STR2WSTR(__FUNCTION__) should work in microsoft mode

[nico]

Bugzilla Link	11789
Resolution	FIXED
Resolved on	Sep 07, 2015 01:26
Version	unspecified
OS	Windows NT
Blocks	llvm/llvm-bugzilla-archive#12477
CC	@majnemer,@zmodem,@tritao

Extended Description

#include <stdio.h>

#define _FOO_STR2WSTR(s) L##s
#define FOO_STR2WSTR(s) _FOO_STR2WSTR(s)

int main() {
//const wchar_t* kFoo = L__FUNCTION__; // doesn't work in cl.exe
const wchar_t* kWFoo = FOO_STR2WSTR(FUNCTION); // works in cl
}

Both gcc and clang don't allow this, but cl.exe does. (gcc and clang support it for e.g. FILE.)

This is used in msvc's c++ standard headers.

[nico]

assigned to @nico

[nico]

(In xstddef, with _HAS_EXCEPTIONS defined to 0.)

[nico]

Stupid -fsyntax-only workaround: change crtdefs.h to #define FUNCTIONW to e.g. _STR2WSTR(FILE) instead.

[llvmbot]

A little more information about this since I ran into the same issue, i.e. when including in VS2010:

See this link: http://blog.shuva.in/index.php?entry=entry080823-063456

Microsoft's compiler turns FOO_STR2WSTR(FUNCTION) into __LPREFIX(FUNCTION). __LPREFIX isn't a macro, it's a weird compiler intrinsic that does what the "L" prefix is meant to do. This seems to be specifically for handling FUNCTION, because FOO_STR2WSTR(ANYTHING_ELSE) just results in LANYTHING_ELSE. Here's my small test case:

// C++ file:
#define __STR2WSTR(str) L##str
#define _STR2WSTR(str) __STR2WSTR(str)
#define FUNCTIONW _STR2WSTR(FUNCTION)
#define UNDEF_TEST _STR2WSTR(NOTFUNCTION)

int main( int argc, char* argv[] ) {
(void)FUNCTIONW;
(void)UNDEF_TEST;
return 0;
}

// Preprocessed output from VC10:
int main( int argc, char* argv[] ) {
(void)LPREFIX( FUNCTION);
(void)L__NOTFUNCTION;
return 0;
}

Maybe Microsoft's preprocessor specifically handles concatenating L ## FUNCTION?

[llvmbot]

I did some more testing. It looks like VC does a bunch of stuff to make it appear like FUNCTION is a macro. This isn't just for the case of the "L" prefix. Note that when prefixing with some garbage, "QWERTY", the preprocessor generates a result similar to the "L" case. It doesn't do anything special for strings though.

// C++ file:
#define __STR2WSTR(str) L##str
#define _STR2WSTR(str) __STR2WSTR(str)
#define FUNCTIONW _STR2WSTR(FUNCTION)
#define UNDEF_TEST _STR2WSTR(NOTFUNCTION)

#define __OTHER_PREFIX(str) QWERTY##str
#define _OTHER_PREFIX(str) __OTHER_PREFIX(str)
#define OTHER_PREFIX _OTHER_PREFIX(FUNCTION)

#define __STRING_PREFIX(str) "a string" str
#define _STRING_PREFIX(str) __STRING_PREFIX(str)
#define STRING_PREFIX _STRING_PREFIX(FUNCTION)

int main( int argc, char* argv[] ) {
(void)FUNCTIONW;
(void)UNDEF_TEST;
(void)OTHER_PREFIX;
(void)STRING_PREFIX;
return 0;
}

// Preprocessed output from VC10:
int main( int argc, char* argv[] ) {
(void)LPREFIX( FUNCTION);
(void)L__NOTFUNCTION;
(void)QWERTY__FSTREXP FUNCTION ;
(void)"a string" FUNCTION ;
return 0;
}

[llvmbot]

So this appears to be something special to do with FUNCTION to make it look like a macro. I would expect this same behavior with FUNCSIG and FUNCDNAME, since they're all supposed to behave like macros, but really are implemented in the compiler and not the preprocessor. The preprocessor specially handles two things: L##FUNCTION, and ##FUNCTION ( representing any token). L##FUNCTION gets turned into __LPREFIX(FUNCTION). Something else, like QWERTY##FUNCTION gets turned into QWERTY__FSTREXP FUNCTION. One other rule is, whenever one of the arguments to the concatenation operator is FUNCTION, a space goes in between them.

