RFR: 8329257: AIX: Switch HOTSPOT_TOOLCHAIN_TYPE from xlc to gcc [v3]

[Joachim Kern]

On Mon, 29 Apr 2024 16:17:13 GMT, Joachim Kern <jkern at openjdk.org> wrote:

>> For the impatient, I suggest adopting mechanism 2, i.e. unconditionally
>> include <alloc.h> in globalDefinitions_gcc.hpp.
>> 
>> We can't include <alloca.h> in shared code, and there is a use in shared code
>> (in the relatively recently added JavaThread::pretouch_stack).
>> 
>> When I questioned whether we needed to include <alloca.h> at all, I referred
>> to a Linux man page I'd found on the internet (the same page mdoerr linked
>> to), which says (in part)
>> 
>> "By default, modern compilers automatically translate all uses of alloca()
>> into the built-in ..."
>> 
>> Apparently I should have kept digging, because it seems that page is
>> old/incorrect. A seemingly more recent Linux man page describes a different
>> way of handling it that is closer to what we're seeing, but still not quite
>> correct.
>> 
>> glibc's <stdlib.h> includes <alloca.h> if __USE_MISC is defined.
>> One of the ways __USE_MISC can become defined is if _GNU_SOURCE is defined,
>> and we define that for both gcc and clang toolchains.
>> 
>> We include <stdlib.h> in globalDefinitions_gcc.hpp. So when building with gcc,
>> globalDefinitions.hpp implicitly includes <alloca.h>.
>> 
>> The glibc definition of alloca is
>> 
>> #ifdef	__GNUC__
>> # define alloca(size)	__builtin_alloca (size)
>> #endif /* GCC.  */
>> 
>> So that explains why we don't need any explicit include of <alloca.h> when
>> building with gcc.  I expect there's something similar going on with Visual
>> Studio and Xcode/clang.  But apparently not with Open XLC clang.
>
> On AIX `stdlib.h` also would define `alloca`, if `__STRICT_ANSI__` wouldn't be set. 
> 
> 
> 780 #if !defined(__xlC__) || defined(__ibmxl__) || defined(__cplusplus)
> 781 #if defined(__IBMCPP__) && !defined(__ibmxl__)
> 782    extern "builtin" char *__alloca (size_t);
> 783 #  define alloca __alloca
> 784 #elif defined(__GNUC__) && !defined(__STRICT_ANSI__)
> 785    #undef  alloca
> 786    #define alloca(size)   __builtin_alloca (size)
> 787 #endif
> 
> 
> A small plain Testprogramm not using all of the flags we used in jdk build, does not set `__STRICT_ANSI__` and then `alloca` is defined correct.

The compiler flag introducing __STRICT_ANSI__ is -std=c++14. If I omit this explicit compiler flag the default is used, which is also c++14. But the default does not set __STRICT_ANSI__ but 2 other defines. I will try a build without -std=c++14 and if this works, we have a solution. Nevertheless i will interrogate IBM what the hell this behavior should be.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/18536#discussion_r1584461665