[llvm-branch-commits] [libcxx] release/18.x: [libcxx] Align `__recommend() + 1` by __endian_factor (#90292) (PR #95264)

[via llvm-branch-commits]

tickerguy wrote:

Its definitely architecture-specific.  The media in question where I first saw this will boot and run on both a Pi3 and Pi4 but on the "4" it never blows up, whether via the original "make" or if I execute the reproducer shell script after it blows on the "3", I shut it down, stick the same card in the "4", boot it and then run the reproducer script.

On the "3" it does blow up, and not just there -- although its intermittent now that I've broken out the routines in the one file that was crashing reliably into three.  This started when those routines, all of which are heavy with OpenSSL function calls (and little else; they perform AES encryption, decryption, and MAC computation) were part of a much larger file and it failed reliably in those routines; I split them out in an attempt to isolate the problem.  Obviously separating a function call from one file with a lot of functions into its own separate ".c" file shouldn't change anything but now it crashes a good part of the time - but not every time - in that if I run the reproducer script it will will crash perhaps half or 3/4 of the time -- but the other times completes and produces the expected .o file.  That its intermittent is even worse, obviously, since you'd expect a compilation of the same file on the same CPU to act in a reasonably-deterministic fashion.

https://github.com/llvm/llvm-project/pull/95264