Re: isl 0.26 C++17 issues when bootstrapping with GCC

[Iain Sandoe <iain@sandoe.co.uk>]

> On 13 May 2024, at 18:46, Richard Biener <richard.guenther@gmail.com> wrote:
> 
> On Mon, May 13, 2024 at 6:00 PM Iain Sandoe <iain@sandoe.co.uk> wrote:
>> 
>> 
>> 
>>> On 13 May 2024, at 16:05, Iain Sandoe via Gcc <gcc@gcc.gnu.org> wrote:
>>>> On 30 Aug 2023, at 00:32, Ben Boeckel via Gcc <gcc@gcc.gnu.org> wrote:
>>>> 
>>>> On Tue, Aug 29, 2023 at 18:57:37 +0200, Richard Biener wrote:
>>>>> I suppose for bootstrapping we could disable ISL during stage1 since
>>>>> it enables an optional feature only.  Other than that GCC only
>>>>> requires a C++11 compiler for building, so ISL breaks that constraint
>>>>> with requiring C++17.
>>>> 
>>>> Note that it doesn't *require* it per sé; the tests that try it are
>>>> compiled if C++17 support was detected at all. The headers seem to just
>>>> have optional header-only `std::any`-using APIs if C++17 is around.
>>>> `isl` supporting a flag to disable the tests would also work, but that
>>>> doesn't fix 0.26. It also doesn't mean it won't start requiring C++17 at
>>>> some point in the future.
>>> 
>>> Perhaps, in the short-term (i.e. before it requires C++ > 11) we can
>>> solve this by ensuring that we pass -std=c++11 to the configure stages
>>> as well as to the build.  ISTM that configure is finding C++17-capability
>>> (because we do not, I think, force C++11 for the configure) and then
>>> the build takes it away by forcing -std=c++11.
>> 
>> That was not right.
>> We add std=c++11 to the compiler command.
>> 
>> However,as noted (earlier in this thread) the isl configure has the idiom
>> - does the compiler do c++17 with no options?
>> - does the compiler do c++17 if we add -std=c++17?
>> 
>> where the second one overrides our setting of std=c++11 in the compiler
>> comand.
>> 
>> (I think that this is a reasonably often used idiom in configures)
>> 
>> However the isl configure _does_ still append CXXFLAGS, and so that if
>> we add -std=c++11 to those, it re-asserts our intent.
>> 
>> Maybe we should just add the -std=c++11 to CXXFLAGS instead of the
>> compiler command?
> 
> I don't understand.  If we set CXX to g++ -std=c++11 and ISL checks
> for -std=c++17 why does it then fail to add that to CXXFLAGS?

This appears to be the underlying bug.

— isl configure.ac does:

AX_CXX_COMPILE_STDCXX_17([], [optional])

….

AM_CONDITIONAL(HAVE_CXX17, test "x$HAVE_CXX17" = "x1”)

— and then Makefile.am adds the c++17-requirements:

if HAVE_CXX17
  noinst_PROGRAMS += isl_test_cpp17 isl_test_cpp17-checked
  TESTS += isl_test_cpp17 isl_test_cpp17-checked
endif

.. this mechanism does not seem to preserve the fact that an additional
-std=c++17 was needed to get the CXX17 (and there is no mention of
CXXFLAGS in Makefile.am) 

Not sure if it’s a bug in isl’s config - or a limitation of 
AX_CXX_COMPILE_STDCXX_17 itself.

— assuming we file a bug and it gets agreed and fixed, we’ll still need 
either to skip broken versions or work around it (I have no specific
preference - although I do build isl in-tree, so far 0.24 has been OK).

Iain


> 
>> Iain
>> 
>>>> In light of that, I feel that skipping it for bootstrap is probably the
>>>> right solution here. Alas, my skill with autotools is closer to the
>>>> caveman-with-club level rather than that of a surgeon.
>>> 
>>> I am not sure we have an easy way to exclude a host module from
>>> stage1 only (but ICBW).
>>> 
>>> Iain
>>> 
>>>> 
>>>> --Ben