[PATCH] Enhance _GLIBCXX_DEBUG constexpr support

François Dumont frs.dumont@gmail.com
Mon Dec 9 08:53:00 GMT 2019


After completing execution of all tests I had to fix implementation of 
__check_singular to get only 1 return statement.



On 12/2/19 8:31 PM, François Dumont wrote:
> Hi
>
>     Here is a patch to enhance constexpr support in _GLIBCXX_DEBUG. I 
> work on std::lower_bound/upper_bound to find out if Debug mode is well 
> prepared. I'll continue on other algos later.
>
>     I initially hope that I could count on the compiler for the 
> valid_range check. But for lower_bound/upper_bound there is no 
> constexpr dedicated implementation like for copy/copy_backward so it 
> implies changing the existing implementation. So I tried to change the 
> while(__len > 0) into a while(__len != 0) and it gave:
>
> constexpr_valid_range_neg.cc:35:20: error: non-constant condition for 
> static assertion
>    35 | static_assert(test1()); // { dg-error "" }
>       |               ~~~~~^~
> In file included from 
> /home/fdt/dev/gcc/install/include/c++/10.0.0/algorithm:61,
>                  from constexpr_valid_range_neg.cc:22:
> constexpr_valid_range_neg.cc:35:20:   in ‘constexpr’ expansion of 
> ‘test1()’
> constexpr_valid_range_neg.cc:30:38:   in ‘constexpr’ expansion of 
> ‘std::lower_bound<const int*, int>(ca0.std::array<int, 12>::end(), 
> ca0.std::array<int, 12>::begin(), 6)’
> /home/fdt/dev/gcc/install/include/c++/10.0.0/bits/stl_algobase.h:1484:32: 
> in ‘constexpr’ expansion of ‘std::__lower_bound<const int*, int, 
> __gnu_cxx::__ops::_Iter_less_val>(__first, __last, (* & __val), 
> __gnu_cxx::__ops::__iter_less_val())’
> /home/fdt/dev/gcc/install/include/c++/10.0.0/bits/stl_algobase.h:1444:7: 
> error: ‘constexpr’ loop iteration count exceeds limit of 262144 (use 
> ‘-fconstexpr-loop-limit=’ to increase the limit)
>  1444 |       while (__len != 0)
>       |       ^~~~~
>
>     It seems ok but it isn't. The compiler had to loop 262144 times to 
> eventually give this status which is not even clear about the fact 
> that begin/end has been inverted. It is a quite heavy operation for a 
> limited result.
>
>     So this patch rather enable _GLIBCXX_DEBUG valid_range check which 
> gives:
>
> constexpr_valid_range_neg.cc:35:20: error: non-constant condition for 
> static assertion
>    35 | static_assert(test1()); // { dg-error "" }
>       |               ~~~~~^~
> In file included from 
> /home/fdt/dev/gcc/install/include/c++/10.0.0/debug/debug.h:90,
>                  from 
> /home/fdt/dev/gcc/install/include/c++/10.0.0/bits/stl_algobase.h:69,
>                  from 
> /home/fdt/dev/gcc/install/include/c++/10.0.0/algorithm:61,
>                  from constexpr_valid_range_neg.cc:22:
> constexpr_valid_range_neg.cc:35:20:   in ‘constexpr’ expansion of 
> ‘test1()’
> constexpr_valid_range_neg.cc:30:38:   in ‘constexpr’ expansion of 
> ‘std::lower_bound<const int*, int>(ca0.std::__debug::array<int, 
> 12>::end(), ca0.std::__debug::array<int, 12>::begin(), 6)’
> /home/fdt/dev/gcc/install/include/c++/10.0.0/bits/stl_algobase.h:1482:7: 
> error: inline assembly is not a constant expression
>  1482 |       __glibcxx_requires_partitioned_lower(__first, __last, 
> __val);
>       |       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> /home/fdt/dev/gcc/install/include/c++/10.0.0/bits/stl_algobase.h:1482:7: 
> note: only unevaluated inline assembly is allowed in a ‘constexpr’ 
> function in C++2a
>
> This time it is done in no time. Of course you can see that the asm 
> trick to generate a non-constant condition is not so nice.
>
> We just don't see the asm call parameter but showing 
> __glibcxx_requires_partitioned_lower is not so bad. For this reason 
> the tests in this patch are not checking for any failure message. 
> We'll see how to adapt when we have the necessary front end help to 
> generate this compilation error.
>
> Of course I could have a nicer compilation error by directly calling 
> __glibcxx_requires_valid_range in the algo and stop doing it within 
> __glibcxx_requires_partitioned_lower. Do you want to expose all macros 
> this way ?
>
>
>     * include/debug/formatter.h (__check_singular): Add C++11 constexpr
>     qualification.
>     * include/debug/helper_functions.h (__check_singular): Likewise. Skip
>     check if constant evaluated.
>     (__valid_range): Remove check skip if constant evaluated.
>     * include/debug/macros.h 
> [_GLIBCXX_HAVE_BUILTIN_IS_CONSTANT_EVALUATED]
>     (_GLIBCXX_DEBUG_VERIFY_COND_AT): Define.
>     * testsuite/25_algorithms/lower_bound/constexpr.cc (test): Add 
> checks on
>     lower_bound results.
>     * testsuite/25_algorithms/upper_bound/constexpr.cc (test): Likewise.
>     * testsuite/25_algorithms/lower_bound/debug/
>     constexpr_partitioned_neg.cc: New.
>     * testsuite/25_algorithms/lower_bound/debug/
>     constexpr_partitioned_pred_neg.cc: New.
>     * testsuite/25_algorithms/lower_bound/debug/
>     constexpr_valid_range_neg.cc: New.
>     * testsuite/25_algorithms/lower_bound/debug/partitioned_neg.cc: New.
>     * testsuite/25_algorithms/lower_bound/debug/partitioned_pred_neg.cc:
>     New.
>     * testsuite/25_algorithms/upper_bound/debug/
>     constexpr_partitioned_neg.cc: New.
>     * testsuite/25_algorithms/upper_bound/debug/
>     constexpr_partitioned_pred_neg.cc: New.
>     * testsuite/25_algorithms/upper_bound/debug/
>     constexpr_valid_range_neg.cc: New.
>     * testsuite/25_algorithms/upper_bound/debug/partitioned_neg.cc: New.
>     * testsuite/25_algorithms/upper_bound/debug/partitioned_pred_neg.cc:
>     New.
>
> Tested under Linux x86_64 normal and debug modes.
>
> Ok to commit ?
>
> Is there any request done to get the necessary front end help ? 
> Ideally a builtin to generate the error at a given source/file line 
> for a given function would be perfect cause we have all those info.
>
> François
>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: debug_constexpr.patch
Type: text/x-patch
Size: 21207 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/libstdc++/attachments/20191209/b3b1988e/attachment.bin>


More information about the Libstdc++ mailing list