martinmoene/expected-lite

Make version with no exception support

flexferrum opened this issue · 9 comments

flexferrum commented

For some special cases (and reasons) it would be perfect to have version of the nonstd::expected implementation with no exception support (like @TartanLlama version).

DBJDBJ commented

Note to self: Ask Martin to have the whole nonstd exceptions free.

  • 👍2
flexferrum commented

Note to self: Ask Martin to have the whole nonstd exceptions free.

Because of nonstd::optional_lite also has an issues with compilation without exceptions enabled.

martinmoene commented

In version 0.4.0.

DBJDBJ commented

Thanks for 0.4.0 ...

Basically this is not how MS STL switches to no C++ exceptions builds.

// expected.hpp #55
// Control presence of exception handling (try and auto discover):
#ifndef nsel_CONFIG_NO_EXCEPTIONS
# if defined(__cpp_exceptions) || defined(__EXCEPTIONS) || defined(_CPPUNWIND)
#  define nsel_CONFIG_NO_EXCEPTIONS  0
# else
#  define nsel_CONFIG_NO_EXCEPTIONS  1
# endif
#endif

MS STL does not use _CPPUNWIND, MS STL uses _HAS_EXCEPTIONS; thus your code using it too will allow users mixing your stuff with MS STL while in no C++ exceptions builds:

// Control presence of exception handling (try and auto discover):
// Now nonstd and MS STL both depend on the same symbol
#ifndef nsel_CONFIG_NO_EXCEPTIONS
# if defined(__cpp_exceptions) || defined(__EXCEPTIONS) || (_HAS_EXCEPTIONS == 1 )
#  define nsel_CONFIG_NO_EXCEPTIONS  0
# else
#  define nsel_CONFIG_NO_EXCEPTIONS  1
# endif
#endif
  • The reasoning is explained in this issue electronicarts/EASTL#399 (comment), I did for EASTL.
  • That change will also make possible using nonstd and MSVC (and MS STL), while using the /kernel switch.
  • ATL (still very usefull) also uses the same symbol

HTH

martinmoene commented

@DBJDBJ Thanks for your input.

Wondering if the line should also contain defined(_HAS_EXCEPTIONS), like:

# if defined(__cpp_exceptions) || defined(__EXCEPTIONS) || (defined(_HAS_EXCEPTIONS) && _HAS_EXCEPTIONS != 0)

Edit:

Apparently easy to forget, according to Preprocessing directives, 15.2 Conditional inclusion the following is fine:

# if defined(__cpp_exceptions) || defined(__EXCEPTIONS) || (_HAS_EXCEPTIONS != 0)
DBJDBJ commented

@martinmoene yes that is right, thanks ...

I think in the WIN32 world _HAS_EXCEPTIONS always exists and it is 0 or 1. This is how I understand the effect of vcruntime.h line 100. Thus this is also how MS STL is built each time whenever and wherever used.

// Depending on _HAS_EXCEPTIONS
// MS STL transforms to using SEH instead of c++ exceptions
// <vcruntime.h> #100
#ifndef _HAS_EXCEPTIONS 
    #ifdef _KERNEL_MODE
        #define _HAS_EXCEPTIONS 0
    #else
        #define _HAS_EXCEPTIONS 1
    #endif // _KERNEL_MODE
#endif // _HAS_EXCEPTIONS
// _CPPUNWIND not mentioned

Building nonstd with or without the /kernel switch, might clarify things considerably.

#include <windows.h> // must include for SEH to work

#ifndef WIN32
#error this is Windows build only
#endif // WIN32

#if _HAS_EXCEPTIONS
#error _HAS_EXCEPTIONS must not be 1 , add the /kernel switch to the cl command line
#endif

#include <vector>
#include <stdio.h>

// also would like to use 
// full nonstd in here
// #include <nonstd/whatever>

/*

Caveat Emptor: how to build with /kernel switch

it is not enough to just provide /kernel switch
one has to manually remove any /EH switch 
usually in  windows builds /EHsc is added by default

This will not stop the build but RTTI (/GR-) has to be  
also explicitly added as it is not switched off 
by using the /kernel switch
*/

static void does_not_compile(void)
{
    /*
    error C2980: C++ exception handling is not supported with /kernel
    try { throw 13; }
    catch (...) {}
    */
}

int main()
{
    __try {
        // also would like to use 
        // full nonstd in here
        std::vector<int> v_{ 1,2,3 };

        // std::out_of_range
        (void)v_.at(42);
    }
    __except ( EXCEPTION_EXECUTE_HANDLER /* aka 0*/) {
        puts("Structured Exception from MS STL was caught");
    }

    return 42;
}

Kind regards ...

DBJDBJ commented

I am slightly worried if I was clear enough, but it is not my fault: one, two or three symbols are made to dance to achieve one "thing".

  1. _CPPUNWIND should not be changed by users, but it should be checked for the presence of c++ exceptions
    1. besides being affected by 2 and 3 bellow it is also affected by compilation without the /EH switch
  2. _HAS_EXCEPTIONS = [ 1 | 0 ] is for users to use and that is what MST STL depends on
  3. _KERNEL_MODE is not for users to use; its value is the effect of the /kernel switch which also effects the value of the _HAS_EXCEPTIONS . And in turn effects the MS STL compilation.

Thus it is a "good thing" nonstd now depends on _HAS_EXCEPTIONS , value of.

HTH