svn commit: r249484 - head/lib

[Tim Kientzle]

On Apr 17, 2013, at 5:05 AM, Tijl Coosemans wrote:

> On 2013-04-17 08:26, Tim Kientzle wrote:
>> On Apr 16, 2013, at 11:06 PM, Juli Mallett wrote:
>>> On Tue, Apr 16, 2013 at 11:00 PM, Tim Kientzle wrote:
>>>> On Apr 16, 2013, at 6:41 AM, Tijl Coosemans wrote:
>>>>> On 2013-04-14 21:13, Tim Kientzle wrote:
>>>>>> Modified: head/lib/Makefile
>>>>>> ==============================================================================
>>>>>> --- head/lib/Makefile        Sun Apr 14 18:36:30 2013        (r249483)
>>>>>> +++ head/lib/Makefile        Sun Apr 14 19:13:51 2013        (r249484)
>>>>>> @@ -252,4 +252,7 @@ _libusbhid=      libusbhid
>>>>>> _libusb=     libusb
>>>>>> .endif
>>>>>> 
>>>>>> +afterinstall:
>>>>>> +    ln -fs ../include ${DESTDIR}/usr/lib/include
>>>>>> +
>>>>>> .include <bsd.subdir.mk>
>>>>> 
>>>>> This breaks with -DNO_CLEAN defined, because then
>>>>> ${DESTDIR}/usr/lib/include/include is created.
>>>> 
>>>> That's a good point.  Would this work better?
>>>> 
>>>> afterinstall:
>>>>    if [ ! -e $(DESTDIR)/usr/lib/include ]; then
>>>>       ln -fs ../include $(DESTDIR)/usr/lib/include
>>>>    fi
> 
> Maybe just: ln -fs ../include $(DESTDIR)/usr/lib/

Ah, yes.  That is the obvious answer.   I'll change it to
that for now.  I'm happy to remove it once we have a
better answer in place.

>>>>> I'm not that fond of this patch by the way, but I don't fully
>>>>> understand the problem it's trying to solve so I won't object.
>>>>> It just looks too much like a hack to me
>>>> 
>>>> It's a subtle issue and I'm not surprised that it raised some
>>>> eyebrows.  I spent a long time looking for a better solution.
>>>> 
>>>> In short, both GCC and Clang make some assumptions
>>>> about the layout of headers used for freestanding compiles.
>>>> (My earlier commit said these assumptions were "undocumented",
>>>> but that's not quite true, they're just rather obscure.)
>>> 
>>> If you're doing a freestanding compile...shouldn't you also be
>>> specifying both include and library paths explicitly?
>> 
>> Yes, of course.  But the correct directories to use vary somewhat
>> across platforms, so we would like to have some reasonably
>> portable way to find the right directory to use for building on
>> a particular system.
>> 
>> Both gcc and clang support a -print-file-name=include option which
>> is supposed to print out the directory containing headers used
>> for freestanding compiles.  You can then take that path and
>> use it as the explicit include directory path for freestanding builds.
>> 
>>> (Or even better, if you're doing a freestanding
>>> compile, but want the default include paths, get the compiler to dump
>>> the default include paths and process that.)
>> 
>> That's precisely what this is for.  I've been working with U-Boot
>> sources which compile on many systems and use
>> -print-file-name=include to identify the directory containing
>> the basic freestanding header files.
> 
> So you compile with -ffreestanding -nostdinc?
> And then add the include path returned by -print-file-name=include?

That's what the U-Boot sources do, yes.

>> The -print-file-name=include option works on Linux, works
>> on MacOS, and --- with this one symlink --- can work on
>> FreeBSD as well.  I've been using it to cross-build U-Boot
>> using the FreeBSD xdev toolchain with both GCC and Clang.
> 
> "clang -E -v - </dev/null" shows it passes "-resource-dir
> /usr/bin/../lib/clang/3.3" to cc1 stage which then complains about
> nonexistent directory "/usr/bin/../lib/clang/3.3/include".
> 
> So how about moving /usr/include/clang/3.3 to
> /usr/lib/clang/3.3/include? That seems to be the location clang
> expects and what lang/clang port uses (in /usr/local).

I would certainly like to see that.  I presumed that there
was some reason this wasn't done in the initial import.

> The path from -resource-dir is also searched by -print-file-name.
> 
> All headers from contrib/llvm/tools/clang/lib/Headers would have
> to be installed there to have a complete freestanding environment,
> but some of those headers would have to be patched to use the base
> system header in the hosted case like the stdint.h header does:
> 
> #if __STDC_HOSTED__ && \
>    defined(__has_include_next) && __has_include_next(<stdint.h>)
> # include_next <stdint.h>
> #else
> ...
> #endif
> 
> In the lang/clang port files/patch-tools_clang_lib_Headers_Makefile
> should be removed I think. It prevents too many useful headers from
> being installed (e.g. avxintrin.h)

That would be great!

I can certainly help with some of this but my time is a
little tight.

While we're talking about freestanding brokenness, is anyone
interested in fixing the fact that FreeBSD/ARM requires freestanding
programs to be linked against libc?

Tim

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 495 bytes
Desc: Message signed with OpenPGP using GPGMail
URL: <http://lists.freebsd.org/pipermail/svn-src-all/attachments/20130417/01b6ed01/attachment.sig>