[llvm-bugs] [Bug 51102] New: [AArch64] Wrong code generated (?) for compare_exchange_strong on __int128_t

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=51102

            Bug ID: 51102
           Summary: [AArch64] Wrong code generated (?) for
                    compare_exchange_strong on __int128_t
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: AArch64
          Assignee: unassignedbugs at nondot.org
          Reporter: efriedma at quicinc.com
                CC: arnaud.degrandmaison at arm.com,
                    llvm-bugs at lists.llvm.org, smithp352 at googlemail.com,
                    Ties.Stuij at arm.com

Consider:

#include <stdatomic.h>
__int128_t ff(_Atomic __int128_t *p) {
    __int128_t e = 0;
    atomic_compare_exchange_strong(p, &e, 0);
    return e;
}
ff:
        mov     x8, x0
.LBB0_1:
        ldaxp   x0, x1, [x8]
        orr     x9, x0, x1
        cbnz    x9, .LBB0_4
        stlxp   w9, xzr, xzr, [x8]
        cbnz    w9, .LBB0_1
        ret
.LBB0_4:
        clrex
        ret


This sequence is subtly wrong in the case where the cmpxchg fails. 
Specifically, the ldaxp is not single-copy atomic, so the returned value might
tear.  At least, that's my understanding of the rules here.  So we need to
replace the "clrex" with a sequence like "stlxp w9, x0, x1, [x8]; cbnz w9,
.LBB0_1".

Not sure if this sort of tearing can actually happen in practice on commonly
used AArch64 chips.

I'll throw together a patch.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20210715/4ab65ba9/attachment.html>