[Bug target/120001] On RISC-V with -O2 and -O3 __sync_or_and_fetch in a loop renders as an endless loop and multiple amoor

[pinskia at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120001

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Vincenzo Romano from comment #6)
> MAybe I have not been clear.
> In cases with -O2 and -O2 the infinite loop is over 2 amoor identical
> instruction.
> Maybe I don't know RISC-V architecture that well, but it looks like a bug to
> me: one amoor should suffice, just like with -Os.


Again GCC unrolled the loop.

That is:

l1:
{
  amoor.d.aqrl  a4,a5,0(a0)
  goto l1;
}

Into:
l1:
{
  amoor.d.aqrl  a4,a5,0(a0)
//  goto l2;
//l2:
  amoor.d.aqrl  a4,a5,0(a0)
goto l1
}

Copied the inner block one time for better alignment/placement in this case.

It does not change the behavior of the function since it is an infinite loop.

Note this behavior shows up on aarch64 too (with -march=armv8.1-a). it is not
riscv specific. It is done because the size of the basic block of the loop is
small and copying it will be decent for the alignment.