https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96402
Fabian Vogt <fab...@ritter-vogt.de> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |fab...@ritter-vogt.de
--- Comment #4 from Fabian Vogt <fab...@ritter-vogt.de> ---
FTR, this was originally reported as
https://bugzilla.opensuse.org/show_bug.cgi?id=1174753 with some more info:
> Assembly dump of CAS16 operation generated:
> 00000000004006f0 <__aarch64_cas16_acq_rel>:
> 4006f0: 90000110 adrp x16, 420000
> <__libc_start_main@GLIBC_2.17>
> 4006f4: 3940e610 ldrb w16, [x16, #57]
> 4006f8: 34000070 cbz w16, 400704
> <__aarch64_cas16_acq_rel+0x14>
> 4006fc: 4860fc82 caspal x0, x1, x2, x3, [x4]
> 400700: d65f03c0 ret
> 400704: aa0003f0 mov x16, x0
> 400708: aa0103f1 mov x17, x1
> 40070c: c87f8480 ldaxp x0, x1, [x4]
> 400710: eb10001f cmp x0, x16
> 400714: fa510020 ccmp x1, x17, #0x0, eq // eq = none
> 400718: 54000061 b.ne 400724 <__aarch64_cas16_acq_rel+0x34>
> // b.any
> 40071c: c82fc490 stlxp w15, x16, x17, [x4]
> 400720: 35ffff6f cbnz w15, 40070c
> <__aarch64_cas16_acq_rel+0x1c>
> 400724: d65f03c0 ret
> 400728: d503201f nop
> 40072c: d503201f nop
>
> CASPAL instruction (at 4006fc) would compare and read 2 64-bit values
> starting at memory specified in register x4 to values stored in x0 and x1,
> and if they compare equal, overwrite values stored at x4 with x2 and x3.
>
> However the fallback instructions (starting at 40070c) writes x0 and x1
> (saved in x16 and x17 at 400704) to memory pointed by x4 rather than x2 and
> x3.
