[ruby-core:83422] [Ruby trunk Feature#12589] VM performance improvement proposal

Issue #12589 has been updated by k0kubun (Takashi Kokubun).


> In any case, if we are going to do some optimizations by ourself (and I see such necessity in the future) not only by GCC/LLVM, we need a convenient IR for this. 

Yeah, I saw rtl_exec.c transforms ISeq dynamically and allows MJIT to have insn that can be inlined easily. I can imagine the same idea will work on stack->RTL->JIT version of your MJIT. For our shared goal (stack->JIT), I agree that having any IR might be a helpful tool.

> Tagging doubles manipulates with exponent and mantissa by constraining exponent range and using a part of exponent field to store few less significant bits of mantissa. Even worse, processing 0.0 makes it even more complicated. Optimizing compilers are not smart enough to see that untaging doubles is a reverse operation to tagging and vice versa.
>
> One simple case where we can avoid untagging is RTL insn with immediate operand (we can use double not VALUE for the immediate operand).

That makes sense. If compiler can't do that with inlined code, we need to do in MJIT level and it would decrease effort in some level if insns are RTL.

> https://github.com/dino-lang/dino/blob/master/DINO/d_inference.c is an example how a basic type inference can be implemented on RTL-like language. It is a different approach to algorithm W in Hindley–Milner type system. The algorithm consists of the following steps
>
> Building a control flow graph (CFG) consisting of basic blocks and control flow edges connecting them.
Calculating available results of RTL instructions – this is a forward data-flow problem on the CFG.
Using the availability information, building def-use chains connecting possible operands and results of RTL instructions and variables.
> Calculating the types of RTL instruction operands and results – this is a forward data-flow problem on the def-use graph.
> The definition of availability and def-use chains can be found practically in any book about optimizing compilers.

Thank you for pointing it out and summary. I'll take a look.

----------------------------------------
Feature #12589: VM performance improvement proposal
https://bugs.ruby-lang.org/issues/12589#change-67389

* 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>