volatile deprecations in c++20
Jon Turney
jon.turney@dronecode.org.uk
Sun Oct 4 20:30:14 GMT 2026
On 03/10/2026 12:00, Christian Franke wrote:
> On Mon, 28 Sep 2026 11:40:44 +0200, Christian Franke wrote:
>> Jon Turney wrote:
>>>
>>> The remaining warnings with gcc 16 (caused by the change to -std=c+
>>> +20 by default) relate to the deprecation of the 'volatile'
>>> qualifier, which fall into two groups:
>>>
>>> ...
>>> I wonder if this is just wrong?
>>
>> The introduction of these errors IMO makes plenty of sense because
>> (volatile_var)++ "looks atomic" but never was. The 'volatile' only
>> ensures that memory accesses are not optimized away.
>>
>> With gcc 14, all the functions below translate to the same x86_64 code
>> if optimization is enabled:
>>
>> volatile unsigned i;
>> void inc0() { ++i; }
>> void inc1() { i++; }
>> void inc2() { i = i + 1; }
>> void inc3() { unsigned t = i; ++t; i = t; }
>>
>> x86_64 code:
>> movl i(%rip), %eax
>> addl $1, %eax
>> movl %eax, i(%rip)
>>
>> Same without 'volatile':
>> addl $1, i(%rip)
>
> aarch64 code (after changing variable declaration to 'extern' to avoid
> usage of '.LANCHOR0'):
> adrp x1, i
> ldr w0, [x1, #:lo12:i]
> add w0, w0, 1
> str w0, [x1, #:lo12:i]
>
> Without 'volatile': same code.
Yeah, this doesn't surprise me.
>>> Since the state of incyg is inspected in other threads to determine
>>> how to deliver a signal, marking these as 'volatile' might have been
>>> enough on single-processor systems, but perhaps we need a memory
>>> barrier here (as we already do in the assembly code which manipulates
>>> incyg), maybe by using InterlockedIncrement/Decrement()?
>>
>> Using 'incyg++' and 'incyg--' was possibly always not safe.
>>
>> InterlockedIncrement/Decrement() would help. These are #defined to
>> intrinsics _InterlockedIncrement/Decrement() which do the same as
>> __atomic_add/sub_fetch().
>>
>> #include <windows.h>
>> unsigned i; // volatile not needed
>>
>> void inc0() { _InterlockedIncrement(&i); }
>> void inc1() { __atomic_add_fetch(&i, 1, __ATOMIC_SEQ_CST); }
>>
>> x86_64 code:
>> lock addl $1, i(%rip)
>
> aarch64 code:
> adrp x0, i
> mov w1, 1
> add x0, x0, :lo12:i
> ldaddal w1, w1, [x0]
>
> There is no 'lock' prefix or 'ldadd*' in winsup/cygwin/scripts/gendef.
> The inc/decrements of 'incyg' are possibly not thread safe.
I don't think thread-safety in increment/decrement is actually an issue
because incyg is part of the TLS data which only the concerned thread
adjusts.
However, a memory barrier is needed here, because other threads may
inspect incyg and make decisions about the state of the thread based on
that (e.g. in signal delivery, and I think the assembler which interacts
with incyg there gets that right, at least in the x86_64 case...)
More information about the Cygwin-developers
mailing list