On Tue, Apr 09, 2024 at 06:52:02PM +0200, Stefan Kanthak wrote:
> "Eric Biggers" <[email protected]> wrote:
>
> > +.macro do_4rounds i, m0, m1, m2, m3
> > +.if \i < 16
> > + movdqu \i*4(DATA_PTR), MSG
> > + pshufb SHUF_MASK, MSG
> > + movdqa MSG, \m0
> > +.else
> > + movdqa \m0, MSG
> > +.endif
> > + paddd \i*4(SHA256CONSTANTS), MSG
>
> To load the round constant independent from and parallel to the previous
> instructions which use \m0 I recommend to change the first lines of the
> do_4rounds macro as follows (this might save 1+ cycle per macro invocation,
> and most obviously 2 lines):
>
> .macro do_4rounds i, m0, m1, m2, m3
> .if \i < 16
> movdqu \i*4(DATA_PTR), \m0
> pshufb SHUF_MASK, \m0
> .endif
> movdqa \i*4(SHA256CONSTANTS), MSG
> paddd \m0, MSG
> ...
Yes, your suggestion looks good. I don't see any performance difference on
Ice Lake, but it does shorten the source code. It belongs in a separate patch
though, since this patch isn't meant to change the output.
- Eric