[Bug target/71245] std::atomic<double> load/store bounces the data to the stack using fild/fistp

["ubizjak at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71245

--- Comment #2 from UroÅ Bizjak <ubizjak at gmail dot com> ---
(In reply to Peter Cordes from comment #0)

> We don't need to allocate any stack space.  We could implement the StoreLoad
> barrier with  lock or $0, -4(%esp) instead of reserving extra stack to avoid
> doing it to our return address (which would introduce extra store-forwarding
> delay before the ret could eventually retire).

This can be trivially implemented in config/i386/sync.md by changing

(define_insn "mfence_nosse"
  [(set (match_operand:BLK 0)
        (unspec:BLK [(match_dup 0)] UNSPEC_MFENCE))
   (clobber (reg:CC FLAGS_REG))]
  "!(TARGET_64BIT || TARGET_SSE2)"
  "lock{%;} or{l}\t{$0, (%%esp)|DWORD PTR [esp], 0}"
  [(set_attr "memory" "unknown")])

Recently x86 linux changed the barrier to what you propose. If it is worth, we
can change it without any problems. OTOH, we have "orl" here - should we change
it to "addl" to be consistent with kernel?