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>