[Bug middle-end/104971] [9/10/11/12 Regression] Optimisation for __builtin_ia32_readeflags corrupts the stack

[andrew.cooper3 at citrix dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104971

--- Comment #3 from Andrew Cooper <andrew.cooper3 at citrix dot com> ---
So yes - my experimentation did start from investigating the memory ordering
behaviour of these builtins, based on a thread on LKML.

The pushf in readflags and popf in writeflags have wildly different ordering
requirements, depending on which flags are wanted/modified.  AC for example
(and IF for kernels) need to not be reordered with respect to any memory
access.

As you observe, readflags in particular needs to not be reordered with any
instruction that modifies the arithmetic flags (which is most of them).

IMO, it would be safe to omit the pushf from readflags if the result is not not
used, because there are no unexpected side effects for pushf.

The same is not true of popf in writeflags, which has side effects even when
written twice with the same value.