llvm dev - Jan 2013 - [LLVMdev] Using C++'11 language features in LLVM itself

Hello,

 I have been following this discussion for a while and I think the
question we should be asking is:

 Why do we want to even bother with all these other broken C++
compilers in a first place?

 Clang is good enough to bootstrap itself on practically
any platform I can think of or it can be cross-bootstrapped
if needed.

 I think usage of any language features should be decided
based on support from a either clang(n-j) release or a
reference version (3.2 being my personal favorite)
rather then on a common unbroken subset from a motley
collection of other compilers.

Just my 0.02$ after 2 months of building clang+llvm.

Paweł

> On Jan 9, 2013, at 2:38 AM, David Chisnall <David.Chisnall at
cl.cam.ac.uk> wrote:
> 
>> On 9 Jan 2013, at 04:49, Marc J. Driftmeyer wrote:
>> 
>>> It's not a coincidence that GCC 4.2.1 is the baseline on
FreeBSD considering the licensing of GPL restrictions on new releases.
>> 
>> [With my FreeBSD hat on]
>> 
>> Our plan for 10.0 is to ship clang only, with gcc 4.2.1 relegated to a
compat package for tier 1 architectures.  This should be x86, x86-64, and
ARMv6/7 (and maybe v8 if we're very lucky, but probably not).  MIPS and
PowerPC are slowly migrating to clang, but will probably take a little bit
longer (although given the progress that these are making, possibly not).  As
these are tier 2 architectures, I don't have a problem requiring an external
compiler or a cross compiler for the initial bootstrap build.
> 
> Ok, that's good news!
Kudos to FreeBSD!
> 
>> 
>> However, our release cycle is (in theory) 6 months, but is out of phase
with LLVM's and so at any given point we are likely to be one or two
releases behind.
>> 
>> As such, it's important to us to be able to build trunk clang with
clang from at least two releases ago, and ideally four.  Losing this ability
makes it very difficult for us to do the bootstrap toolchain build
Since Thursday, February 4, 2010 we can always do
clang(n-j) ... clang(n-1)clang(n) -> clang(n+1)
> 
> 
> Can you make this more concrete for me?  If we made this change in LLVM
3.3, what version of Clang would you need us to be compatible with?  Would it be
ok to require Clang 3.1 or later, or do you need clang 3.0 support?
> 
> -Chris

On Thu, Jan 10, 2013 at 1:11 AM, Pawel Wodnicki <root at 32bitmicro.com>
wrote:
> Hello,
>
>  I have been following this discussion for a while and I think the
> question we should be asking is:
>
>  Why do we want to even bother with all these other broken C++
> compilers in a first place?
>
Because not everyone that uses LLVM also uses Clang for every C++ compile.
 If someone wants to use LLVM as their optimizer/code generator for a new
language, why restrict them to using Clang for their entire project.  This
is especially crucial in HPC applications, where you may want to use LLVM
to compile a domain-specific language and the compiler is part of a larger
project that you want to compile using a vendor compiler for performance
reasons.

>
>  Clang is good enough to bootstrap itself on practically
> any platform I can think of or it can be cross-bootstrapped
> if needed.
>
Being able to bootstrap Clang on a platform is different from Clang being
100% compatible with a platform.  Clang can be built on Windows, but it's
far from usable as a system compiler on that platform.  Last time I
checked, Clang could not parse all of the SDK headers.  And forcing MinGW
isn't a solution either, not to mention the quagmire of incompatible
versions/distributions.  Please correct me if I am wrong, but I do not
believe you can build Clang with MSVC and then build Clang itself using the
Microsoft headers/libraries.  The C++ ABI support for MSVC in Clang is
still not 100% functional, I believe.

>
>  I think usage of any language features should be decided
> based on support from a either clang(n-j) release or a
> reference version (3.2 being my personal favorite)
> rather then on a common unbroken subset from a motley
> collection of other compilers.
>
While I would love to see this happen, I just don't think it's
practical/feasible currently.

