Re: [Qemu-devel] [PATCH] target-arm: use clz32() instead of a for loop

[Aurelien Jarno <aurelien@aurel32.net>]

Stuart Brady a écrit :
> On Fri, Oct 23, 2009 at 09:04:53AM +0200, Aurelien Jarno wrote:
>> Stuart Brady a écrit :
>>> Just a quick note that the implementation of clz, ctz and popcnt is
>>> still listed in the TCG TODO list.  The last time I looked, I noticed
>>> that quite a few architectures have clz/ctz instructions:
>>>
>>>    http://lkml.indiana.edu/hypermail/linux/kernel/0601.3/1683.html
>> OTOH, a dump shows that those instruction are not used than often, so I
>> am not sure it worth implementing it.
> 
> Really?  I'm surprised, as I gather that optimised ffs/fls/hweight
> functions in the kernel do give a modest gain...  I suppose I'll have
> to try it on several different targets and see! :-)

I gave a quick look at MIPS, and at least here, it is used often.

>>> For those that don't, I think a combination the following two hacks at
>>> http://graphics.stanford.edu/~seander/bithacks.html could be used:
>> The best is probably to use an helper in that case, calling clz32(x).
> 
> Yes, you're right.
> 
> There are several other places that should also call clz32()/ctz32().
> The ones that I can see are helper_neon_cls_s32() for ARM, helper_bsf()
> and helper_bsr() for X86, helper_ff1() for M68K.  (I'm not sure about
> 'do_clz8' and 'do_clz16', though.)
> 
> At some point, possibly next weekend, I'll submit patches to add clz
> and ctz helpers to tcg-runtime.c, and to convert Alpha, ARM, CRIS, M68K,
> MIPS, PowerPC and x86 (any others I've missed?) to use those helpers.

The main problem I see for a TCG implementation is the definition of
clz/ctz. Some targets define that clz(0) or ctz(0) returns 32, some
other define it as being "undefined".

If we go for the common denominator for the TCG op, that is clz(0) =
undefined, it means that a test with brcond has to be added in the
targets using clz(0) = 32, and this is likely to give more slow down
than speed gain.

If we go for clz(0) = 32, it means the test has to be implemented in
TCG, which might be complicated for some hosts.

-- 
Aurelien Jarno                          GPG: 1024D/F1BCDB73
aurelien@aurel32.net                 http://www.aurel32.net