OPAL_PATCHER_BEGIN/OPAL_PATCHER_END is broken in some cases

[pinskia]

Forwarded from https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95692 :

ompi/opal/mca/patcher/patcher.h

Line 31 in 0bccfcd

# define OPAL_PATCHER_BEGIN \

The inline-asm needs to clobber memory to avoid moving the END before a different restore the r2.

That is BEGIN should be:

asm volatile ("std 2, %0" : "=m" (toc_save) :: "memory");
asm volatile ("nop; nop; nop; nop; nop" ::: "memory");;

While END should be:

asm volatile ("ld  2, %0" : : "m" (toc_save) : "memory");;

[devreal]

Thanks @pinskia for tracking this (over 4 years)! Would you be able to file a PR that adds the clobber?

[pinskia]

Thanks @pinskia for tracking this (over 4 years)! Would you be able to file a PR that adds the clobber?

I don't have a way to test it fully. I was just using godbolt (https://godbolt.org/) to do quick testing by looking at the generated assembly output.

[jsquyres]

@bosilca @janjust Can Nvidia do some testing here, and file a PR if it all works as expected? You're the main consumers of patcher these days.