[clang][Headers] Avoid compiler warnings in builtin headers

[ddcc]

Summary

While debugging module support using -Wsystem-headers, we discovered that if
-Werror, and -Wundef or -Wmacro-redefined are specified, they can cause errors
to be generated in these builtin headers.

[aaron.ballman]

Thanks for catching this! As far as these changes go, they're pretty reasonable, but don't seem like they hit all of the problem areas. There's 20+ occurrences of #if __STDC_VERSION__ and 10+ of #if __cplusplus in lib/Headers; shouldn't they all be updated?

clang/lib/Headers/stdint.h

94–95

It seems to me that a cleaner approach to this is to #undef the macros we expect to be redefining before we redefine them instead of disabling the definition warning entirely. For example, it would still be useful for a user to find out that our stdint.h redefines some symbol from their own header files from an earlier inclusion.

503

Why does this usage not need to be changed?

[ddcc]

Undef macros instead, handle other header files

[tra]

clang/lib/Headers/__clang_cuda_builtin_vars.h

37 ↗

(On Diff #449118)

Are there actual use cases where CUDA headers are used from a non-C++ compilation? I do not think they will compile with non-C++ compilation at all.

I'm curious how we managed to run into a situation where CUDA headers may be used w/o C++.

I think a better fix here and other CUDA headers may be to add

#if !defined(__cplusplus)
#error CUDA headers must be used from a CUDA/C++ compilation.
#endif

[ddcc]

clang/lib/Headers/__clang_cuda_builtin_vars.h

37 ↗

(On Diff #449118)

Sure, I can do that. I'm not familiar with CUDA or its headers, so I previously just updated them all.

[ddcc]

Error out on undef __cplusplus in CUDA headers

[tra]

clang/lib/Headers/__clang_cuda_builtin_vars.h

14 ↗

(On Diff #449403)

Nit: the error message sounds a bit odd.
Perhaps rephrase it as This CUDA header requires C++.

clang/lib/Headers/__clang_hip_cmath.h

381 ↗

(On Diff #449403)

HIP headers are also not expected to compile w/o C++ and we may want a #errorhere, too.

Actually, I'm not even sure if we shopuld/need to add #if defined(__cplusplus)in files that are clearly C++ -- undefined macro warning will be the least of one's worries if they attempt to compile it w/o C++.

Perhaps it would make sense to limit these changes to the headers intended to be used from both C and C++? Both CUDA and HIP are expected to never be included from a non-C++ compilation and will never trigger the undefined macro warning on __cplusplus.

@yaxunl - FYI.

[ddcc]

clang/lib/Headers/__clang_hip_cmath.h

381 ↗

(On Diff #449403)

Sure, I'm fine with leaving those headers alone since I'm not familiar with them.

[ddcc]

Drop changes to CUDA/HIP/OpenMP headers

[iana]

clang/lib/Headers/stdint.h

102–103

What are you seeing that's defining __int_least64_t and all these other ones by the time you get here? It's surprising to me that so much preprocessor state exists when you hit this point considering that this file doesn't include anything else. Is this some kind of artifact with how the OS module map is making a module for stdint.h?

[iana]

clang/lib/Headers/stdint.h

102–103

Oh I see, it's not these ones that are the problem, it's the redefinitions below. I guess it's a bigger change, but I wonder if flipping the order would be cleaner? e.g.

#ifdef __INT32_TYPE__
# define __int_least32_t int32_t
#endif

#ifdef __INT64_TYPE__
# ifndef __int_least32_t
#  define __int_least32_t int64_t
# endif
#endif

I guess it's an extra line per type doing it that way, but maybe it's more obvious? (Maybe I just don't do a lot of preprocessor programming, but #undef feels like a code smell to me)

[ddcc]

clang/lib/Headers/stdint.h

102–103

I'm a little reluctant to change this file too much since it's tedious and error-prone, but yeah if we can agree on a solution, I can go ahead and make the change.

