Allow reinterpret_cast in constant expressions in microsoft mode

[nico]

Bugzilla Link	12116
Resolution	WONTFIX
Resolved on	Aug 29, 2018 10:24
Version	unspecified
OS	All
Blocks	llvm/llvm-bugzilla-archive#12477
CC	@DougGregor,@efriedma-quic,@tritao,@zygoloid

Extended Description

When compiling ms headers with -nobuiltininc, clang picks up offsetof() from msvc's stddef.h ( http://msdn.microsoft.com/en-us/library/dz4y9b9a(v=vs.80).aspx ). If it's used with chrome's STATIC_ASSERT macro, clang complains about "reinterpret_cast not allowed in constant expressions".

I came up with an offsetof() macro that exhibits the same error:

hummer:clang thakis$ cat test.cc
#define offsetof(s,m)
(long)(void*)&((reinterpret_cast<s*>(0)->m))

template
struct GpuCompileAssert {
};

struct SetToken {
int a, token;
};

GpuCompileAssert<(bool(offsetof(SetToken, token) == 4))> foo;

hummer:clang thakis$ ../../Release+Asserts/bin/clang -c test.cc -fms-extensions -fms-compatibility
here4!
here3!
test.cc:12:18: error: non-type template argument of type 'bool' is not an integral constant expression
GpuCompileAssert<(bool(offsetof(SetToken, token) == 4))> foo;
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
test.cc:12:24: note: cast which performs the conversions of a reinterpret_cast is not allowed in a constant expression
GpuCompileAssert<(bool(offsetof(SetToken, token) == 4))> foo;
^
test.cc:2:3: note: expanded from macro 'offsetof'
(long)(void*)&((reinterpret_cast<s*>(0)->m))
^
1 error generated.

We should allow reinterpret_cast<>()s in constant expressions in microsoft mode.

[efriedma-quic]

How about "don't use -nobuiltininc" instead? We are never going to support MSVC's stdarg.h, and given that, we really just need to make our builtin includes either work or forward appropriately.

[zygoloid]

Sorry, I'm not prepared to break the constant expression code this much. Accepting (say) reinterpret_casts of null pointers is far from enough to get this treated as a constant expression. "((S*)0)->m" isn't allowed in a constant expression, because you can't access members of null pointers in a constant expression. And the cast from void* to long isn't allowed in a constant expression either.

However, the expression is foldable. If Eli's suggestion is unworkable for some reason and you really need this code pattern to compile, I think it'd be acceptable to make either -fms-extensions or -fms-compatibility allow constant folding in non-type template arguments.

[nico]

Ok.

[timurrrr]

mentioned in issue llvm/llvm-bugzilla-archive#12477