openmp: Align kmp_int64 and kmp_uint64 to 64 bits.

[pcc]

Summary

On 32-bit x86, both of these types have an alignment of 32 bits, which means
that atomic operations involving these types require a runtime call. We can
avoid the overhead of the runtime call, as well as a dependency on libatomic,
by aligning both types to 64 bits.

As far as I can tell these types are not part of the library's public
interface, so this change should be fine from an ABI perspective.

[AndreyChurbanov]

Not sure this is a good change, e.g. gcc produces tons of warnings:

warning: ignoring attributes on template argument 'kmp_uint64 {aka long long unsigned int}' [-Wignored-attributes]
  std::atomic<kmp_uint64> next_ticket;
                        ^

and apparently ignores the attribute. Strictly, C++ does not allow to specify alignment in typedef, AFAIK. Though many compilers may support this.

Maybe it would be better to align objects, not types. There should not be many 64-bit atomic objects in the code those need explicit alignment, I think. And there are places where (poorly aligned) pair of 4-byte integers is changes atomically as a single 8-bype object. We may need to reorder structure fields then, as aligned type won't help here.

Just my 2c.

openmp/runtime/src/kmp_os.h

 #if KMP_OS_UNIX
 #define KMP_END_OF_LINE "\n"
 typedef char kmp_int8;
 typedef unsigned char kmp_uint8;
 typedef short kmp_int16;
 typedef unsigned short kmp_uint16;
 typedef int kmp_int32;
 typedef unsigned int kmp_uint32;
-typedef long long kmp_int64;
-typedef unsigned long long kmp_uint64;
+typedef __attribute__((aligned(8))) long long kmp_int64;
+typedef __attribute__((aligned(8))) unsigned long long kmp_uint64;
 #define KMP_INT32_SPEC "d"
 #define KMP_UINT32_SPEC "u"
 #define KMP_INT64_SPEC "lld"
 #define KMP_UINT64_SPEC "llu"
 #endif /* KMP_OS_UNIX */
 
 #if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS
 #define KMP_SIZE_T_SPEC KMP_UINT32_SPEC