I think this is reasonable to implement in the preprocessor. I'll have to do some more research into how the "FSTREXP" identifiers get treated in the compiler -- I want to see if they get turned into anything else there.

[nico]

Does MSVC compile this program?:

void f() {
const char* f = "foo" FUNCTION;
}

One possible way to make this work would be to have the parser tell the preprocessor about the current function name every time it enters a function, and have the preprocessor handle FUNCTION like FILE (i.e. like a normal string) except expanding it to the current "current function name" it got from the parser.

This might get tricky with template functions, where the parameter types aren't known when that function is getting parsed.

What does FUNCTION expand to in template functions in msvc? Does msvc support PRETTY_FUNCTION? What does that expand to?

[llvmbot]

MSVC does compile that program, and the preprocessed output is the same as the input.

MSVC doesn't support PRETTY_FUNCTION. Here's a list of what the preprocessor does support: http://msdn.microsoft.com/library/b0084kay.aspx

I think we should handle this in the preprocessor, as the regular FUNCTION and equivalent macros are currently handled.

[llvmbot]

It looks like the FSTREXP cases aren't worth compiling:

#include
using namespace std;

const char* const QWERTY = "qwerty";

#define __OTHER_PREFIX(str) QWERTY##str
#define _OTHER_PREFIX(str) __OTHER_PREFIX(str)
#define OTHER_PREFIX _OTHER_PREFIX(FUNCTION)

#define __STRING_PREFIX(str) "a string" str
#define _STRING_PREFIX(str) __STRING_PREFIX(str)
#define STRING_PREFIX _STRING_PREFIX(FUNCTION)

int main() {
cout << OTHER_PREFIX << endl; // error here
cout << STRING_PREFIX << endl;
return 0;
}

As noted above, OTHER_PREFIX expands to QWERTY__FSTREXP FUNCTION. However, the compiler then complains that it doesn't recognize the identifier "QWERTY__FSTREXP". I'm not sure why the preprocessor emits this, then, maybe it's an internal implementation detail. Given this, I think it's only worth trying to implement the __LPREFIX thing.

[llvmbot]

I'm tackling the problem right now. I think easiest way would be to

1. Support wchar_t variant of FUNCTION (I'm calling it FUNCTIONW as I didn't realize that MS defines a macro of that name. It actually works fine this way, still, it might be better to rename it to something that doesn't collide).
2. In MS mode, predefine L__FUNCTION__ to FUNCTIONW.

The ability of MS compilers to paste string literals and predef variables would have to be handled separately.

I have a partial implementation of FUNCTIONW (without codegen, only for -fsyntax-only): https://github.com/avakar/clang/tree/bug11789

[tritao]

Hi Martin.

There has been some patches and discussion regarding this in the list:

http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20120611/058838.html

[llvmbot]

If you see my discussion with Richard Smith about my patch, it sounds like he would prefer a very simple patch, rather than emulating the exact behavior of MSVC, which I can understand. Given that, it might make sense to not even bother #defining L__FUNCTION__ to FUNCTIONW or anything like that. Just add a predefined expression L__FUNCTION__ that is the wide string function name, only available in ms compatibility mode. Thoughts?

[llvmbot]

If you see my discussion with Richard Smith about my patch, it sounds like he
would prefer a very simple patch, rather than emulating the exact behavior of
MSVC,

Yes, I've seen your original patch submission and found it too complex as well.

Given that, it might make sense to not even
bother #defining L__FUNCTION__ to FUNCTIONW or anything like that. Just add
a predefined expression L__FUNCTION__ that is the wide string function name,
only available in ms compatibility mode. Thoughts?

It would be sufficient for compatibility's sake, however, there is currently no way to get wchar_t version of FUNCTION and since it will have to be supported for the sake of MS, I'd make it available for non-MS modes as well. And, of course, we wouldn't want it to be called L__FUNCTION__ there. That's why I went with FUNCTIONW and a predefined macro.

[nico]

r159060

See http://thread.gmane.org/gmane.comp.compilers.clang.scm/52277 for discussion.

[timurrrr]

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