On Tue, 19 Nov 2024, Maciej W. Rozycki wrote:
> > time I think the RMW sequence on _Atomic variables is a clear
> > bug that you'd need to fix also for -mno-safe-bwa.
>
> That's weird indeed, GCC internals manual clearly says:
>
> "'atomic_storeMODE'
> This pattern implements an atomic store operation with memory model
> semantics. Operand 0 is the memory address being stored to.
> Operand 1 is the value to be written. Operand 2 is the memory
> model to be used for the operation.
>
> If not present, the '__atomic_store' built-in function will attempt
> to perform a normal store and surround it with any required memory
> fences. If the store would not be atomic, then an
> '__atomic_exchange' is attempted with the result being ignored."
>
> and while we do not have `atomic_storeqi' nor `atomic_storehi' operations
> we do have `atomic_exchangeqi' and `atomic_exchangehi' ones defined in the
> Alpha backend (for byte and word operations respectively). So presumably
> the middle end is not aware for some reason that on non-BWX Alpha normal
> QI and HI mode stores are not atomic.
FTR the reason is the middle end universally assumes, in the absence of
an 'atomic_storeMODE' pattern, that stores are atomic for object sizes of
up to the target's word size, which does not stand for non-BWX Alpha.
> I do hope this obviously obscure but trivial GCC bug was not the *sole*
> reason to drop non-BWX support from Linux.
I found no previous bug report, so this is now PR target/117759[1]. An
obvious bug fix has been posted for review as well[2].
References:
[1] "Alpha: `_Atomic' keyword not respected for !BWX and 8-bit/16-bit stores",
<https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117759>
[2] "Alpha: Respect `_Atomic' keyword for HI and QI data on !BWX [PR117759]",
<https://inbox.sourceware.org/gcc-patches/[email protected]/>
Maciej
