wrong detection of ALIGNED_64K_SUPPORTED

[Rasmus Villemoes]

When cross-compiling for arm with gcc >= 5.4 but binutils 2.25, the assembler complains

  /tmp/ccH0h4Uw.s:14: Error: alignment too large: 15 assumed

Using binutils 2.26 seems to be ok. I assume the binutils commit db2ed2e (Rationalize ARM .align) is what makes the difference. In any case, I think it would be better to replace the preprocessor checks with a configure-time check of whether the entire toolchain will accept __aligned(65536).

[Colin Ian King (colin-king)]

Fix committed:

http://kernel.ubuntu.com/git/cking/stress-ng.git/commit/?id=9c7d1c449945538d35874cffe713e8d259e75beb

Please can you pull the latest changes and test build this and let me know if this fixes this for you. Thanks!

[Colin Ian King (colin-king)]

Before you do a build; please "make clean" first to allow stress-ng to regenerate it's configuration settings.

[Rasmus Villemoes (villemoes)]

I'll check tomorrow, but I highly doubt this is correct. You need and want to check for 64K, aka 2^16 aka 65536, not just 64 byte alignment. The binutils commit I mentioned removed an explicit bound of 15 for the max log2 alignment.

[Colin Ian King (colin-king)]

Hi Rasmus, yep, I made a thinko here.. I've modified this to do 64K alignment now.

See: http://kernel.ubuntu.com/git/cking/stress-ng.git/commit/?id=2c03776f1b20007ecc94d9d6c21658ac27dd905d

[Rasmus Villemoes (villemoes)]

This seems to work. But I think the modifications of stress-ng.h are buggy - defining ALIGN64 or not should _not_ depend on HAVE_ALIGNED_64K - it should still just depend on the compiler being non-ancient (and implicitly assume that the rest of the toolchain, assembler and linker, always supports 64 bytes alignment - which isn't necessarily true, nor is a cache line always 64 bytes, but in practice this should be good enough, at least until someone actually files a bug report).

[Colin Ian King (colin-king)]

I've added a configuration 64 byte alignment check too.