Re: master: Replace mem-reg XADD with load+add+store

What will actually break if it's:
ADD QWORD PTR [R13+16], 16       ; thread.binding-stack-pointer
MOV RDX, [R13+16]                ; thread.binding-stack-pointer

?

On Sat, Aug 29, 2026 at 1:35 AM Stas Boukarev <[email protected]> wrote:
>
> It might theoretically be faster, but the latency might be eaten by
> something else, making shorter code a net win.
> Have you observed any improvements empirically?
>
> On Sat, Aug 29, 2026 at 1:31 AM snuglas via Sbcl-commits
> <[email protected]> wrote:
> >
> > The branch "master" has been updated in SBCL:
> >        via  33c68e006d3a0a30df0e6e1007743e2a9e58e6d2 (commit)
> >       from  db35d456122e0be7795fd50e713ce0b76a8f996d (commit)
> >
> > - Log -----------------------------------------------------------------
> > commit 33c68e006d3a0a30df0e6e1007743e2a9e58e6d2
> > Author: Douglas Katzman <[email protected]>
> > Date:   Fri Aug 28 17:40:59 2026 -0400
> >
> >     Replace mem-reg XADD with load+add+store
> >
> >     llvm-mca confirms that 3 instructions is better than 2 in this case.
> >     XADD is mainly used with :LOCK so if you don't need :LOCK, you don't need XADD
> > ---
> >  src/compiler/x86-64/tls.lisp | 18 ++++++++++--------
> >  1 file changed, 10 insertions(+), 8 deletions(-)
> >
> > diff --git a/src/compiler/x86-64/tls.lisp b/src/compiler/x86-64/tls.lisp
> > index a19717654..6824348b9 100644
> > --- a/src/compiler/x86-64/tls.lisp
> > +++ b/src/compiler/x86-64/tls.lisp
> > @@ -369,18 +369,19 @@
> >      (inst mov (thread-tls-ea tls-index) val)))
> >
> >  (defun bind (bsp symbol tmp)
> > -  (inst mov bsp (* binding-size n-word-bytes))
> > -  (inst xadd (thread-slot-ea thread-binding-stack-pointer-slot) bsp)
> > +  (load-binding-stack-pointer bsp)
> > +  (inst add bsp (* binding-size n-word-bytes))
> > +  (store-binding-stack-pointer bsp)
> >    (let* ((tls-index (load-time-tls-offset symbol))
> >           (tls-cell (thread-tls-ea tls-index)))
> >      ;; Too bad we can't use "XCHG [thread + disp], val" to write new value
> >      ;; and read the old value in one step. It will violate the constraints
> >      ;; prescribed in the internal documentation on special binding.
> >      (inst mov tmp tls-cell)
> > -    (storew tmp bsp binding-value-slot)
> > +    (storew tmp bsp (- binding-value-slot binding-size))
> >      ;; Indices are small enough to be written as :DWORDs which avoids
> >      ;; a REX prefix if 'bsp' happens to be any of the low 8 registers.
> > -    (inst mov :dword (ea (ash binding-symbol-slot word-shift) bsp) tls-index)
> > +    (inst mov :dword (object-slot-ea bsp (- binding-symbol-slot binding-size) 0) tls-index)
> >      (values tls-cell tmp)))
> >
> >  (define-vop (bind) ; bind a known symbol
> > @@ -419,10 +420,11 @@
> >                             &optional (newval nil newvalp)
> >                             &aux (value-ea (ea 1 index-temp)))
> >    (inst mov val-temp (ea index-temp thread-tn))
> > -  (inst mov bsp (* binding-size n-word-bytes))
> > -  (inst xadd (thread-slot-ea thread-binding-stack-pointer-slot) bsp)
> > -  (inst mov (ea (ash binding-value-slot word-shift) bsp) val-temp)
> > -  (inst mov :dword (ea (ash binding-symbol-slot word-shift) bsp) index-temp)
> > +  (load-binding-stack-pointer bsp)
> > +  (inst add bsp (* binding-size n-word-bytes))
> > +  (store-binding-stack-pointer bsp)
> > +  (storew val-temp bsp (- binding-value-slot binding-size))
> > +  (inst mov :dword (object-slot-ea bsp (- binding-symbol-slot binding-size) 0) index-temp)
> >    ;; (usually) update the indirect pointer
> >    (cond ((not symbol) ; compile-time-unknown if indirection word exists
> >           (assemble ()
> >
> > -----------------------------------------------------------------------
> >
> >
> > hooks/post-receive
> > --
> > SBCL
> >
> >
> > _______________________________________________
> > Sbcl-commits mailing list
> > [email protected]
> > https://lists.sourceforge.net/lists/listinfo/sbcl-commits


_______________________________________________
Sbcl-commits mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/sbcl-commits