[clang] P2558R2 not implemented, despite claim to contrary on cxx_status.html

[zygoloid]

Clang appears to have not implemented P2558R2, despite claiming to have done so.

This example is mishandled:

#define FOO$ 10
#define STR(x) #x
#define STR2(x) STR(x)
const char *p = STR2(FOO$)

This is valid under P2558R2, and p should be initialized to the string $ 10$:

• The first line defines a macro named FOO whose expansion is $ 10.
• The STR2(FOO$) macro expansion's argument expands to $ 10$.
• Therefore the result of stringization should be "$ 10$".

Under [lex.pptoken], adding $ et al to the basic source character set means that when $ appears in a source file (outside of a string or character literal), it always forms a single-character preprocessing-token. Ultimately, the issue here is that the GNU "dollars in identifiers" extension is a non-conforming extension post-P2558R2, so presumably we should stop enabling it by default in our conforming mode. Maybe it should only be enabled in -std=gnu++2c, not in -std=c++2c?

Also, the handling of

#define STR(x) #x
#define STR2(x) STR(x)
const char *p = STR2(\u0024);

is incorrect. This is accepted in all C++ language modes prior to Clang 17, and rejected in all C++ language modes in Clang 17 onwards. Both behaviors are wrong: it should be accepted in C++23 and earlier, and rejected in C++2c, because $ is part of the basic source character set only in C++2c onwards. (Even if this were correct, cxx_status.html should say this paper is implemented in Clang 17, not simply "Yes".)

[zygoloid]

The first issue is also demonstrating that we fail to implement [cpp.replace.general]/4. But the first issue can also be exhibited without violating that rule:

#define STR(X) #X
#define STR2 STR(
const char *q = STR2$10);

... is rejected instead of expanding the macro STR2.

[tavianator]

I presume you mean P2558R2, not P2588R2?

[zygoloid]

I presume you mean P2558R2, not P2588R2?

Thanks, yes I did!

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Richard Smith (zygoloid)

Clang appears to have not implemented P2558R2, despite claiming to have done so.

This example is mishandled:

#define FOO$ 10
#define STR(x) #x
#define STR2(x) STR(x)
const char *p = STR2(FOO$)

This is valid under P2558R2, and p should be initialized to the string $ 10$:

• The first line defines a macro named FOO whose expansion is $ 10.
• The STR2(FOO$) macro expansion's argument expands to $ 10$.
• Therefore the result of stringization should be "$ 10$".

Under [lex.pptoken], adding $ et al to the basic source character set means that when $ appears in a source file (outside of a string or character literal), it always forms a single-character preprocessing-token. Ultimately, the issue here is that the GNU "dollars in identifiers" extension is a non-conforming extension post-P2558R2, so presumably we should stop enabling it by default in our conforming mode. Maybe it should only be enabled in -std=gnu++2c, not in -std=c++2c?

Also, the handling of

#define STR(x) #x
#define STR2(x) STR(x)
const char *p = STR2(\u0024);

is incorrect. This is accepted in all C++ language modes prior to Clang 17, and rejected in all C++ language modes in Clang 17 onwards. Both behaviors are wrong: it should be accepted in C++23 and earlier, and rejected in C++2c, because $ is part of the basic source character set only in C++2c onwards. (Even if this were correct, cxx_status.html should say this paper is implemented in Clang 17, not simply "Yes".)