Re: [PATCH 3/7] Remove atomic_bit_set/bit_test_set

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 06/07/22 17:56, H.J. Lu via Libc-alpha wrote:
> On Wed, Jul 6, 2022 at 1:31 PM Noah Goldstein via Libc-alpha
> <libc-alpha@sourceware.org> wrote:
>>
>> On Wed, Jul 6, 2022 at 1:14 PM Wilco Dijkstra <Wilco.Dijkstra@arm.com> wrote:
>>>
>>> Hi Noah,
>>>
>>> The goal here is to move to the standard atomics, not to invent our own set for
>>> convenience or just for fun to be different etc. There were 2 uses of atomic_bit_set
>>> in all of GLIBC, and I cannot believe that atomic_fetch_or (&x, 1 << bit) could cause
>>> any confusion given that it is very similar to x |= 1 << bit. We don't wrap such
>>> expressions in macros like bitset (x, bit) either - it just doesn't make sense.
>>>
>>> As I mentioned, I'll use EXITING_BITMASK rather than 1 << EXITING_BIT.
>>
>> Alright.
> 
> Before GCC 12,
> 
> (__atomic_fetch_or (&x, MASK, __ATOMIC_ACQUIRE) & MASK) != 0;
> 
> will be optimized to "lock btsl" only if x is unsigned.  But cancelhandling in
> 
> if (atomic_bit_test_set (&pd->cancelhandling, TERMINATED_BIT) == 0)
> 
> is signed which leads to the much worse code when GCC 11 or older are
> used.
> 

I think it should be safe to change cancelhandling to be unsigned,
although this is really a micro-optimization that we should handle
in the compiler instead roll-out our own atomics.