[aaron.ballman]

These changes look reasonable to me, thanks!

clang/lib/Headers/stdint.h

102–103

I'm not toooooo sad about the #undef though this does make preprocessing this file do more work. However, rearranging the macros in this file might be a challenge given how much these macros interact with one another. From the comments just above:

* These macros are defined using the same successive-shrinking approach as
* the type definitions above. It is likewise important that macros are defined
* in order of decending width.

clang/lib/Headers/stddef.h

 namespace std { typedef decltype(nullptr) nullptr_t; }
 using ::std::nullptr_t;
 #endif
 #endif
 #undef __need_NULL
 #endif /* defined(__need_NULL) */
 
 #if defined(__need_STDDEF_H_misc)
-#if __STDC_VERSION__ >= 201112L || __cplusplus >= 201103L
+#if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L) ||              \
+    (defined(__cplusplus) && __cplusplus >= 201103L)
 #include "__stddef_max_align_t.h"
 #endif
 #define offsetof(t, d) __builtin_offsetof(t, d)
 #undef __need_STDDEF_H_misc
 #endif  /* defined(__need_STDDEF_H_misc) */
 
 /* Some C libraries expect to see a wint_t here. Others (notably MinGW) will use
 __WINT_TYPE__ directly; accommodate both by requiring __need_wint_t */

clang/lib/Headers/stdint.h

  * suboptimal.
  *
  * In violation of the standard, some targets do not implement a type that is
  * wide enough to represent all of the required widths (8-, 16-, 32-, 64-bit).
  * To accommodate these targets, a required minimum-width type is only
  * defined if there exists an exact-width type of equal or greater width.
  */
 
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wmacro-redefined"
+
 #ifdef __INT64_TYPE__
 # ifndef __int8_t_defined /* glibc sys/types.h also defines int64_t*/
 typedef __INT64_TYPE__ int64_t;
 # endif /* __int8_t_defined */
 typedef __UINT64_TYPE__ uint64_t;
 # define __int_least64_t int64_t
 # define __uint_least64_t uint64_t
 # define __int_least32_t int64_t
 # define __uint_least32_t uint64_t
 # define __int_least16_t int64_t
 # define __uint_least16_t uint64_t
 # define __int_least8_t int64_t
 # define __uint_least8_t uint64_t
 #endif /* __INT64_TYPE__ */
 

 # define INT_LEAST64_MAX   __INT_LEAST64_MAX
 # define UINT_LEAST64_MAX __UINT_LEAST64_MAX
 # define INT_FAST64_MIN    __INT_LEAST64_MIN
 # define INT_FAST64_MAX    __INT_LEAST64_MAX
 # define UINT_FAST64_MAX  __UINT_LEAST64_MAX
 
 /* FIXME: This is using the placeholder dates Clang produces for these macros
    in C2x mode; switch to the correct values once they've been published. */
 #if __STDC_VERSION__ >= 202000L
 # define UINT_LEAST64_WIDTH __UINT_LEAST64_WIDTH
 # define INT_LEAST64_WIDTH  UINT_LEAST64_WIDTH
 # define UINT_FAST64_WIDTH  __UINT_LEAST64_WIDTH
 # define INT_FAST64_WIDTH   UINT_FAST64_WIDTH
 #endif /* __STDC_VERSION__ */
 #endif /* __INT_LEAST64_MIN */
 
 

 #if __STDC_VERSION__ >= 202000L
 #define PTRDIFF_WIDTH    __PTRDIFF_WIDTH__
 #define SIG_ATOMIC_WIDTH __SIG_ATOMIC_WIDTH__
 #define SIZE_WIDTH       __SIZE_WIDTH__
 #define WCHAR_WIDTH      __WCHAR_WIDTH__
 #define WINT_WIDTH       __WINT_WIDTH__
 #endif
 
+#pragma GCC diagnostic pop
+
 #endif /* __STDC_HOSTED__ */
 #endif /* __CLANG_STDINT_H */