Builtin::BIalloca not defined under MSVC

[llvmbot]

Bugzilla Link	4513
Resolution	FIXED
Resolved on	Mar 12, 2010 00:55
Version	unspecified
OS	Windows XP
Reporter	LLVM Bugzilla Contributor
CC	@asl

Extended Description

When compiling under MSVC, Builtin::BIalloca is not declared. This causes CGBuiltin.cpp to fail to compile.

The problem is that the system headers have this:

#if     !__STDC__
/* Non-ANSI names for compatibility */
#define alloca  _alloca
#endif  /* __STDC__*/

When Builtin.h #includes Builtins.def it does so with BUILTIN defined as:

#define BUILTIN(ID, TYPE, ATTRS) BI##ID,

LIBBUILTIN is not defined so Builtins.def supplies a default definition:

#if defined(BUILTIN) && !defined(LIBBUILTIN)
#  define LIBBUILTIN(ID, TYPE, ATTRS, HEADER) BUILTIN(ID, TYPE, ATTRS)
#endif

Then the line

LIBBUILTIN(alloca, "v*z",         "f",     "stdlib.h")

expands to BI_alloca, instead of the intended BIalloca,

This happens because LIBBUILTIN is defined in terms of BUILTIN. The preprocessor rescans and expands the parameter before the concat. If the line was BUILTIN(alloca, ...) then it would work properly.

Any of the following should fix this:

1. Explicitly #undef alloca
2. Change Builtins.h to explictly #define LIBBUILTIN and remove the default definition from Builtins.def
3. Change the default definition in Builtins.def to:

#  define LIBBUILTIN(ID, TYPE, ATTRS, HEADER) BUILTIN(##ID, TYPE, ATTRS)

but this isn't portable. I don't know if there's a portable way to stop the preprocessor from rescanning.

[lattner]

#undef'ing alloca is the right fix. Please prepare a patch for this and I'd be happy to apply it. I would do it, but I can't test on VC++.

[llvmbot]

#undef alloca for VC++
Here's a patch. Feel free to change/delete the comment.

[lattner]

The #undef shouldn't go in the .def file, please add it to the .cpp or .h file that #includes the .def file. Thanks, sorry for the confusion.

[llvmbot]

#undef alloca for VC++ (take 2)
This puts the #undef in Builtins.h. Builtins.def is also #included by Builtins.cpp but there's no issue there because it provides a #define for LIBBUILTIN.

[lattner]

Looks great, applied, thanks!
http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20090706/018833.html