plans for 64bit support

[Daniel Oberhoff]

Am 11.12.2007 um 09:50 schrieb Ryan Schmidt:

>
> On Dec 11, 2007, at 02:35, Daniel Oberhoff wrote:
>
>> Am 10.12.2007 um 01:33 schrieb Ryan Schmidt:
>>
>>> On Dec 9, 2007, at 14:43, Daniel Oberhoff wrote:
>>>
>>>> Now that Leopard is out and already at 10.5.1 will macports be  
>>>> supporting 64bit libraries? It's just I need 64bit in my octave  
>>>> installation. I pull my octave from octave.org's cvs, but it  
>>>> needs quite a lot of support libraries. From what I gather it  
>>>> should be possible on Leopard to build fat libraries, i.e. ones  
>>>> that contain 64 and 32 bit code (i think it works using -arch  
>>>> x84_64 -arch i686 as gcc flags). Or will this be left to the  
>>>> separate ports?
>>>
>>> MacPorts is supposed to build libraries for whatever system it's  
>>> running on. So I would have thought that if you're on a 64-bit  
>>> Intel system, it should build 64-bit Intel libraries. Is it  
>>> building 32-bit Intel libraries for you?
>>>
>>> We have the +universal variant for building 2-way (32-bit, Intel  
>>> and PowerPC) universal binaries. We are still in the process of  
>>> getting this to work with many of the ports. It could be changed  
>>> to build 4-way (32-bit and 64-bit, Intel and PowerPC) universal  
>>> binaries. This should be possible on Tiger too, as far as I know.  
>>> It won't fix any ports that are having trouble building 2-way  
>>> universal binaries. Not sure if it would mess up any ports that  
>>> are already working. Are all the ports that you need already  
>>> working as 2-way universal binaries?
>>>
>>> I haven't heard anyone suggest building libraries that contain 32- 
>>> bit and 64-bit code for just one processor family before (in  
>>> relation to MacPorts). It would of course be possible, but I think  
>>> it would make most sense to continue along our current path:  
>>> software should be default install for the architecture you're on,  
>>> and if you need multiple architectures, then you need the  
>>> +universal variant.
>>>
>>> It has been said before that maybe 64-bit binaries aren't all that  
>>> helpful, but the Ars Technica review of Leopard explains that  
>>> while 64-bit binaries aren't that helpful on the PowerPC  
>>> architecture, they really are quite good for secondary reasons on  
>>> the Intel architecture. The 32-bit Intel architecture has often  
>>> been called inferior to the 32-bit PowerPC architecture, but the  
>>> 64-bit Intel architecture seems to fix many of the issues. Also,  
>>> maybe Leopard being a full 64-bit system makes 64-bit binaries  
>>> more relevant.
>>>
>>> Perhaps we could do 4-way universal binaries only when MacPorts is  
>>> running on Leopard.... but that might be a bad idea, since only  
>>> people running Leopard could  then develop and test this.
>>>
>>> We could introduce a new automatic variant... +universal4?  
>>> +universal64? People could test with this new variant and if any  
>>> problems are encountered it would not prevent anyone from using  
>>> the existing 2-way 32-bit +universal variant. I'm wary of this  
>>> though... I wouldn't want, say, "universal64" directives to start  
>>> appearing in portfiles, if we want to eventually fold +universal64  
>>> into +universal.
>>>
>>> Just some thoughts off the top of my head.
>>
>> Hmm, ok. That makes sense, since I upgraded with an installed  
>> macports base. I suppose I should really reinstall all then, since  
>> otherwise I will end up with a crazy mix?
>
> Yes, if you're moving from one architecture to another (G4->G5, PPC- 
> >Intel, 32-bit->64-bit, etc.) you should really throw away your old  
> MacPorts (saving any conf files or databases or other data) and  
> rebuild all your ports.
>
>> Though, per default on Leopard compilation seems to be 32 bit. When  
>> you just compile something without further options pointer size is  
>> 32bit. Only when you specify -arch x86_64 or -m64 do you get 64bit.
>
> Ah, I see. I didn't know. I haven't had a 64-bit machine.
>
>> So it would really make sense to have both (but not ppc). So I  
>> would vote for an autamatic variant to get intel only 32/64bit.
>
> So by default we build 32-bit binaries for the host architecture.  
> And we have a +universal variant for building 32-bit binaries for  
> both architectures. You want an additional variant for building fat  
> 32-bit and 64-bit Intel binaries. Do we then also need a separate  
> variant for building fat 32-bit and 64-bit PowerPC binaries? And  
> what about a fat universal variant with all 4 options? This  
> proliferation of variants is what I want to avoid. Most users will  
> not need it and it will confuse some of them. It's much extra work  
> for some portfile authors, because some software (rather a lot,  
> actually) does not build successful universal binaries using the  
> simple tricks, and much more work is required to build universal.  
> See for example the openssl port (which, to build universal, you'll  
> have to back-date to version 0.9.8e because of bug #12911) which  
> goes through crazy hoops to build first for one architecture, then  
> the other, then lipo them together. Similar (though perhaps not  
> identical) hoops would have to be gone through for fat binaries or 4- 
> way binaries. Well, that is until MacPorts base grows some universal- 
> build helpers. (I still really think MacPorts base should have an  
> easy way to build a port multiple times, once for each architecture,  
> and lipo it together, automatically, selected with just a single  
> line in the portfile).
>

Yah, that doesn't sound fun. Would it be feasible to make these  
incremental? After all the mechanism to combine architectures is  
always the same (use -arch flags, lipo...). I.e. make it possible to  
say +64 +32 +ppc +intel to get the 4way build or just +ppc +intel to  
get a two way build. Or at least have something like that behind the  
scenes, and just offer specializations if some of the variants need  
some special care.

Alternatively maybe have two macports trees, one with 64bit, one with  
32bit. Kinda like those lib64 dirs on some linux flavours. Is that  
possible right now?

>
> Why, by the way, do you need 64-bit octave? Just curious.
>

Various reasons:

1) I do get data clumps in the gigabyte range sometimes. Yes, it does  
lead to some paging, but that's better than a crash with out of  
memory :).
2) this is more of a hack: I use Judy for some sparse double precision  
structs, and since Judy constrains data size to sizeof(void*) I need  
64 bit.
3) afaik 64bit has more registers than 32bit on intel architecture,  
avoiding register spillage, which sometimes snails some of my code.

all in all, it is not critical right now. But every now and then 64bit  
would help a lot. Maybe not always though. Also due to the policy on  
Leopard to default to 32bit would probably make it hard to make do  
with an exclusive 64bit set of libraries.

Hmm, thinking of that: is it possible to convince port _not_ to switch  
over to 64bit now? I'd rather have a clean set of 32 bit libs :).

Best

Daniel