>
> Just my 0.02$ after 2 months of building clang+llvm.
>
> Paweł
>
>
> > On Jan 9, 2013, at 2:38 AM, David Chisnall <David.Chisnall at
> cl.cam.ac.uk> wrote:
> >
> >> On 9 Jan 2013, at 04:49, Marc J. Driftmeyer wrote:
> >>
> >>> It's not a coincidence that GCC 4.2.1 is the baseline on
FreeBSD
> considering the licensing of GPL restrictions on new releases.
> >>
> >> [With my FreeBSD hat on]
> >>
> >> Our plan for 10.0 is to ship clang only, with gcc 4.2.1 relegated
to a
> compat package for tier 1 architectures.  This should be x86, x86-64, and
> ARMv6/7 (and maybe v8 if we're very lucky, but probably not).  MIPS and
> PowerPC are slowly migrating to clang, but will probably take a little bit
> longer (although given the progress that these are making, possibly not).
>  As these are tier 2 architectures, I don't have a problem requiring an
> external compiler or a cross compiler for the initial bootstrap build.
> >
> > Ok, that's good news!
>
> Kudos to FreeBSD!
>
> >
> >>
> >> However, our release cycle is (in theory) 6 months, but is out of
phase
> with LLVM's and so at any given point we are likely to be one or two
> releases behind.
> >>
> >> As such, it's important to us to be able to build trunk clang
with
> clang from at least two releases ago, and ideally four.  Losing this
> ability makes it very difficult for us to do the bootstrap toolchain build
>
> Since Thursday, February 4, 2010 we can always do
> clang(n-j) ... clang(n-1)clang(n) -> clang(n+1)
>
> >
> >
> > Can you make this more concrete for me?  If we made this change in
LLVM
> 3.3, what version of Clang would you need us to be compatible with?  Would
> it be ok to require Clang 3.1 or later, or do you need clang 3.0 support?
> >
> > -Chris
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>


-- 

Thanks,

Justin Holewinski
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130110/46b8b776/attachment.html>

On 1/10/2013 7:32 AM, Justin Holewinski wrote:
> On Thu, Jan 10, 2013 at 1:11 AM, Pawel Wodnicki <root at
32bitmicro.com> wrote:
> 
>> Hello,
>>
>>  I have been following this discussion for a while and I think the
>> question we should be asking is:
>>
>>  Why do we want to even bother with all these other broken C++
>> compilers in a first place?
>>
> 
> Because not everyone that uses LLVM also uses Clang for every C++ compile.
>  If someone wants to use LLVM as their optimizer/code generator for a new
> language, why restrict them to using Clang for their entire project.  This
> is especially crucial in HPC applications, where you may want to use LLVM
> to compile a domain-specific language and the compiler is part of a larger
> project that you want to compile using a vendor compiler for performance
> reasons.
> 
In LLVM+platform_C++_compiler scenario the safest way would be to use
no C++11 features in LLVM at all and just keep the status quo.
> 
>>
>>  Clang is good enough to bootstrap itself on practically
>> any platform I can think of or it can be cross-bootstrapped
>> if needed.
>>
> 
> Being able to bootstrap Clang on a platform is different from Clang being
> 100% compatible with a platform.  Clang can be built on Windows, but
it's
> far from usable as a system compiler on that platform.  Last time I
> checked, Clang could not parse all of the SDK headers.  And forcing MinGW
> isn't a solution either, not to mention the quagmire of incompatible
> versions/distributions.  Please correct me if I am wrong, but I do not
> believe you can build Clang with MSVC and then build Clang itself using the
> Microsoft headers/libraries.  The C++ ABI support for MSVC in Clang is
> still not 100% functional, I believe.
> 
 I do not think you can legally make clang a system compiler on
Windows without licensing headers/libraries from Microsoft.
So inability to parse all of the SDK headers is a non issue to me,
can't use them anyway.

 Seems the only sensible open source way to use clang on Windows
platform is to use either MinGW or Cygwin environment which works fine!
Clang can bootstrap itself on either of them using clang(n-1).

