Hi Jerry!
Am 03.01.25 um 05:21 schrieb Jerry D:
On 1/2/25 12:04 PM, Harald Anlauf wrote:
Dear all,
this patch addresses overeager optimization of Cray pointers when
used in comparisons. Cray pointers are non-standard, and odd in a
sense that they were introduced before modern Fortran pointers.
Comparisons with e.g. a "NULL" pointer are actually comparisons
with integer zero etc., which means that while they are references
they can actually be "NULL" to mimic a disassociated pointer.
The only solution I could find was treating them locally as volatile
when used in a comparison. If someone has a better solution, please
share!
As this is a local solution, and a real-world legacy code using Cray
pointers would likely never use such a test in a vectorizable loop,
I expect negligible (performance and code-size) impact.
Regtested on x86_64-pc-linux-gnu. OK for mainline?
This PR is marked as a regression (since gcc-7), is this OK for
a (limited?) backport?
Thanks,
Harald
The hack is fairly isolated and simple. The problem is a quirk from the
past. The only way to expose an issue is to get it into the real world
and see if anyone notices a problem.
I do not have access to any code using Cray pointers beyond what is in
the testsuite. They might get used in some SPEC benchmark, we'll see.
OK for trunk and maybe 14. If you think farther back, up to you.
Pushed to mainline as r15-6545 so far.
Thanks for the feedback!
Harald
Jerry
