Unconditional call to memset in value_initialized()

[Aleksey Gurtovoy]

The call is there to workaround bugs in a few specific compilers, but is applied unconditionally, cannot be optimized away, and thus significantly limits applicability of the component in a library context.

See ​http://old.nabble.com/boost::mpl::for_each-and-value_initialized-td25246848.html for the full discussion.

[niels_dekker]

Adds BOOST_VALUE_INITIALIZATION_NEEDS_MEMSET to trunk/boost/config/compiler

[niels_dekker]

Fernando (the author of value_init) mailed me that it's okay to him if I add an #ifdef to utility/value_init.hpp:

  #ifdef BOOST_VALUE_INITIALIZATION_NEEDS_MEMSET
       std::memset(&x, 0, sizeof(x));
  #endif

But therefor, this macro needs to be defined in <boost/config/compiler/...>, for the relevant compiler versions. Hereby I have attached my proposed additions to the trunk version of boost/config/compiler: ​config_compiler.patch Please check if it's okay.

[John Maddock]

The patch is OK as far as it goes, but we should have documentation and tests for each new macro, please see: http://www.boost.org/doc/libs/1_41_0/libs/config/doc/html/boost_config/guidelines_for_boost_authors.html#boost_config.guidelines_for_boost_authors.adding_new_defect_macros

Thanks, John.

[niels_dekker]

Thanks, John. I read that there's a BOOST_NO_SOMETHING naming convention. So now I'm considering to name the macro BOOST_NO_PROPER_VALUE_INITIALIZATION_FOR_ALL_TYPES. (Instead of BOOST_VALUE_INITIALIZATION_NEEDS_MEMSET.) So that the unit test should show that compilers that have this macro defined do indeed do an improper value-initialization, for at least one type.

I'll have a closer look later...

[niels_dekker]

Hereby I have just added a new patch for boost/config (to be applied to the trunk): ​https://svn.boost.org/trac/boost/attachment/ticket/3869/no_complete_value_initialization.patch It adds a new macro, BOOST_NO_COMPLETE_VALUE_INITIALIZATION, and includes documentation and tests. Please tell me what you think.

[niels_dekker]

New proposed patch, adding BOOST_NO_COMPLETE_VALUE_INITIALIZATION. Replaces the old config_compiler.patch (Update: removed Boost.Test dependency, as suggested by John Maddock.)

[niels_dekker]

Made memset call conditional, based on BOOST_NO_COMPLETE_VALUE_INITIALIZATION

[niels_dekker]

The ​value_init_workaround.patch I just attached should resolve this ticket. It assumes that the defect macro BOOST_NO_COMPLETE_VALUE_INITIALIZATION is defined for those compilers that need it (ticket #4080). Note that it adds an extra macro, BOOST_DETAIL_VALUE_INIT_WORKAROUND (defined as either 0 or 1), which controls whether the memset workaround is applied. This macro is merely intended as an implementation detail: I would like to have an extra test which toggles the value of the macro, in order to check if the workaround is both effective and necessary.

[niels_dekker]

(In [62307]) Made memset call in value_init conditional, see #3869. Updated the section "compiler issues" of its documentation.

[niels_dekker]

(In [63014]) Added value_init_workaround_test, reviewed by Fernando Cacciola, see #3869

[niels_dekker]

(In [63638]) Merged value_init fixes (extra tests + documentation) from trunk, see #3472, #3869.

[viboes]

Is there something else to be added? If not, could you close the issue?

[niels_dekker]

Closed, as suggested by Vicente (viboes).