"quagmire of incompatible versions/distributions" - describes very
well the situation I would rather avoid by using clang.
> 
>>
>>  I think usage of any language features should be decided
>> based on support from a either clang(n-j) release or a
>> reference version (3.2 being my personal favorite)
>> rather then on a common unbroken subset from a motley
>> collection of other compilers.
>>
> 
> While I would love to see this happen, I just don't think it's
> practical/feasible currently.
> 
> 
At some point it is no longer a technical question as
all the effort spent on ensuring compatibility with
the other compilers could be redirected towards making
clang better!
>>
>> Just my 0.02$ after 2 months of building clang+llvm.
>>
>> Paweł
>>
>>
>>> On Jan 9, 2013, at 2:38 AM, David Chisnall <David.Chisnall at
>> cl.cam.ac.uk> wrote:
>>>
>>>> On 9 Jan 2013, at 04:49, Marc J. Driftmeyer wrote:
>>>>
>>>>> It's not a coincidence that GCC 4.2.1 is the baseline
on FreeBSD
>> considering the licensing of GPL restrictions on new releases.
>>>>
>>>> [With my FreeBSD hat on]
>>>>
>>>> Our plan for 10.0 is to ship clang only, with gcc 4.2.1
relegated to a
>> compat package for tier 1 architectures.  This should be x86, x86-64,
and
>> ARMv6/7 (and maybe v8 if we're very lucky, but probably not).  MIPS
and
>> PowerPC are slowly migrating to clang, but will probably take a little
bit
>> longer (although given the progress that these are making, possibly
not).
>>  As these are tier 2 architectures, I don't have a problem
requiring an
>> external compiler or a cross compiler for the initial bootstrap build.
>>>
>>> Ok, that's good news!
>>
>> Kudos to FreeBSD!
>>
>>>
>>>>
>>>> However, our release cycle is (in theory) 6 months, but is out
of phase
>> with LLVM's and so at any given point we are likely to be one or
two
>> releases behind.
>>>>
>>>> As such, it's important to us to be able to build trunk
clang with
>> clang from at least two releases ago, and ideally four.  Losing this
>> ability makes it very difficult for us to do the bootstrap toolchain
build
>>
>> Since Thursday, February 4, 2010 we can always do
>> clang(n-j) ... clang(n-1)clang(n) -> clang(n+1)
>>
>>>
>>>
>>> Can you make this more concrete for me?  If we made this change in
LLVM
>> 3.3, what version of Clang would you need us to be compatible with? 
Would
>> it be ok to require Clang 3.1 or later, or do you need clang 3.0
support?
>>>
>>> -Chris
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>
> 
> 
>

Pawel Wodnicki <root at 32bitmicro.com> writes:
>  Clang is good enough to bootstrap itself on practically
> any platform I can think of or it can be cross-bootstrapped
> if needed.
You're completely ignoring communities that compile code outside
llvm+clang.  Any compiler chosen by that group has to be able to
correctly compile hundres if not thousands of source code files.  It is
not easy to find such a compiler and that makes it an arduous process to
upgrade.  Of course upgrades are done, but they're done conservatively.

                            -David

On 1/10/2013 12:11 PM, dag at cray.com wrote:
> Pawel Wodnicki <root at 32bitmicro.com> writes:
> 
>>  Clang is good enough to bootstrap itself on practically
>> any platform I can think of or it can be cross-bootstrapped
>> if needed.
> 
> You're completely ignoring communities that compile code outside
> llvm+clang.  Any compiler chosen by that group has to be able to
> correctly compile hundres if not thousands of source code files.  It is
> not easy to find such a compiler and that makes it an arduous process to
> upgrade.  Of course upgrades are done, but they're done conservatively.
> 
>                             -David
> 
> 
 As you have pointed out the reality is that for a long time
there will be a lot of different C++ code to maintain and for
practical reasons it makes no sens to throw away time tested tools
just to use some modern C++ style.  At the same time I am not
convinced that it is worth the effort to introduce some C++11
features into LLVM and try to stay compatible with all these other
compilers.

 So if the goal is to be able to use *any* however broken C++ compiler
to build LLVM and stay compatible with the platform compiler
then perhaps we should approach it differently. Ideally source
translation in clang could be used to "lower" C++11 code to
either C++03 or even C++98, worst case just use CPP backend.

 I realize that Cfront approach does not have a good
track record but this could be a C++front translating
C++11 to older version of the C++ language.

Paweł

On Jan 10, 2013, at 11:11 AM, <dag at cray.com>
 wrote:
> Pawel Wodnicki <root at 32bitmicro.com> writes:
> 
>> Clang is good enough to bootstrap itself on practically
>> any platform I can think of or it can be cross-bootstrapped
>> if needed.
> 
> You're completely ignoring communities that compile code outside
> llvm+clang.  Any compiler chosen by that group has to be able to
> correctly compile hundres if not thousands of source code files.  It is
> not easy to find such a compiler and that makes it an arduous process to
> upgrade.  Of course upgrades are done, but they're done conservatively.
> 
You don't have to select one compiler.  We've got many production
compilers installed (Intel, PGI, Cray, numerous versions of  GCC),  a number of
official  `friendly' versions maintained, and dozens of unofficial builds
(e.g. clang).

Given the fast flux of the LLVM interfaces, getting hung-up on the lowest common
denominator for a C++ compiler seems weird to me.

Marcus