[Vm-dev] Interpreter>>isContextHeader: optimization

[Igor Stasenko]

2009/2/21 Eliot Miranda <eliot.miranda at gmail.com>:
>
> Hi Igor,
>
> On Fri, Feb 20, 2009 at 11:37 PM, Igor Stasenko <siguctua at gmail.com> wrote:
>>
>> Here the method:
>>
>> isContextHeader: aHeader
>>        self inline: true.
>>        ^ ((aHeader >> 12) bitAnd: 16r1F) = 13                  "MethodContext"
>>                or: [((aHeader >> 12) bitAnd: 16r1F) = 14               "BlockContext"
>>                or: [((aHeader >> 12) bitAnd: 16r1F) = 4]]      "PseudoContext"
>>
>> i think it wouldn't hurt to rewrite it as:
>>
>> isContextHeader: aHeader
>>        self inline: true.
>>  | hdr |
>>  hdr := aHeader bitAnd: (16r1F << 12).
>>        ^ hdr = (13 << 12)                      "MethodContext"
>>                or: [ hdr = (14 << 12)          "BlockContext"
>>                or: [ hdr = (4 << 12)]]  "PseudoContext"
>>
>> which will allow GCC to optimize it more easily.
>> I'm not sure if it can optimize it in its current state.
>> This may impact a small speedup of copy operations and any other
>> operations which need to determine a number of pointer fields in
>> object (users of #lastPointerOf:)
>
>
> First you should look at the assembly that gcc generates to be sure anything is needed.  e.g.
> cat >t.c <<END
> long isContext(long aHeader) {
>     return ((aHeader >> 12) & 0x1F) == 13
>         || ((aHeader >> 12) & 0x1F) == 14
>         || ((aHeader >> 12) & 0x1F) == 4;
> }
> END
> gcc -O3 -S -fomit-frame-pointer t.c; cat t.s
>     .text
> .globl _isContext
> _isContext:
>     movl    4(%esp), %edx
>     sarl    $12, %edx
>     andl    $31, %edx
>     leal    -13(%edx), %eax
>     cmpl    $1, %eax
>     jbe L2
>     cmpl    $4, %edx
>     je  L2
>     xorl    %eax, %eax
>     ret
> L2:
>     movl    $1, %eax
>     ret
>     .subsections_via_symbols
>
> So you don't need to do anything; it has done everything for you.
> However, one point is important.  Using 16r1F << 12 et al as your masks and constants to compare against is much worse on many systems, most importantly x86, than shifting down by 12 and comparing against small constants, because the instruction set can encode small constants far more compactly, and that means better code density in the icache which is significant for performance.  e.g. on x86 a constant in the range -128 to 127 typically takes a byte whereas anything else will take 4.
> But what I really think is that this is too low a level to worry about.  Much more important to focus on
> - context to stack mapping
> - in-line cacheing via a JIT
> - exploiting multicore via Hydra
> and beyond (e.g. speculative inlining)
> than worrying about tiny micro-optimizations like this :)

Thanks Eliot.
In fact, this method drawn my attention because of its number of
checks. Typically, all code which dealing with object formats contain
many branches. And from places where this method is called, there are
additional checks surrounding it.
So, the real problem is the overwhelming number of checks to do
something, and i think this having own impacts on performance.
I hope that a new object format with 64 bit header, which you plan to
use, will allow us to avoid so many branches in code which having high
usage frequency.

> Best
> Eliot
>>
>>
>> --
>> Best regards,
>> Igor Stasenko AKA sig.
>


-- 
Best regards,
Igor Stasenko AKA sig.