On Sat, May 11, 2024 at 10:08:50PM +0200, Arnd Bergmann wrote:
> On Sat, May 11, 2024, at 21:37, Paul E. McKenney wrote:
> > On Sat, May 11, 2024 at 08:49:08PM +0200, John Paul Adrian Glaubitz wrote:
> >
> > The pre-EV56 Alphas have no byte store instruction, correct?
> >
> > If that is in fact correct, what code is generated for a volatile store
> > to a single byte for those CPUs? For example, for this example?
> >
> > char c;
> >
> > ...
> >
> > WRITE_ONCE(c, 3);
> >
> > The rumor I heard is that the compilers will generate a non-atomic
> > read-modify-write instruction sequence in this case, first reading the
> > 32-bit word containing that byte into a register, then substituting the
> > value to be stored into corresponding byte of that register, and finally
> > doing a 32-bit store from that register.
> >
> > Is that the case, or am I confused?
>
> I think it's slightly worse: gcc will actually do a 64-bit
> read-modify-write rather than a 32-bit one, and it doesn't
> use atomic ll/sc when storing into an _Atomic struct member:
>
> echo '#include <stdatomic.h>^M struct s { _Atomic char c; _Atomic char s[7]; long l; }; void f(struct s *s) { atomic_store(&s->c, 3); }' | alpha-linux-gcc-14 -xc - -S -o- -O2 -mcpu=ev5
>
> f:
> .frame $30,0,$26,0
> $LFB0:
> .cfi_startproc
> .prologue 0
> mb
> lda $1,3($31)
> insbl $1,$16,$1
> ldq_u $2,0($16)
> mskbl $2,$16,$2
> bis $1,$2,$1
> stq_u $1,0($16)
> bis $31,$31,$31
> mb
> ret $31,($26),1
> .cfi_endproc
> $LFE0:
> .end f
>
> compared to -mcpu=ev56:
>
> f:
> .frame $30,0,$26,0
> $LFB0:
> .cfi_startproc
> .prologue 0
> mb
> lda $1,3($31)
> stb $1,0($16)
> bis $31,$31,$31
> mb
> ret $31,($26),1
> .cfi_endproc
> $LFE0:
> .end f
Thank you, Arnd!
And that breaks things because it can clobber concurrent stores to
other bytes in that enclosing machine word.
Thanx, Paul
