missed warning for ODR violation.

[pawelsopensource]

Bugzilla Link	35566
Version	5.0
OS	Linux
CC	@adrian-prantl,@dwblaikie,@pogo59

Extended Description

% cat tu1.cpp
struct S { int i = 0; };
S foo() { return S(); }

% cat tu2.cpp
struct S { double i = 0.0; };
S bar() { return S(); }

% clang++ -Wodr -std=c++1y -fPIC -flto tu1.cpp tu2.cpp -c
% clang++ -Wodr -std=c++1y -fPIC -flto -shared tu1.o tu2.o -o tu.so -fuse-ld=lld

from the other side, g++ 6.3.1 reports ODR violation.

% g++ -Wodr -std=c++1y -fPIC -flto tu1.cpp tu2.cpp -c
% g++ -Wodr -std=c++1y -fPIC -flto -shared tu1.o tu2.o -o tu.so -fuse-ld=gold
tu1.cpp:1:8: warning: type ‘struct S’ violates the C++ One Definition Rule [-Wodr]
struct S { int i = 0; };
^
tu2.cpp:1:8: note: a different type is defined in another translation unit
struct S { double i = 0.0; };
^
tu1.cpp:1:20: note: the first difference of corresponding definitions is field ‘i’
struct S { int i = 0; };
^
tu2.cpp:1:23: note: a field of same name but different type is defined in another translation unit
struct S { double i = 0.0; };
^
tu1.cpp:1:8: note: type ‘int’ should match type ‘double’
struct S { int i = 0; };
^

[dwblaikie]

LLVM tends not to emit diagnostics from the backend (everything's up-front before LLVM IR is produced) which means things like this aren't handled.

Not impossible to do, but rare/exceptional.

Certainly the GCC behavior here is rather nice/I wouldn't mind having it in LLVM, but I doubt it's high on anyone's list.

(CC'ing LTO & debug info folks in case it catches their eye)

[shafik]

CC @AaronBallman @erichkeane

[erichkeane]

Hmm... we don't have any cross-TU diagnostics in Clang at all. We'd end up having to do this in LTO, which happens at a layer below Clang, so this wouldn't be something that Clang could do anything about anyway.

[thesamesam]

(I'm not sure why clang:diagnostics was removed, given it is described as "New warning... in Clang".)