gamedevlists-windows

[GD-Windows] VC8.0 and CMOV

[Alen L. <ale...@cr...>]

Hi all,

Is there any way to make MSVC8.0 to generate CMOV instructions? I have
some performance-critical loops, with a lot of small ifs, where it
generates branches, and I'd like it to put cmov/fcmov there instead.
Allegedly, it is able to do so, but /O2 doesn't seem to help. I can't
even find any intrinsic to do that. Any ideas appreciated.

Thanks,
Alen

Re: [GD-Windows] VC8.0 and CMOV

[Andrew F. <pb...@gm...>]

Does using "__asm" help?  You could hand-code the instructions, then.
confer:  <http://msdn2.microsoft.com/EN-US/library/45yd4tzz(VS.80).aspx>,
which is the VC8.0 specific page about using __asm anywhere a
statement could appear.

It could be that the scheduler doesn't know that you are guaranteeing
that the program will be run on a Pentium Pro or better, system.
(Pentium Pro is when that instruction was added.)  Of course, now that
I go to look for a reference to this on the microsoft web site, this
may not be a concern.  I'm showing my age. :)

--andy

On Nov 23, 2007 4:45 AM, Alen Ladavac <ale...@cr...> wrote:
> Hi all,
>
> Is there any way to make MSVC8.0 to generate CMOV instructions? I have
> some performance-critical loops, with a lot of small ifs, where it
> generates branches, and I'd like it to put cmov/fcmov there instead.
> Allegedly, it is able to do so, but /O2 doesn't seem to help. I can't
> even find any intrinsic to do that. Any ideas appreciated.
>
> Thanks,
> Alen
>
>
> -------------------------------------------------------------------------
> This SF.net email is sponsored by: Microsoft
> Defy all challenges. Microsoft(R) Visual Studio 2005.
> http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
> _______________________________________________
> Gamedevlists-windows mailing list
> Gam...@li...
> https://lists.sourceforge.net/lists/listinfo/gamedevlists-windows
> Archives:
> http://sourceforge.net/mailarchive/forum.php?forum_id=555
>

Re: [GD-Windows] VC8.0 and CMOV

[Alen L. <ale...@cr...>]

Well, yes I tried that. Creating an inline function with _asm to
compare two numbers does work, but it seems to mess up with the
compiler's optimizations of the rest of the function, because I can't
input and output floats or ALU flags in any other way than through
memory.

Regarding the architecture, I'm optimizing this with /arch:SSE which,
according to MSDN, implies using CMOV as well, but the complier never
uses that instruction by itself.

(BTw, sorry if this seems a bit off-topic, but google doesn't help,
which seems like people usually don't really care. I hoped someone in
the gamedev would already have experienced this.)

Thanks,
Alen


Andrew wrote:
> Does using "__asm" help?  You could hand-code the instructions, then.
> confer: 
> <http://msdn2.microsoft.com/EN-US/library/45yd4tzz(VS.80).aspx>,
> which is the VC8.0 specific page about using __asm anywhere a
> statement could appear.

> It could be that the scheduler doesn't know that you are guaranteeing
> that the program will be run on a Pentium Pro or better, system.
> (Pentium Pro is when that instruction was added.)  Of course, now that
> I go to look for a reference to this on the microsoft web site, this
> may not be a concern.  I'm showing my age.  

> --andy

> On Nov 23, 2007 4:45 AM, Alen Ladavac <ale...@cr...> wrote:
>> Hi all,
>>
>> Is there any way to make MSVC8.0 to generate CMOV instructions? I have
>> some performance-critical loops, with a lot of small ifs, where it
>> generates branches, and I'd like it to put cmov/fcmov there instead.
>> Allegedly, it is able to do so, but /O2 doesn't seem to help. I can't
>> even find any intrinsic to do that. Any ideas appreciated.
>>
>> Thanks,
>> Alen
>>
>>
>> -------------------------------------------------------------------------
>> This SF.net email is sponsored by: Microsoft
>> Defy all challenges. Microsoft(R) Visual Studio 2005.
>> http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
>> _______________________________________________
>> Gamedevlists-windows mailing list
>> Gam...@li...
>> https://lists.sourceforge.net/lists/listinfo/gamedevlists-windows
>> Archives:
>> http://sourceforge.net/mailarchive/forum.php?forum_id=555
>>




-- 
Alen

Re: [GD-Windows] VC8.0 and CMOV

[Jan W. <jan...@gm...>]

Hello,

On Fri, 23 Nov 2007 11:45:21 +0100, Alen Ladavac <ale...@cr...> =
 =

wrote:
> Is there any way to make MSVC8.0 to generate CMOV instructions?
> Allegedly, it is able to do so,
Confirmed, I've seen them in VC2005 SP1's output. Not sure whether that =
 =

was due to hand-generated snippets being pasted in by the "optimizer" or=
  =

because the code generator has been improved.
Even after quite some experimentation, no reliable way to provoke CMOV w=
as  =

found.

One thing that did work (surprisingly enough) is a carry flag-based tric=
k:
int mask =3D (condition)? ~0 : 0;
This actually works out to a SBB, which can be used to select one of two=
  =

values or increment/decrement an index, etc. Whether this is useful  =

depends on the exact contents of your if() statements.

> but /O2 doesn't seem to help
Code Generation -> Enhanced Instruction Set will probably be necessary a=
s  =

well.


On Fri, 23 Nov 2007 16:13:21 +0100, Andrew Finkenstadt <pb...@gm...=
>  =

wrote:
> Does using "__asm" help?  You could hand-code the instructions, then.
Unfortunately this approach suffers from VC's inability to  =

peephole-optimize the 'boundaries' of asm code. You'll see back-to-back =
 =

stores and loads of variables you pass to/from asm, which often eats awa=
y  =

any gains attained. CPUs have forwarding hardware for this, but it still=
  =

hurts.

HTH+HAND
Jan Wassenberg