I have reproduced the https://svn.boost.org/trac10/ticket/4115 problem by applying boostorg/lexical_cast#31 and #307 on VC9.0/10.0/11.0, the fix for LexicalCast proposed in boostorg/lexical_cast#32.

I'm curious to know why you think injecting typeinfo into the namespace in which it really should be defined anyway is such a bad idea?
Though we should certainly restrict the workaround to pre-msvc-12, as it's only required for obsolete compilers anyway.

I did not say it is a bad idea, it is convenient but costly.

The IAR compiler for Arm uses Dinkumware and if I disable exceptions, which is admittedly I guess a non-standard extension, then it fails to compile this injection because it does not have ::typeinfo.
So, I would be in favour of limiting this kind of thing to only the compilers where it is known to work, rather than assuming it works in general.

Digging around more on this, I'm not sure that #307 is the right solution after all.

I fixed this another way by a very simple modification to dinkumware.hpp on line 89, which currently is:

#if BOOST_MSVC < 1800

but I think it should be:

#if defined(_MSC_VER) && BOOST_MSVC < 1800

What do you think, @jzmaddock? I believe you added that line 6 months ago. :)

You're correct: that's supposed to detect old MSVC versions, checking for defined(BOOST_MSVC) should do the trick also?

#307 is the right way, because Boost.TypeInfo is a portable solution to the problem the code is solving.

You're correct: that's supposed to detect old MSVC versions, checking for defined(BOOST_MSVC) should do the trick also?

defined(__ghs__) is clearly detecting something different than MSVC

You're correct: that's supposed to detect old MSVC versions, checking for defined(BOOST_MSVC) should do the trick also?

Yes to defined(BOOST_MSVC) since the whole point of BOOST_MSVC is avoid checking _MSC_VER which MSVC-pretenders define.

You're correct: that's supposed to detect old MSVC versions, checking for defined(BOOST_MSVC) should do the trick also?

Sure, I was just suggesting what I already saw used repeatedly within that file, but now I realize that they were testing the version of _MSC_VER whereas this is BOOST_MSVC.
I trust you all know these macros better than I do.

Even though we might agree about that line 89 needing a defined(BOOST_MSVC), it is actually tangential to this issue, which is about unconditionally including <typeinfo>, when there is an existing cross-library solution for it in Boost.Core.

I'm inclined to agree with @Kojoley that their solution is the right way to go, but I think the PR needs to also remove all the references to BOOST_NO_STD_TYPEINFO from the tests, etc? Since I think Dinkumware is the only place that defines it.

So, I would be super happy about someone putting the defined(BOOST_MSVC) in line 89 in the interim, to at least fix the broken conditional.

Even though we might agree about that line 89 needing a defined(BOOST_MSVC), it is actually tangential to this issue, which is about unconditionally including <typeinfo>, when there is an existing cross-library solution for it in Boost.Core.

I'm inclined to agree with @Kojoley that their solution is the right way to go, but I think the PR needs to also remove all the references to BOOST_NO_STD_TYPEINFO from the tests, etc? Since I think Dinkumware is the only place that defines it.

So, I would be super happy about someone putting the defined(BOOST_MSVC) in line 89 in the interim, to at least fix the broken conditional.

I does confirm problem (IAR) and possibly solution works.