clang fails to reject invalid typeid expression and so crashes later on

[llvmbot]

Bugzilla Link	9208
Resolution	FIXED
Resolved on	May 09, 2011 18:14
Version	trunk
OS	All
Reporter	LLVM Bugzilla Contributor
CC	@DougGregor,@RealLitb

Extended Description

#include

template void foo() {};

int main() {
typeid(foo);
}

The above is invalid (at least according to g++) but since clang accepts it clang tries to generate the type's name using its name mangling logic and therefore hits the following assert:

Assertion failed: (false && "Overloaded and dependent types shouldn't get to name mangling"), function mangleType, file ItaniumMangle.cpp, line 1303.

g++'s output is:

typeinfo_assert.c++: In function ‘int main()’:
typeinfo_assert.c++:6: error: insufficient contextual information to determine type

[RealLitb]

This one and http://llvm.org/bugs/show_bug.cgi?id=7211 are probably related.

[RealLitb]

This one and http://llvm.org/bugs/show_bug.cgi?id=7211 are probably related.

It should be noted that the recent resolution of #7877 makes this code and similar code well-formed (I didn't test this specific case though). It interprets 14.1p1 which says "An overloaded function name shall not be used without arguments in contexts other than those listed." to not apply when the overloaded function name is "foo", as it resolves the template-id to an lvalue later and then makes it so this rule does not apply anymore to the then resolved lvalue, ignoring that the name once was the name of an overloaded function according to 14.1p1. As a result, it also allows statement-expressions whose mere constitute is a "foo;".

I don't know what the proper interpretation of the rule is, but I figured I would add a comment on this. It would IMO be bad if it accepts "foo;" but rejects "typeid(foo);" though, since I can't see any rule that would justify such a difference. Either accept both or reject both.

[llvmbot]

The typeid case should be fixed by my patch for #7877 . I tried checking it, but when i include on my windows 7 64 bit, VS2010 build, i get other errors. I know the sizeof, and decltype case works.

Also my initial patch did not address a bug in clang that would accept the following code:

void two(); void two(int);
int main()
{
two; // this should be an error
}

I just submitted a tiny patch to address it.
http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20110214/039095.html

Thanks for your feedback Johannes!

[llvmbot]

I've verified that this code does parse now, however there's still an assert in ItaniumMangle.cpp when clang tries to figure out what the type's name should be.

If I use a version with asserts disabled then an executable is generated, but code such as:

cout << typeid(foo).name();

simply prints out an empty string. Since gcc considers this code invalid I don't know what a correct name would be. I don't know if it'd be right or possible to use the type of lvalue that the template-id is resolved to, but if so then I guess I'd expect the above to print "FvvE".

[RealLitb]

The typeid case should be fixed by my patch for #7877 . I tried checking it,
but when i include on my windows 7 64 bit, VS2010 build, i get other
errors. I know the sizeof, and decltype case works.

Also my initial patch did not address a bug in clang that would accept the
following code:

void two(); void two(int);
int main()
{
two; // this should be an error
}

I just submitted a tiny patch to address it.
http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20110214/039095.html

Thanks for your feedback Johannes!

Hello Faisal, I tested your patch. It seems to reject

template
void f();

int main() {
f != 0;
}

Function pointers are allowed as operands of op!=. Was that an oversight or is this intentionally rejected? Comeau/EDG accept it. GC4.6 and GCC4.5.2 also accept "typeid(foo)". Overview about the various expressions and what compilers do (my silly tests may be wrong. please double-check):

                Clang     Comeau/EDG  GCC (4.5.2)

-----------------+------------------------------------
typeid(f); OK (?) OK OK
+f; OK OK ERROR
!f; OK OK ERROR
&f; OK OK OK
f == 0; ERROR OK ERROR
f != 0; ERROR OK ERROR
f && 0; OK OK ERROR
f || 0; OK OK ERROR
f ? f OK OK ERROR (only 1st op?)
: f
(f, f) OK OK ERROR
f; OK (?) OK OK

EDG accepts them all but GCC accepts only some and rejects others for no obvious reason.

[llvmbot]

Hi Johannes,
Thanks as usual for your encyclopedic attention to detail. Based on your feedback, I have submitted a patch to address some more cases in which uses of single function templateids should be resolvable.
http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20110307/039785.html

Thanks!

[DougGregor]

Clang trunk is correctly accepting all of these